diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_100k_semisupervised_1a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_100k_semisupervised_1a.sh
index e95de232304..87cd42a0a74 100644
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_100k_semisupervised_1a.sh
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_100k_semisupervised_1a.sh
@@ -43,12 +43,12 @@
 
 set -u -e -o pipefail
 
-stage=0   # Start from -1 for supervised seed system training
+stage=0
 train_stage=-100
 nj=80
 test_nj=50
 
-# The following 3 options decide the output directory for semi-supervised 
+# The following 3 options decide the output directory for semi-supervised
 # chain system
 # dir=${exp_root}/chain${chain_affix}/tdnn${tdnn_affix}
 
@@ -89,7 +89,7 @@ echo "$0 $@"  # Print the command line for logging
 if [ -f ./path.sh ]; then . ./path.sh; fi
 . ./utils/parse_options.sh
 
-# The following can be replaced with the versions that model
+# The following can be replaced with the versions that do not model
 # UNK using phone LM. $sup_lat_dir should also ideally be changed.
 unsup_decode_lang=data/lang_test_poco_sup100k_unk
 unsup_decode_graph_affix=_poco_sup100k_unk
@@ -141,6 +141,8 @@ if [ $stage -le 2 ]; then
 
   steps/make_mfcc.sh --nj $nj --cmd "$train_cmd" \
     --mfcc-config conf/mfcc_hires.conf data/${unsupervised_set}_sp_hires || exit 1
+  steps/compute_cmvn_stats.sh data/${unsupervised_set}_sp_hires
+  utils/fix_data_dir.sh data/${unsupervised_set}_sp_hires
 fi
 unsupervised_set_perturbed=${unsupervised_set}_sp
 
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_50k_semisupervised_1a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_50k_semisupervised_1a.sh
index 2d5b2f8480e..c5b7a822991 100755
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_50k_semisupervised_1a.sh
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_50k_semisupervised_1a.sh
@@ -109,7 +109,7 @@ test_graph_affix=_poco_unk
 
 unsup_rescore_lang=${unsup_decode_lang}_big
 
-dir=$exp_root/chain${chain_affix}/tdnn${tdnn_affix}
+dir=$exp_root/chain${chain_affix}/tdnn_lstm${tdnn_affix}
 
 if ! cuda-compiled; then
   cat <<EOF && exit 1
@@ -144,6 +144,8 @@ fi
 if [ $stage -le 2 ]; then
   utils/data/perturb_data_dir_speed_3way.sh data/${unsupervised_set} \
     data/${unsupervised_set_perturbed}_hires
+  utils/data/perturb_data_dir_volume.sh \
+    data/${unsupervised_set_perturbed}_hires
 
   steps/make_mfcc.sh --cmd "$train_cmd" --nj $nj \
     --mfcc-config conf/mfcc_hires.conf \
@@ -177,7 +179,7 @@ fi
 # Rescore undeterminized lattices with larger LM
 if [ $stage -le 5 ]; then
   steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-    --acwt 0.1 --beam 8.0  --skip-scoring true \
+    --acwt 0.1 --beam 8.0 --skip-scoring true \
     $unsup_decode_lang $unsup_rescore_lang \
     data/${unsupervised_set_perturbed}_hires \
     $sup_chain_dir/decode_${unsupervised_set_perturbed} \
@@ -433,6 +435,7 @@ if [ $stage -le 18 ]; then
       steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
         --nj $num_jobs --cmd "$decode_cmd" ${decode_iter:+--iter $decode_iter} \
         --online-ivector-dir $ivector_root_dir/ivectors_${decode_set}_hires \
+        --frames-per-chunk 160 \
         $test_graph_dir data/${decode_set}_hires \
         $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || touch $dir/.error
     ) &
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_ex250k_semisupervised_conf_smbr_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_ex250k_semisupervised_conf_smbr_a.sh
deleted file mode 100644
index 6f0a5f932fb..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_ex250k_semisupervised_conf_smbr_a.sh
+++ /dev/null
@@ -1,523 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_ex250k_1e_smbr  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-train_extra_opts="--chain.mmi-factor-schedule=1.0,1.0@0.2,0.5@0.2,0.5 --chain.smbr-factor-schedule=0.0,0.0@0.2,0.2@0.2,0.2"
-chain_smbr_extra_opts="--one-silence-class"
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_ex250k_unk
-unsup_rescore_lang=${unsup_decode_lang}_big
-unsup_decode_graph_affix=_poco_ex250k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    $train_extra_opts --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  --lang data/lang_chain_unk || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_ex250k_semisupervised_conf_smbr_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_ex250k_semisupervised_conf_smbr_b.sh
deleted file mode 100644
index 37362657651..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_ex250k_semisupervised_conf_smbr_b.sh
+++ /dev/null
@@ -1,529 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-smbr_leaky_hmm_coefficient=0.00001
-mmi_factor_schedule="output-0=1.0,1.0 output-1=0.2,0.2"
-smbr_factor_schedule="output-0=0.0,0.0 output-1=0.4,0.4"
-
-# Semi-supervised options
-comb_affix=comb_250k_ex250k_1b_smbr  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-chain_smbr_extra_opts="--one-silence-class"
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_ex250k_unk
-unsup_rescore_lang=${unsup_decode_lang}_big
-unsup_decode_graph_affix=_poco_ex250k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --chain.smbr-leaky-hmm-coefficient $smbr_leaky_hmm_coefficient \
-    --chain.mmi-factor-schedule="$mmi_factor_schedule" \
-    --chain.smbr-factor-schedule="$smbr_factor_schedule" \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  --lang data/lang_chain_unk || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_a.sh
deleted file mode 100644
index ebf52fa8b40..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_a.sh
+++ /dev/null
@@ -1,513 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=40
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7b  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_a
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_1a  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=3,2
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain
-unsup_decode_lang=data/lang_poco_test_sup100k
-unsup_decode_graph_affix=_poco_sup100k
-
-test_lang=data/lang_poco_test
-test_graph_affix=_poco
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-  fi
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. 2 \
-    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_c.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_c.sh
deleted file mode 100644
index f41374e4593..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_c.sh
+++ /dev/null
@@ -1,517 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=40
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7b  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_a
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_1c  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=768
-cell_dim=768
-projection_dim=192
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain
-unsup_decode_lang=data/lang_poco_test_sup100k
-unsup_decode_graph_affix=_poco_sup100k
-
-test_lang=data/lang_poco_test
-test_graph_affix=_poco
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-  fi
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_d.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_d.sh
deleted file mode 100644
index f5f41fd67c1..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_d.sh
+++ /dev/null
@@ -1,517 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7c  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_1d  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=768
-cell_dim=768
-projection_dim=192
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-  fi
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_e.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_e.sh
deleted file mode 100644
index 61c55686efe..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_e.sh
+++ /dev/null
@@ -1,516 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_1e  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_f.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_f.sh
deleted file mode 100644
index dfdc36d6428..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_f.sh
+++ /dev/null
@@ -1,516 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7g  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_e
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_1f  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-  
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode_semisup.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay 
-  output name=output-1 input=output.affine@$label_delay 
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay 
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay 
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true --constrained false \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true --constrained false $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 64 \
-    --lang2weight $supervision_weights --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_g.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_g.sh
deleted file mode 100644
index 9dcfb693eda..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_g.sh
+++ /dev/null
@@ -1,516 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7f  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_f
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_1g  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain
-unsup_decode_lang=data/lang_poco_test_sup100k
-unsup_decode_graph_affix=_poco_sup100k
-
-test_lang=data/lang_poco_test
-test_graph_affix=_poco
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if true || [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode_semisup.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 --sub-split 1 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    $treedir ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay 
-  output name=output-1 input=output.affine@$label_delay 
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay 
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true --constrained false \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true --constrained false $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 64 \
-    --lang2weight $supervision_weights --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_smbr_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_smbr_a.sh
deleted file mode 100644
index 23aa3531377..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_smbr_a.sh
+++ /dev/null
@@ -1,512 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb_250k_1b_smbr  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-chain_smbr_extra_opts="--one-silence-class"
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --chain.smbr-leaky-hmm-coefficient 0.00001 \
-    --chain.mmi-factor-schedule="output-0=1.0,1.0 output-1=0.5,0.5" \
-    --chain.smbr-factor-schedule="output-0=0.0,0.0 output-1=0.2,0.2" \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  --lang data/lang_chain_unk || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_smbr_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_smbr_b.sh
deleted file mode 100644
index ce8bf87c87d..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_250k_semisupervised_conf_smbr_b.sh
+++ /dev/null
@@ -1,437 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k
-semisup_train_set=    # semisup100k_250k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-smbr_leaky_hmm_coefficient=0.00001
-mmi_factor_schedule="output-0=1.0,1.0 output-1=0.2,0.2"
-smbr_factor_schedule="output-0=0.0,0.0 output-1=0.4,0.4"
-
-# Semi-supervised options
-comb_affix=comb_250k_1b_smbr  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-chain_smbr_extra_opts="--one-silence-class"
-lm_weights=3,2
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay
-  output name=output-1 input=output.affine@$label_delay
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 64 \
-    --lang2weight $supervision_weights --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --chain.smbr-leaky-hmm-coefficient $smbr_leaky_hmm_coefficient \
-    --chain.mmi-factor-schedule="$mmi_factor_schedule" \
-    --chain.smbr-factor-schedule="$smbr_factor_schedule" \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  --lang data/lang_chain_unk || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_ex500k_semisupervised_smbr_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_ex500k_semisupervised_smbr_a.sh
deleted file mode 100644
index ec5d2138730..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_ex500k_semisupervised_smbr_a.sh
+++ /dev/null
@@ -1,531 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_500k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_500k
-semisup_train_set=    # semisup100k_500k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-smbr_leaky_hmm_coefficient=0.00001
-mmi_factor_schedule="output-0=1.0,1.0 output-1=0.2,0.2"
-smbr_factor_schedule="output-0=0.0,0.0 output-1=0.4,0.4"
-
-# Semi-supervised options
-comb_affix=comb_500k_ex500k_1a_smbr  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-chain_smbr_extra_opts="--one-silence-class"
-lm_weights=3,1
-num_copies=2,1
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_test_poco_ex500k_unk
-unsup_rescore_lang=data/lang_test_poco_ex500k_big_unk
-unsup_decode_graph_affix=_poco_ex500k_unk
-
-test_lang=data/lang_test_poco_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-    steps/compute_cmvn_stats.sh data/${dset}_sp_hires
-    utils/fix_data_dir.sh data/${dset}_sp_hires
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode_semisup.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-decode_affix=${decode_affix}_big
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay 
-  output name=output-1 input=output.affine@$label_delay 
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay 
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay 
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 128 \
-    --lang2weight $supervision_weights --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --chain.smbr-leaky-hmm-coefficient $smbr_leaky_hmm_coefficient \
-    --chain.mmi-factor-schedule="$mmi_factor_schedule" \
-    --chain.smbr-factor-schedule="$smbr_factor_schedule" \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  --lang data/lang_chain_unk || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_kl_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_kl_a.sh
deleted file mode 100644
index e2a94495332..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_kl_a.sh
+++ /dev/null
@@ -1,442 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_500k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_500k
-semisup_train_set=    # semisup100k_500k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-mmi_factor_schedule="output-0=1.0,1.0 output-1=0.5,0.5"
-kl_factor_schedule="output-0=0.0,0.0 output-1=0.5,0.5"
-
-# Semi-supervised options
-comb_affix=comb_500k_1a_kl0.5  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=3,1
-num_copies=2,1
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-    steps/compute_cmvn_stats.sh data/${dset}_sp_hires
-    utils/fix_data_dir.sh data/${dset}_sp_hires
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode_semisup.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay 
-  output name=output-1 input=output.affine@$label_delay 
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay 
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay 
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 128 \
-    --lang2weight $supervision_weights --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --chain.mmi-factor-schedule="$mmi_factor_schedule" \
-    --chain.kl-factor-schedule="$kl_factor_schedule" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_smbr_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_smbr_b.sh
deleted file mode 100644
index 281a6e4d88d..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_smbr_b.sh
+++ /dev/null
@@ -1,518 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_500k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_500k
-semisup_train_set=    # semisup100k_500k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-smbr_leaky_hmm_coefficient=0.00001
-mmi_factor_schedule="output-0=1.0,1.0 output-1=0.2,0.2"
-smbr_factor_schedule="output-0=0.0,0.0 output-1=0.4,0.4"
-
-# Semi-supervised options
-comb_affix=comb_500k_1b_smbr  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-chain_smbr_extra_opts="--one-silence-class"
-lm_weights=3,1
-num_copies=
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-    steps/compute_cmvn_stats.sh data/${dset}_sp_hires
-    utils/fix_data_dir.sh data/${dset}_sp_hires
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode_semisup.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay 
-  output name=output-1 input=output.affine@$label_delay 
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay 
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay 
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 128 \
-    --lang2weight $supervision_weights --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --chain.smbr-leaky-hmm-coefficient $smbr_leaky_hmm_coefficient \
-    --chain.mmi-factor-schedule="$mmi_factor_schedule" \
-    --chain.smbr-factor-schedule="$smbr_factor_schedule" \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  --lang data/lang_chain_unk || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_smbr_c.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_smbr_c.sh
deleted file mode 100644
index f29b65c6e7b..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_500k_semisupervised_conf_smbr_c.sh
+++ /dev/null
@@ -1,446 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_500k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=80
-exp=exp/semisup_100k
-
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_500k
-semisup_train_set=    # semisup100k_500k
-
-tdnn_affix=7d_h1024  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_c
-
-nnet3_affix=    # affix for nnet3 and chain dir -- relates to i-vector used
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-smbr_leaky_hmm_coefficient=0.00001
-mmi_factor_schedule="output-0=1.0,1.0 output-1=0.2,0.2"
-smbr_factor_schedule="output-0=0.0,0.0 output-1=0.4,0.4"
-
-# Semi-supervised options
-comb_affix=comb_500k_1c_smbr  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-chain_smbr_extra_opts="--one-silence-class"
-lm_weights=3,1
-num_copies=2,1
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_100k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_sup100k_unk
-unsup_decode_graph_affix=_poco_sup100k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-    steps/compute_cmvn_stats.sh data/${dset}_sp_hires
-    utils/fix_data_dir.sh data/${dset}_sp_hires
-  fi
-
-  if [ $stage -le 4 ]; then
-    steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-      data/${dset}_sp_hires data/${dset}_sp_max2_hires
-
-    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $decode_nj \
-      data/${dset}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-      $exp/nnet3${nnet3_affix}/ivectors_${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode_semisup.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-  ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}/ || true
-done
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}_sp${decode_affix}/frame_subsampling_factor
-fi
-
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri4a
-
-treedir=$exp/chain${nnet3_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${nnet3_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-#if [ $stage -le 9 ]; then
-#  steps/subset_ali_dir.sh --cmd "$train_cmd" \
-#    data/${unsupervised_set} data/${unsupervised_set}_sp_hires \
-#    $chaindir/best_path_${unsupervised_set}_sp${decode_affix} \
-#    $chaindir/best_path_${unsupervised_set}${decode_affix}
-#  echo $frame_subsampling_factor > $chaindir/best_path_${unsupervised_set}${decode_affix}/frame_subsampling_factor
-#fi
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm4 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay 
-  output name=output-1 input=output.affine@$label_delay 
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay 
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay 
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${nnet3_affix}/tri4a_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${unsupervised_set}_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 128 \
-    --lang2weight $supervision_weights --lang2num-copies "$num_copies" \
-    2 $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --chain.smbr-leaky-hmm-coefficient $smbr_leaky_hmm_coefficient \
-    --chain.mmi-factor-schedule="$mmi_factor_schedule" \
-    --chain.smbr-factor-schedule="$smbr_factor_schedule" \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  --lang data/lang_chain_unk || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_b.sh
deleted file mode 100755
index e686d977ded..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_b.sh
+++ /dev/null
@@ -1,225 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7b
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_a
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
-lang=data/lang_chain
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_c.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_c.sh
deleted file mode 100755
index 1854a4a86e1..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_c.sh
+++ /dev/null
@@ -1,225 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7c
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_c
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
-lang=data/lang_chain_unk
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang_unk $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang_unk $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-EOF
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_d.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_d.sh
deleted file mode 100755
index 265a8c05a11..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_d.sh
+++ /dev/null
@@ -1,228 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7d
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_c
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
-lang=data/lang_chain_unk
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang_unk $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang_unk $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm4 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-EOF
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_f.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_f.sh
deleted file mode 100755
index 9bc98d90934..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_f.sh
+++ /dev/null
@@ -1,228 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7f
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_e
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
-lang=data/lang_chain
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm4 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-EOF
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true --constrained false" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_b.sh
deleted file mode 100755
index 988299a4621..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_b.sh
+++ /dev/null
@@ -1,244 +0,0 @@
-#!/bin/bash
-set -e
-set -u
-
-# This is oracle experiment for semi-supervised training with 100 hours 
-# of supervised data and 250 hours of unsupervised data
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7b_oracle
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k_n10k
-base_train_set=train_oracle100k_250k_n10k
-tree_affix=bi_a
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if [ $stage -le 6 ]; then
-  utils/combine_data.sh data/${base_train_set}_sp_hires \
-    data/${supervised_set}_sp_hires data/${unsupervised_set}_sp_hires || exit 1
-
-  utils/combine_data.sh data/${base_train_set} \
-    data/${supervised_set} data/${unsupervised_set} || exit 1
-
-  utils/data/perturb_data_dir_speed_3way.sh data/${base_train_set} \
-    data/${base_train_set}_sp || exit 1
-fi
-
-if [ $stage -le 7 ]; then
-  steps/make_mfcc.sh --nj 40 data/${base_train_set}_sp || exit 1
-  steps/compute_cmvn_stats.sh data/${base_train_set}_sp || exit 1
-  utils/fix_data_dir.sh data/${base_train_set}_sp
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${base_train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${base_train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires
-lang=data/lang_chain
-
-if [ $stage -le 8 ]; then
-  steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-    data/${base_train_set}_sp_hires data/${base_train_set}_sp_max2_hires
-
-  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 30 \
-    data/${base_train_set}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-    $exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires || exit 1;
-fi
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${base_train_set}_sp \
-    data/lang $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: Could not find $treedir/final.mdl"
-  exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-EOF
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_c.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_c.sh
deleted file mode 100755
index b21dd72a37a..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_c.sh
+++ /dev/null
@@ -1,247 +0,0 @@
-#!/bin/bash
-set -e
-set -u
-
-# This is oracle experiment for semi-supervised training with 100 hours 
-# of supervised data and 250 hours of unsupervised data
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7c_oracle
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k_n10k
-base_train_set=train_oracle100k_250k_n10k
-tree_affix=bi_a
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if [ $stage -le 6 ]; then
-  utils/combine_data.sh data/${base_train_set}_sp_hires \
-    data/${supervised_set}_sp_hires data/${unsupervised_set}_sp_hires || exit 1
-
-  utils/combine_data.sh data/${base_train_set} \
-    data/${supervised_set} data/${unsupervised_set} || exit 1
-
-  utils/data/perturb_data_dir_speed_3way.sh data/${base_train_set} \
-    data/${base_train_set}_sp || exit 1
-fi
-
-if [ $stage -le 7 ]; then
-  steps/make_mfcc.sh --nj 40 data/${base_train_set}_sp || exit 1
-  steps/compute_cmvn_stats.sh data/${base_train_set}_sp || exit 1
-  utils/fix_data_dir.sh data/${base_train_set}_sp
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${base_train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${base_train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires
-lang=data/lang_chain
-
-if [ $stage -le 8 ]; then
-  steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-    data/${base_train_set}_sp_hires data/${base_train_set}_sp_max2_hires
-
-  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 30 \
-    data/${base_train_set}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-    $exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires || exit 1;
-fi
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${base_train_set}_sp \
-    data/lang $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: Could not find $treedir/final.mdl"
-  exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=dnn2 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=dnn4 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=dnn6 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-EOF
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_d.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_d.sh
deleted file mode 100755
index 876633fedd6..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_oracle_d.sh
+++ /dev/null
@@ -1,247 +0,0 @@
-#!/bin/bash
-set -e
-set -u
-
-# This is oracle experiment for semi-supervised training with 100 hours 
-# of supervised data and 250 hours of unsupervised data
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7d_oracle
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k_n10k
-base_train_set=train_oracle100k_250k_n10k
-tree_affix=bi_a
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if [ $stage -le 6 ]; then
-  utils/combine_data.sh data/${base_train_set}_sp_hires \
-    data/${supervised_set}_sp_hires data/${unsupervised_set}_sp_hires || exit 1
-
-  utils/combine_data.sh data/${base_train_set} \
-    data/${supervised_set} data/${unsupervised_set} || exit 1
-
-  utils/data/perturb_data_dir_speed_3way.sh data/${base_train_set} \
-    data/${base_train_set}_sp || exit 1
-fi
-
-if [ $stage -le 7 ]; then
-  steps/make_mfcc.sh --nj 40 data/${base_train_set}_sp || exit 1
-  steps/compute_cmvn_stats.sh data/${base_train_set}_sp || exit 1
-  utils/fix_data_dir.sh data/${base_train_set}_sp
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${base_train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${base_train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires
-lang=data/lang_chain
-
-if [ $stage -le 8 ]; then
-  steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-    data/${base_train_set}_sp_hires data/${base_train_set}_sp_max2_hires
-
-  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 30 \
-    data/${base_train_set}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-    $exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires || exit 1;
-fi
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${base_train_set}_sp \
-    data/lang $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: Could not find $treedir/final.mdl"
-  exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm4 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm4 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm4 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-EOF
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1b.sh
new file mode 100644
index 00000000000..edf9f2eaac2
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1b.sh
@@ -0,0 +1,456 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script is semi-supervised recipe with 100 hours of supervised data
+# and 250 hours unsupervised data with naive splitting.
+# Based on "Semi-Supervised Training of Acoustic Models using Lattice-Free MMI",
+# Vimal Manohar, Hossein Hadian, Daniel Povey, Sanjeev Khudanpur, ICASSP 2018
+# http://www.danielpovey.com/files/2018_icassp_semisupervised_mmi.pdf
+# local/semisup/run_100k.sh shows how to call this.
+
+# This version of script uses only supervised data for i-vector extractor
+# training as against using the combined data as in run_tdnn_50k_semisupervised.sh.
+# We use 3-gram LM trained on 100 hours of supervised data. We do not have
+# enough data to do 4-gram LM rescoring as in run_tdnn_50k_semisupervised.sh.
+
+# This script uses phone LM to model UNK.
+# This script uses the same tree as that for the seed model.
+# See the comments in the script about how to change these.
+
+# Unsupervised set: train_unsup100k_250k (250 hour subset of Fisher excluding 100 hours for supervised)
+# unsup_frames_per_eg=150
+# Deriv weights: Lattice posterior of best path pdf
+# Unsupervised weight: 1.0
+# Weights for phone LM (supervised, unsupervised): 3,2
+# LM for decoding unsupervised data: 4gram
+# Supervision: Naive split lattices
+
+set -u -e -o pipefail
+
+stage=0
+train_stage=-100
+nj=80
+test_nj=50
+
+# The following 3 options decide the output directory for semi-supervised
+# chain system
+# dir=${exp_root}/chain${chain_affix}/tdnn_lstm${tdnn_affix}
+
+exp_root=exp/semisup_100k
+chain_affix=    # affix for chain dir
+tdnn_affix=_semisup_1b  # affix for semi-supervised chain system
+
+# Datasets -- Expects data/$supervised_set and data/$unsupervised_set to be
+# present
+supervised_set=train_sup
+unsupervised_set=train_unsup100k_250k
+
+# Input seed system
+sup_chain_dir=exp/semisup_100k/chain/tdnn_lstm_1b_sp  # supervised chain system
+sup_lat_dir=exp/semisup_100k/chain/tri4a_train_sup_unk_lats  # Seed model options
+sup_tree_dir=exp/semisup_100k/chain/tree_bi_a  # tree directory for supervised chain system
+ivector_root_dir=exp/semisup_100k/nnet3  # i-vector extractor root directory
+
+# Semi-supervised options
+supervision_weights=1.0,1.0   # Weights for supervised, unsupervised data egs.
+                              # Can be used to scale down the effect of unsupervised data
+                              # by using a smaller scale for it e.g. 1.0,0.3
+lm_weights=3,2  # Weights on phone counts from supervised, unsupervised data for denominator FST creation
+
+sup_egs_dir=   # Supply this to skip supervised egs creation
+unsup_egs_dir=  # Supply this to skip unsupervised egs creation
+unsup_egs_opts=  # Extra options to pass to unsupervised egs creation
+
+# Neural network opts
+hidden_dim=1024
+cell_dim=1024
+projection_dim=256
+
+# training options
+num_epochs=2
+minibatch_size=64,32
+chunk_left_context=40
+chunk_right_context=0
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=5
+
+# decode options
+extra_left_context=50
+extra_right_context=0
+frames_per_chunk_decoding=160
+
+decode_iter=  # Iteration to decode with
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
+. ./utils/parse_options.sh
+
+# The following can be replaced with the versions that do not model
+# UNK using phone LM. $sup_lat_dir should also ideally be changed.
+unsup_decode_lang=data/lang_test_poco_sup100k_unk
+unsup_decode_graph_affix=_poco_sup100k_unk
+test_lang=data/lang_test_poco_unk
+test_graph_affix=_poco_unk
+
+dir=$exp_root/chain${chain_affix}/tdnn_lstm${tdnn_affix}
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+supervised_set_perturbed=${supervised_set}_sp
+unsupervised_set_perturbed=${unsupervised_set}_sp
+
+sup_ivector_dir=$ivector_root_dir/ivectors_${supervised_set_perturbed}_hires
+
+graphdir=$sup_chain_dir/graph${unsup_decode_graph_affix}
+
+for f in data/${supervised_set_perturbed}/feats.scp \
+  data/${supervised_set_perturbed}_hires/feats.scp \
+  $ivector_root_dir/extractor/final.ie $sup_ivector_dir/ivector_online.scp \
+  $sup_lat_dir/lat.1.gz $sup_tree_dir/ali.1.gz \
+  $unsup_decode_lang/G.fst; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find file $f"
+    exit 1
+  fi
+done
+
+if [ ! -f $graphdir/HCLG.fst ]; then
+  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $sup_chain_dir $graphdir
+fi
+
+# Prepare the speed-perturbed unsupervised data directory
+if [ $stage -le 2 ]; then
+  if [ -f data/${unsupervised_set}_sp_hires/feats.scp ]; then
+    echo "$0: data/${unsupervised_set}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
+    exit 1
+  fi
+
+  utils/data/perturb_data_dir_speed_3way.sh data/$unsupervised_set data/${unsupervised_set}_sp_hires
+  utils/data/perturb_data_dir_volume.sh data/${unsupervised_set}_sp_hires
+
+  steps/make_mfcc.sh --nj $nj --cmd "$train_cmd" \
+    --mfcc-config conf/mfcc_hires.conf data/${unsupervised_set}_sp_hires || exit 1
+  steps/compute_cmvn_stats.sh data/${unsupervised_set}_sp_hires
+  utils/fix_data_dir.sh data/${unsupervised_set}_sp_hires
+fi
+unsupervised_set_perturbed=${unsupervised_set}_sp
+
+# Extract i-vectors for the unsupervised data
+if [ $stage -le 3 ]; then
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+    data/${unsupervised_set_perturbed}_hires data/${unsupervised_set_perturbed}_max2_hires
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+    data/${unsupervised_set_perturbed}_max2_hires $ivector_root_dir/extractor \
+    $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires || exit 1
+fi
+
+# Decode unsupervised data and write lattices in non-compact
+# undeterminized format
+# Set --skip-scoring to false in order to score the unsupervised data
+if [ $stage -le 5 ]; then
+  echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
+  steps/nnet3/decode_semisup.sh --num-threads 4 --nj $nj --cmd "$decode_cmd" \
+            --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
+            --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+            --frames-per-chunk 160 \
+            --extra-left-context $extra_left_context \
+            --extra-right-context $extra_right_context \
+            --extra-left-context-initial 0 --extra-right-context-final 0 \
+            --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
+            $graphdir data/${unsupervised_set_perturbed}_hires $sup_chain_dir/decode_${unsupervised_set_perturbed}
+fi
+
+# Get best path alignment and lattice posterior of best path alignment to be
+# used as frame-weights in lattice-based training
+if [ $stage -le 8 ]; then
+  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
+    data/${unsupervised_set_perturbed}_hires \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed} \
+    $sup_chain_dir/best_path_${unsupervised_set_perturbed}
+fi
+
+frame_subsampling_factor=1
+if [ -f $sup_chain_dir/frame_subsampling_factor ]; then
+  frame_subsampling_factor=$(cat $sup_chain_dir/frame_subsampling_factor)
+fi
+cmvn_opts=$(cat $sup_chain_dir/cmvn_opts) || exit 1
+
+diff $sup_tree_dir/tree $sup_chain_dir/tree || { echo "$0: $sup_tree_dir/tree and $sup_chain_dir/tree differ"; exit 1; }
+
+# Uncomment the following lines if you need to build new tree using both
+# supervised and unsupervised data. This may help if amount of
+# supervised data used to train the seed system tree is very small.
+# unsupervised data
+
+# tree_affix=bi_semisup_a
+# treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+# if [ -f $treedir/final.mdl ]; then
+#   echo "$0: $treedir/final.mdl exists. Remove it and run again."
+#   exit 1
+# fi
+#
+# if [ $stage -le 9 ]; then
+#   # This is usually 3 for chain systems.
+#   echo $frame_subsampling_factor > \
+#     $sup_chain_dir/best_path_${unsupervised_set_perturbed}/frame_subsampling_factor
+#
+#   # This should be 1 if using a different source for supervised data alignments.
+#   # However alignments in seed tree directory have already been sub-sampled.
+#   echo $frame_subsampling_factor > \
+#     $sup_tree_dir/frame_subsampling_factor
+#
+#   # Build a new tree using stats from both supervised and unsupervised data
+#   steps/nnet3/chain/build_tree_multiple_sources.sh \
+#     --use-fmllr false --context-opts "--context-width=2 --central-position=1" \
+#     --frame-subsampling-factor $frame_subsampling_factor \
+#     7000 $lang \
+#     data/${supervised_set_perturbed} \
+#     ${sup_tree_dir} \
+#     data/${unsupervised_set_perturbed} \
+#     $chaindir/best_path_${unsupervised_set_perturbed} \
+#     $treedir || exit 1
+# fi
+#
+# sup_tree_dir=$treedir   # Use the new tree dir for further steps
+
+# Train denominator FST using phone alignments from
+# supervised and unsupervised data
+if [ $stage -le 10 ]; then
+  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
+    ${sup_tree_dir} ${sup_chain_dir}/best_path_${unsupervised_set_perturbed} \
+    $dir
+fi
+
+if [ $stage -le 11 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $sup_tree_dir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=40"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
+
+  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+
+  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+  # We use separate outputs for supervised and unsupervised data
+  # so we can properly track the train and valid objectives.
+
+  output name=output-0 input=output.affine
+  output name=output-1 input=output.affine
+
+  output name=output-0-xent input=output-xent.log-softmax
+  output name=output-1-xent input=output-xent.log-softmax
+EOF
+
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+. $dir/configs/vars
+
+left_context=$[model_left_context + chunk_left_context]
+right_context=$[model_right_context + chunk_right_context]
+left_context_initial=$model_left_context
+right_context_final=$model_right_context
+
+egs_left_context=$(perl -e "print int($left_context + $frame_subsampling_factor / 2)")
+egs_right_context=$(perl -e "print int($right_context + $frame_subsampling_factor / 2)")
+egs_left_context_initial=$(perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)")
+egs_right_context_final=$(perl -e "print int($right_context_final + $frame_subsampling_factor / 2)")
+
+if [ -z "$sup_egs_dir" ]; then
+  sup_egs_dir=$dir/egs_${supervised_set_perturbed}
+  frames_per_eg=$(cat $sup_chain_dir/egs/info/frames_per_eg)
+
+  if [ $stage -le 12 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
+    fi
+    mkdir -p $sup_egs_dir/
+    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the supervised data"
+    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
+               --left-context $egs_left_context --right-context $egs_right_context \
+               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+               --frame-subsampling-factor $frame_subsampling_factor \
+               --alignment-subsampling-factor $frame_subsampling_factor \
+               --frames-per-eg $frames_per_eg \
+               --frames-per-iter 1500000 \
+               --cmvn-opts "$cmvn_opts" \
+               --online-ivector-dir $sup_ivector_dir \
+               --generate-egs-scp true \
+               data/${supervised_set_perturbed}_hires $dir \
+               $sup_lat_dir $sup_egs_dir
+  fi
+else
+  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
+fi
+
+unsup_frames_per_eg=150  # Using a frames-per-eg of 150 for unsupervised data
+                         # was found to be better than allowing smaller chunks
+                         # (160,140,110,80) like for supervised system
+lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices when
+                      # creating numerator supervision
+lattice_prune_beam=4.0  # beam for pruning the lattices prior to getting egs
+                        # for unsupervised data
+tolerance=1   # frame-tolerance for chain training
+
+unsup_lat_dir=${sup_chain_dir}/decode_${unsupervised_set_perturbed}
+if [ -z "$unsup_egs_dir" ]; then
+  unsup_egs_dir=$dir/egs_${unsupervised_set_perturbed}
+
+  if [ $stage -le 13 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
+    fi
+    mkdir -p $unsup_egs_dir
+    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the unsupervised data"
+    steps/nnet3/chain/get_egs.sh \
+      --cmd "$decode_cmd" --alignment-subsampling-factor 1 \
+      --left-tolerance $tolerance --right-tolerance $tolerance \
+      --left-context $egs_left_context --right-context $egs_right_context \
+      --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+      --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
+      --frame-subsampling-factor $frame_subsampling_factor \
+      --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
+      --lattice-prune-beam "$lattice_prune_beam" \
+      --deriv-weights-scp $sup_chain_dir/best_path_${unsupervised_set_perturbed}/weights.scp \
+      --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+      --generate-egs-scp true $unsup_egs_opts \
+      data/${unsupervised_set_perturbed}_hires $dir \
+      $unsup_lat_dir $unsup_egs_dir
+  fi
+fi
+
+comb_egs_dir=$dir/comb_egs
+if [ $stage -le 14 ]; then
+  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
+    --block-size 128 \
+    --lang2weight $supervision_weights 2 \
+    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
+  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
+fi
+
+if [ $train_stage -le -4 ]; then
+  # This is to skip stages of den-fst creation, which was already done.
+  train_stage=-4
+fi
+
+if [ $stage -le 15 ]; then
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$comb_egs_dir" \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $sup_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights true \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.chunk-width $frames_per_eg \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.chunk-left-context-initial 0 \
+    --egs.chunk-right-context-final 0 \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.num-chunk-per-minibatch "$minibatch_size" \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.max-param-change 2.0 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --cleanup.remove-egs false \
+    --feat-dir data/${supervised_set_perturbed}_hires \
+    --tree-dir $sup_tree_dir \
+    --lat-dir $sup_lat_dir \
+    --dir $dir || exit 1;
+fi
+
+test_graph_dir=$dir/graph${test_graph_affix}
+if [ $stage -le 17 ]; then
+  # Note: it might appear that this $lang directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --self-loop-scale 1.0 ${test_lang} $dir $test_graph_dir
+fi
+
+if [ $stage -le 18 ]; then
+  rm -f $dir/.error
+  for decode_set in dev test; do
+    (
+      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
+      if [ $num_jobs -gt $test_nj ]; then num_jobs=$test_nj; fi
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+        --nj $num_jobs --cmd "$decode_cmd" ${decode_iter:+--iter $decode_iter} \
+        --online-ivector-dir $ivector_root_dir/ivectors_${decode_set}_hires \
+        --frames-per-chunk $frames_per_chunk_decoding \
+        --extra-left-context $extra_left_context \
+        --extra-right-context $extra_right_context \
+        --extra-left-context-initial 0 --extra-right-context-final 0 \
+        $test_graph_dir data/${decode_set}_hires \
+        $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || touch $dir/.error
+    ) &
+  done
+  wait;
+  if [ -f $dir/.error ]; then
+    echo "$0: Decoding failed. See $dir/decode${test_graph_affix}_*/log/*"
+    exit 1
+  fi
+fi
+
+exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1d.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1d.sh
new file mode 100644
index 00000000000..466f4fb7c91
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1d.sh
@@ -0,0 +1,456 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script is semi-supervised recipe with 100 hours of supervised data
+# and 250 hours unsupervised data with naive splitting.
+# Based on "Semi-Supervised Training of Acoustic Models using Lattice-Free MMI",
+# Vimal Manohar, Hossein Hadian, Daniel Povey, Sanjeev Khudanpur, ICASSP 2018
+# http://www.danielpovey.com/files/2018_icassp_semisupervised_mmi.pdf
+# local/semisup/run_100k.sh shows how to call this.
+
+# This version of script uses only supervised data for i-vector extractor
+# training as against using the combined data as in run_tdnn_50k_semisupervised.sh.
+# We use 3-gram LM trained on 100 hours of supervised data. We do not have
+# enough data to do 4-gram LM rescoring as in run_tdnn_50k_semisupervised.sh.
+
+# This script uses phone LM to model UNK.
+# This script uses the same tree as that for the seed model.
+# See the comments in the script about how to change these.
+
+# Unsupervised set: train_unsup100k_250k (250 hour subset of Fisher excluding 100 hours for supervised)
+# unsup_frames_per_eg=150
+# Deriv weights: Lattice posterior of best path pdf
+# Unsupervised weight: 1.0
+# Weights for phone LM (supervised, unsupervised): 3,2
+# LM for decoding unsupervised data: 4gram
+# Supervision: Naive split lattices
+
+set -u -e -o pipefail
+
+stage=0
+train_stage=-100
+nj=80
+test_nj=50
+
+# The following 3 options decide the output directory for semi-supervised
+# chain system
+# dir=${exp_root}/chain${chain_affix}/tdnn_lstm${tdnn_affix}
+
+exp_root=exp/semisup_100k
+chain_affix=    # affix for chain dir
+tdnn_affix=_semisup_1d  # affix for semi-supervised chain system
+
+# Datasets -- Expects data/$supervised_set and data/$unsupervised_set to be
+# present
+supervised_set=train_sup
+unsupervised_set=train_unsup100k_250k
+
+# Input seed system
+sup_chain_dir=exp/semisup_100k/chain/tdnn_lstm_1d_sp  # supervised chain system
+sup_lat_dir=exp/semisup_100k/chain/tri4a_train_sup_lats  # Seed model options
+sup_tree_dir=exp/semisup_100k/chain/tree_bi_c  # tree directory for supervised chain system
+ivector_root_dir=exp/semisup_100k/nnet3  # i-vector extractor root directory
+
+# Semi-supervised options
+supervision_weights=1.0,1.0   # Weights for supervised, unsupervised data egs.
+                              # Can be used to scale down the effect of unsupervised data
+                              # by using a smaller scale for it e.g. 1.0,0.3
+lm_weights=3,2  # Weights on phone counts from supervised, unsupervised data for denominator FST creation
+
+sup_egs_dir=   # Supply this to skip supervised egs creation
+unsup_egs_dir=  # Supply this to skip unsupervised egs creation
+unsup_egs_opts=  # Extra options to pass to unsupervised egs creation
+
+# Neural network opts
+hidden_dim=1024
+cell_dim=1024
+projection_dim=256
+
+# training options
+num_epochs=2
+minibatch_size=64,32
+chunk_left_context=40
+chunk_right_context=0
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=5
+
+# decode options
+extra_left_context=50
+extra_right_context=0
+frames_per_chunk_decoding=160
+
+decode_iter=  # Iteration to decode with
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
+. ./utils/parse_options.sh
+
+# The following can be replaced with the versions that do not model
+# UNK using phone LM. $sup_lat_dir should also ideally be changed.
+unsup_decode_lang=data/lang_test_poco_sup100k
+unsup_decode_graph_affix=_poco_sup100k
+test_lang=data/lang_test_poco
+test_graph_affix=_poco
+
+dir=$exp_root/chain${chain_affix}/tdnn_lstm${tdnn_affix}
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+supervised_set_perturbed=${supervised_set}_sp
+unsupervised_set_perturbed=${unsupervised_set}_sp
+
+sup_ivector_dir=$ivector_root_dir/ivectors_${supervised_set_perturbed}_hires
+
+graphdir=$sup_chain_dir/graph${unsup_decode_graph_affix}
+
+for f in data/${supervised_set_perturbed}/feats.scp \
+  data/${supervised_set_perturbed}_hires/feats.scp \
+  $ivector_root_dir/extractor/final.ie $sup_ivector_dir/ivector_online.scp \
+  $sup_lat_dir/lat.1.gz $sup_tree_dir/ali.1.gz \
+  $unsup_decode_lang/G.fst; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find file $f"
+    exit 1
+  fi
+done
+
+if [ ! -f $graphdir/HCLG.fst ]; then
+  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $sup_chain_dir $graphdir
+fi
+
+# Prepare the speed-perturbed unsupervised data directory
+if [ $stage -le 2 ]; then
+  if [ -f data/${unsupervised_set}_sp_hires/feats.scp ]; then
+    echo "$0: data/${unsupervised_set}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
+    exit 1
+  fi
+
+  utils/data/perturb_data_dir_speed_3way.sh data/$unsupervised_set data/${unsupervised_set}_sp_hires
+  utils/data/perturb_data_dir_volume.sh data/${unsupervised_set}_sp_hires
+
+  steps/make_mfcc.sh --nj $nj --cmd "$train_cmd" \
+    --mfcc-config conf/mfcc_hires.conf data/${unsupervised_set}_sp_hires || exit 1
+  steps/compute_cmvn_stats.sh data/${unsupervised_set}_sp_hires
+  utils/fix_data_dir.sh data/${unsupervised_set}_sp_hires
+fi
+unsupervised_set_perturbed=${unsupervised_set}_sp
+
+# Extract i-vectors for the unsupervised data
+if [ $stage -le 3 ]; then
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+    data/${unsupervised_set_perturbed}_hires data/${unsupervised_set_perturbed}_max2_hires
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+    data/${unsupervised_set_perturbed}_max2_hires $ivector_root_dir/extractor \
+    $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires || exit 1
+fi
+
+# Decode unsupervised data and write lattices in non-compact
+# undeterminized format
+# Set --skip-scoring to false in order to score the unsupervised data
+if [ $stage -le 5 ]; then
+  echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
+  steps/nnet3/decode_semisup.sh --num-threads 4 --nj $nj --cmd "$decode_cmd" \
+            --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
+            --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+            --frames-per-chunk 160 \
+            --extra-left-context $extra_left_context \
+            --extra-right-context $extra_right_context \
+            --extra-left-context-initial 0 --extra-right-context-final 0 \
+            --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
+            $graphdir data/${unsupervised_set_perturbed}_hires $sup_chain_dir/decode_${unsupervised_set_perturbed}
+fi
+
+# Get best path alignment and lattice posterior of best path alignment to be
+# used as frame-weights in lattice-based training
+if [ $stage -le 8 ]; then
+  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
+    data/${unsupervised_set_perturbed}_hires \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed} \
+    $sup_chain_dir/best_path_${unsupervised_set_perturbed}
+fi
+
+frame_subsampling_factor=1
+if [ -f $sup_chain_dir/frame_subsampling_factor ]; then
+  frame_subsampling_factor=$(cat $sup_chain_dir/frame_subsampling_factor)
+fi
+cmvn_opts=$(cat $sup_chain_dir/cmvn_opts) || exit 1
+
+diff $sup_tree_dir/tree $sup_chain_dir/tree || { echo "$0: $sup_tree_dir/tree and $sup_chain_dir/tree differ"; exit 1; }
+
+# Uncomment the following lines if you need to build new tree using both
+# supervised and unsupervised data. This may help if amount of
+# supervised data used to train the seed system tree is very small.
+# unsupervised data
+
+# tree_affix=bi_semisup_a
+# treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+# if [ -f $treedir/final.mdl ]; then
+#   echo "$0: $treedir/final.mdl exists. Remove it and run again."
+#   exit 1
+# fi
+#
+# if [ $stage -le 9 ]; then
+#   # This is usually 3 for chain systems.
+#   echo $frame_subsampling_factor > \
+#     $sup_chain_dir/best_path_${unsupervised_set_perturbed}/frame_subsampling_factor
+#
+#   # This should be 1 if using a different source for supervised data alignments.
+#   # However alignments in seed tree directory have already been sub-sampled.
+#   echo $frame_subsampling_factor > \
+#     $sup_tree_dir/frame_subsampling_factor
+#
+#   # Build a new tree using stats from both supervised and unsupervised data
+#   steps/nnet3/chain/build_tree_multiple_sources.sh \
+#     --use-fmllr false --context-opts "--context-width=2 --central-position=1" \
+#     --frame-subsampling-factor $frame_subsampling_factor \
+#     7000 $lang \
+#     data/${supervised_set_perturbed} \
+#     ${sup_tree_dir} \
+#     data/${unsupervised_set_perturbed} \
+#     $chaindir/best_path_${unsupervised_set_perturbed} \
+#     $treedir || exit 1
+# fi
+#
+# sup_tree_dir=$treedir   # Use the new tree dir for further steps
+
+# Train denominator FST using phone alignments from
+# supervised and unsupervised data
+if [ $stage -le 10 ]; then
+  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
+    ${sup_tree_dir} ${sup_chain_dir}/best_path_${unsupervised_set_perturbed} \
+    $dir
+fi
+
+if [ $stage -le 11 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $sup_tree_dir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=40"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
+
+  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+
+  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+  # We use separate outputs for supervised and unsupervised data
+  # so we can properly track the train and valid objectives.
+
+  output name=output-0 input=output.affine
+  output name=output-1 input=output.affine
+
+  output name=output-0-xent input=output-xent.log-softmax
+  output name=output-1-xent input=output-xent.log-softmax
+EOF
+
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+. $dir/configs/vars
+
+left_context=$[model_left_context + chunk_left_context]
+right_context=$[model_right_context + chunk_right_context]
+left_context_initial=$model_left_context
+right_context_final=$model_right_context
+
+egs_left_context=$(perl -e "print int($left_context + $frame_subsampling_factor / 2)")
+egs_right_context=$(perl -e "print int($right_context + $frame_subsampling_factor / 2)")
+egs_left_context_initial=$(perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)")
+egs_right_context_final=$(perl -e "print int($right_context_final + $frame_subsampling_factor / 2)")
+
+if [ -z "$sup_egs_dir" ]; then
+  sup_egs_dir=$dir/egs_${supervised_set_perturbed}
+  frames_per_eg=$(cat $sup_chain_dir/egs/info/frames_per_eg)
+
+  if [ $stage -le 12 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
+    fi
+    mkdir -p $sup_egs_dir/
+    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the supervised data"
+    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
+               --left-context $egs_left_context --right-context $egs_right_context \
+               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+               --frame-subsampling-factor $frame_subsampling_factor \
+               --alignment-subsampling-factor $frame_subsampling_factor \
+               --frames-per-eg $frames_per_eg \
+               --frames-per-iter 1500000 \
+               --cmvn-opts "$cmvn_opts" \
+               --online-ivector-dir $sup_ivector_dir \
+               --generate-egs-scp true \
+               data/${supervised_set_perturbed}_hires $dir \
+               $sup_lat_dir $sup_egs_dir
+  fi
+else
+  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
+fi
+
+unsup_frames_per_eg=150  # Using a frames-per-eg of 150 for unsupervised data
+                         # was found to be better than allowing smaller chunks
+                         # (160,140,110,80) like for supervised system
+lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices when
+                      # creating numerator supervision
+lattice_prune_beam=4.0  # beam for pruning the lattices prior to getting egs
+                        # for unsupervised data
+tolerance=1   # frame-tolerance for chain training
+
+unsup_lat_dir=${sup_chain_dir}/decode_${unsupervised_set_perturbed}
+if [ -z "$unsup_egs_dir" ]; then
+  unsup_egs_dir=$dir/egs_${unsupervised_set_perturbed}
+
+  if [ $stage -le 13 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
+    fi
+    mkdir -p $unsup_egs_dir
+    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the unsupervised data"
+    steps/nnet3/chain/get_egs.sh \
+      --cmd "$decode_cmd" --alignment-subsampling-factor 1 \
+      --left-tolerance $tolerance --right-tolerance $tolerance \
+      --left-context $egs_left_context --right-context $egs_right_context \
+      --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+      --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
+      --frame-subsampling-factor $frame_subsampling_factor \
+      --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
+      --lattice-prune-beam "$lattice_prune_beam" \
+      --deriv-weights-scp $sup_chain_dir/best_path_${unsupervised_set_perturbed}/weights.scp \
+      --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+      --generate-egs-scp true $unsup_egs_opts \
+      data/${unsupervised_set_perturbed}_hires $dir \
+      $unsup_lat_dir $unsup_egs_dir
+  fi
+fi
+
+comb_egs_dir=$dir/comb_egs
+if [ $stage -le 14 ]; then
+  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
+    --block-size 128 \
+    --lang2weight $supervision_weights 2 \
+    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
+  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
+fi
+
+if [ $train_stage -le -4 ]; then
+  # This is to skip stages of den-fst creation, which was already done.
+  train_stage=-4
+fi
+
+if [ $stage -le 15 ]; then
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$comb_egs_dir" \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $sup_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights true \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.chunk-width $frames_per_eg \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.chunk-left-context-initial 0 \
+    --egs.chunk-right-context-final 0 \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.num-chunk-per-minibatch "$minibatch_size" \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.max-param-change 2.0 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --cleanup.remove-egs false \
+    --feat-dir data/${supervised_set_perturbed}_hires \
+    --tree-dir $sup_tree_dir \
+    --lat-dir $sup_lat_dir \
+    --dir $dir || exit 1;
+fi
+
+test_graph_dir=$dir/graph${test_graph_affix}
+if [ $stage -le 17 ]; then
+  # Note: it might appear that this $lang directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --self-loop-scale 1.0 ${test_lang} $dir $test_graph_dir
+fi
+
+if [ $stage -le 18 ]; then
+  rm -f $dir/.error
+  for decode_set in dev test; do
+    (
+      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
+      if [ $num_jobs -gt $test_nj ]; then num_jobs=$test_nj; fi
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+        --nj $num_jobs --cmd "$decode_cmd" ${decode_iter:+--iter $decode_iter} \
+        --online-ivector-dir $ivector_root_dir/ivectors_${decode_set}_hires \
+        --frames-per-chunk $frames_per_chunk_decoding \
+        --extra-left-context $extra_left_context \
+        --extra-right-context $extra_right_context \
+        --extra-left-context-initial 0 --extra-right-context-final 0 \
+        $test_graph_dir data/${decode_set}_hires \
+        $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || touch $dir/.error
+    ) &
+  done
+  wait;
+  if [ -f $dir/.error ]; then
+    echo "$0: Decoding failed. See $dir/decode${test_graph_affix}_*/log/*"
+    exit 1
+  fi
+fi
+
+exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_smbr_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_smbr_a.sh
deleted file mode 100755
index 1806303f319..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_smbr_a.sh
+++ /dev/null
@@ -1,233 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7smbr_a
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_c
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-label_delay=5
-extra_opts="--chain.mmi-factor-schedule=1.0,1.0@0.1,0.5@0.2,0.5 --chain.smbr-factor-schedule=0.0,0.0@0.1,0.5@0.2,0.5"
-chain_smbr_extra_opts=
-smbr_leaky_hmm_coefficient=0.00001
-leaky_hmm_coefficient=0.1
-l2_regularize=0.0  # 00005
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
-lang=data/lang_chain_unk
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang_unk $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang_unk $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient $leaky_hmm_coefficient \
-    --chain.smbr-leaky-hmm-coefficient $smbr_leaky_hmm_coefficient \
-    --chain.l2-regularize $l2_regularize \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --cleanup.preserve-model-interval 10 \
-    --dir $dir --lang $lang $extra_opts || exit 1;
-fi
-
-graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_smbr_a.sh.orig b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_smbr_a.sh.orig
deleted file mode 100755
index a7505376a19..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_smbr_a.sh.orig
+++ /dev/null
@@ -1,240 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7smbr_a
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_c
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-label_delay=5
-extra_opts="--chain.mmi-factor-schedule=1.0,1.0@0.1,0.5@0.2,0.5 --chain.smbr-factor-schedule=0.0,0.0@0.1,0.5@0.2,0.5"
-chain_smbr_extra_opts=
-smbr_leaky_hmm_coefficient=0.00001
-leaky_hmm_coefficient=0.1
-<<<<<<< Updated upstream
-l2_regularize=0.0  # 00005
-=======
->>>>>>> Stashed changes
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
-lang=data/lang_chain_unk
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang_unk $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang_unk $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient $leaky_hmm_coefficient \
-    --chain.smbr-leaky-hmm-coefficient $smbr_leaky_hmm_coefficient \
-<<<<<<< Updated upstream
-    --chain.l2-regularize $l2_regularize \
-=======
-    --chain.l2-regularize 0.0 \
->>>>>>> Stashed changes
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --cleanup.preserve-model-interval 10 \
-    --dir $dir --lang $lang $extra_opts || exit 1;
-fi
-
-graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_a.sh
deleted file mode 100755
index bf1e4878c8e..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_a.sh
+++ /dev/null
@@ -1,232 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-# This is similar to _d, but uses a phone LM UNK model
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7a
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup15k
-unsupervised_set=train_unsup100_250k
-semisup_train_set=semisup15k_250k
-tree_affix=bi_i
-nnet3_affix=_semi15k_250k
-chain_affix=_semi15k_250k
-exp=exp/semisup_15k
-gmm=tri3
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=10
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires
-lang=data/lang_chain_unk
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --unsup-train-set $unsupervised_set \
-                                  --semisup-train-set $semisup_train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang_unk $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang_unk $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_b.sh
deleted file mode 100755
index 3c9ab27a353..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_b.sh
+++ /dev/null
@@ -1,232 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-# This is similar to _d, but uses a phone LM UNK model
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7b
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup15k
-unsupervised_set=train_unsup100_250k
-semisup_train_set=semisup15k_250k
-tree_affix=bi_j
-nnet3_affix=_semi15k_250k
-chain_affix=_semi15k_250k
-exp=exp/semisup_15k
-gmm=tri3
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=10
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires
-lang=data/lang_chain
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --unsup-train-set $unsupervised_set \
-                                  --semisup-train-set $semisup_train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_oracle_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_oracle_a.sh
deleted file mode 100755
index 997a17a5329..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_oracle_a.sh
+++ /dev/null
@@ -1,244 +0,0 @@
-#!/bin/bash
-set -e
-set -u
-
-# This is oracle experiment for semi-supervised training with 100 hours 
-# of supervised data and 250 hours of unsupervised data
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7a_oracle
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-supervised_set=train_sup
-unsupervised_set=train_unsup100k_250k_n10k
-base_train_set=train_oracle100k_250k_n10k
-tree_affix=bi_i
-nnet3_affix=_semi15k_250k
-chain_affix=_semi15k_250k
-exp=exp/semisup_15k
-gmm=tri3
-
-hidden_dim=1024
-cell_dim=1024
-projection_dim=256
-
-# training options
-num_epochs=4
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if [ $stage -le 6 ]; then
-  utils/combine_data.sh data/${base_train_set}_sp_hires \
-    data/${supervised_set}_sp_hires data/${unsupervised_set}_sp_hires || exit 1
-
-  utils/combine_data.sh data/${base_train_set} \
-    data/${supervised_set} data/${unsupervised_set} || exit 1
-
-  utils/data/perturb_data_dir_speed_3way.sh data/${base_train_set} \
-    data/${base_train_set}_sp || exit 1
-fi
-
-if [ $stage -le 7 ]; then
-  steps/make_mfcc.sh --nj 40 data/${base_train_set}_sp || exit 1
-  steps/compute_cmvn_stats.sh data/${base_train_set}_sp || exit 1
-  utils/fix_data_dir.sh data/${base_train_set}_sp
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${base_train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${base_train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires
-lang=data/lang_chain
-
-if true || [ $stage -le 8 ]; then
-  steps/online/nnet2/copy_data_dir.sh --utts-per-spk-max 2 \
-    data/${base_train_set}_sp_hires data/${base_train_set}_sp_max2_hires
-
-  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 30 \
-    data/${base_train_set}_sp_max2_hires $exp/nnet3${nnet3_affix}/extractor \
-    $exp/nnet3${nnet3_affix}/ivectors_${base_train_set}_sp_hires || exit 1;
-fi
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${base_train_set}_sp \
-    data/lang $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: Could not find $treedir/final.mdl"
-  exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-EOF
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_a.sh
deleted file mode 100644
index 145b4c0e178..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_a.sh
+++ /dev/null
@@ -1,507 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=40
-exp=exp/semisup_15k
-
-unsupervised_set=train_unsup100k_250k  # set this to your choice of unsupervised data
-supervised_set=train_sup15k
-semisup_train_set=semisup15k_250k   
-
-tdnn_affix=7a  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_i
-
-nnet3_affix=_semi15k_250k    # affix for nnet3 and chain dir -- relates to i-vector used
-chain_affix=_semi15k_250k
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb1a  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=5,2
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024 
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_15k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain
-unsup_decode_lang=data/lang_poco_test_ex250k
-unsup_rescore_lang=${unsup_decode_lang}_big
-unsup_decode_graph_affix=_poco_ex250k
-
-test_lang=data/lang_poco_test
-test_graph_affix=_poco
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-done
-
-decode_affix=${decode_affix}_big
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-fi
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri3
-
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${chain_affix}/tri3_${supervised_set}_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. 2 \
-    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs 4 \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_b.sh
deleted file mode 100644
index 681a46212c9..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_b.sh
+++ /dev/null
@@ -1,509 +0,0 @@
-#!/bin/bash
-
-# This script uses phone LM to model UNK.
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=40
-exp=exp/semisup_15k
-
-unsupervised_set=train_unsup100k_250k  # set this to your choice of unsupervised data
-supervised_set=train_sup15k
-semisup_train_set=semisup15k_250k
-
-tdnn_affix=7a  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_i
-
-nnet3_affix=_semi15k_250k    # affix for nnet3 and chain dir -- relates to i-vector used
-chain_affix=_semi15k_250k
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb1b  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=5,2
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024 
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_15k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_ex250k_unk
-unsup_rescore_lang=${unsup_decode_lang}_big
-unsup_decode_graph_affix=_poco_ex250k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-done
-
-decode_affix=${decode_affix}_big
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-fi
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri3
-
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay 
-  output name=output-1 input=output.affine@$label_delay
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay 
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${chain_affix}/tri3_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --block-size 64 \
-    --lang2weight $supervision_weights 2 \
-    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs 4 \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_c.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_c.sh
deleted file mode 100644
index 01c0191be83..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_semisupervised_conf_c.sh
+++ /dev/null
@@ -1,507 +0,0 @@
-#!/bin/bash
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=40
-exp=exp/semisup_15k
-
-unsupervised_set=train_unsup100k_250k  # set this to your choice of unsupervised data
-supervised_set=train_sup15k
-semisup_train_set=semisup15k_250k   
-
-tdnn_affix=7b  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_j
-
-nnet3_affix=_semi15k_250k    # affix for nnet3 and chain dir -- relates to i-vector used
-chain_affix=_semi15k_250k
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb1c  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=5,2
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024 
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_15k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain
-unsup_decode_lang=data/lang_poco_test_ex250k
-unsup_rescore_lang=${unsup_decode_lang}_big
-unsup_decode_graph_affix=_poco_ex250k
-
-test_lang=data/lang_poco_test
-test_graph_affix=_poco
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-done
-
-decode_affix=${decode_affix}_big
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-fi
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/tri3
-
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${chain_affix}/tri3_${supervised_set}_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. 2 \
-    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs 4 \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_smbr_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_smbr_a.sh
deleted file mode 100755
index aff735560e0..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_15k_smbr_a.sh
+++ /dev/null
@@ -1,232 +0,0 @@
-#!/bin/bash
-set -e
-
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-# This is similar to _d, but uses a phone LM UNK model
-
-# configs for 'chain'
-stage=0
-tdnn_affix=7smbr_a
-train_stage=-10
-get_egs_stage=-10
-decode_iter=
-train_set=train_sup15k
-unsupervised_set=train_unsup100_250k
-semisup_train_set=semisup15k_250k
-tree_affix=bi_i
-nnet3_affix=_semi15k_250k
-chain_affix=_semi15k_250k
-exp=exp/semisup_15k
-gmm=tri3
-hidden_dim=512
-cell_dim=512
-projection_dim=128
-
-# training options
-num_epochs=10
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-label_delay=5
-extra_opts="--chain.mmi-factor-schedule=1.0,1.0@0.1,0.5@0.2,0.5 --chain.smbr-factor-schedule=0.0,0.0@0.1,0.5@0.2,0.5"
-chain_smbr_extra_opts=
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-remove_egs=false
-common_egs_dir=
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. ./cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires
-lang=data/lang_chain_unk
-
-# The iVector-extraction and feature-dumping parts are the same as the standard
-# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
-# run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
-                                  --speed-perturb true \
-                                  --train-set $train_set \
-                                  --unsup-train-set $unsupervised_set \
-                                  --semisup-train-set $semisup_train_set \
-                                  --nnet3-affix "$nnet3_affix" || exit 1
-
-if [ $stage -le 9 ]; then
-  # Get the alignments as lattices (gives the chain training more freedom).
-  # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
-    --generate-ali-from-lats true data/${train_set}_sp \
-    data/lang_unk $gmm_dir $lat_dir || exit 1;
-  rm $lat_dir/fsts.*.gz # save space
-fi
-
-if [ $stage -le 10 ]; then
-  # Create a version of the lang/ directory that has one state per phone in the
-  # topo file. [note, it really has two states.. the first one is only repeated
-  # once, the second one has zero or more repeats.]
-  rm -rf $lang
-  cp -r data/lang_unk $lang
-  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
-  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
-  # Use our special topology... note that later on may have to tune this
-  # topology.
-  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
-fi
-
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
-fi
-
-if [ $stage -le 12 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-if [ $stage -le 13 ]; then
-  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
-    utils/create_split_dir.pl \
-     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
-  fi
-
-  mkdir -p $dir/egs
-  touch $dir/egs/.nodelete # keep egs around when that run dies.
-
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$common_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $train_ivector_dir \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.0 \
-    --chain.apply-deriv-weights false \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
-    --egs.chunk-width 160,140,110,80 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch 64,32 \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs $num_epochs \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs $remove_egs \
-    --feat-dir $train_data_dir \
-    --tree-dir $treedir \
-    --lat-dir $lat_dir \
-    --chain.smbr-extra-opts="$chain_smbr_extra_opts" \
-    --dir $dir --lang $lang $extra_opts || exit 1;
-fi
-
-graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
-fi
-
-decode_suff=
-if [ $stage -le 15 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    iter_opts=" --iter $decode_iter "
-  fi
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
-          --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
-      ) &
-  done
-fi
-wait;
-exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1a.sh
similarity index 73%
rename from egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_a.sh
rename to egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1a.sh
index 6bafc30f3aa..74d7fdb76e9 100755
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_a.sh
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1a.sh
@@ -1,35 +1,44 @@
 #!/bin/bash
+
+# Copyright 2017-18  Vimal Manohar
+# Apache 2.0
+
 set -e
+set -o pipefail
 
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-# This is similar to _d, but uses a phone LM UNK model
+# This is fisher chain recipe for training a model on a subset of around
+# 15-50 hours of supervised data.
+# This system uses phone LM to model UNK.
+# local/semisup/run_50k.sh and local/semisup/run_100k.sh show how to call this.
 
 # configs for 'chain'
 stage=0
-tdnn_affix=7a
 train_stage=-10
 get_egs_stage=-10
-decode_iter=
+exp_root=exp/semisup_50k
+
+nj=30
+tdnn_affix=_1a
 train_set=train_sup50k
-unsupervised_set=train_unsup100_250k
-semisup_train_set=semisup50k_250k
-tree_affix=bi_i
-nnet3_affix=_semi50k_250k
-chain_affix=_semi50k_250k
-exp=exp/semisup_50k
-gmm=tri4a
+ivector_train_set=   # dataset for training i-vector extractor
+
+nnet3_affix=  # affix for nnet3 dir -- relates to i-vector used
+chain_affix=  # affix for chain dir
+tree_affix=bi_a
+gmm=tri4a  # Expect GMM model in $exp/$gmm for alignment
+
+# Neural network opts
 hidden_dim=1024
 cell_dim=1024
 projection_dim=256
 
 # training options
-num_epochs=10
+num_epochs=4
 minibatch_size=64,32
 chunk_left_context=40
 chunk_right_context=0
 dropout_schedule='0,0@0.20,0.3@0.50,0'
 xent_regularize=0.025
-self_repair_scale=0.00001
 label_delay=5
 
 # decode options
@@ -37,15 +46,16 @@ extra_left_context=50
 extra_right_context=0
 
 remove_egs=false
-common_egs_dir=
+common_egs_dir=   # if provided, will skip egs generation
+common_treedir=   # if provided, will skip the tree building stage
 
-# End configuration section.
+decode_iter=
 
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
 . ./cmd.sh
-. ./path.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
 . ./utils/parse_options.sh
 
 if ! cuda-compiled; then
@@ -56,35 +66,48 @@ where "nvcc" is installed.
 EOF
 fi
 
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
+gmm_dir=$exp_root/$gmm   # used to get training lattices (for chain supervision)
+treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+lat_dir=$exp_root/chain${chain_affix}/${gmm}_${train_set}_sp_unk_lats  # training lattices directory
+dir=$exp_root/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
 train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires
+train_ivector_dir=$exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
 lang=data/lang_chain_unk
 
 # The iVector-extraction and feature-dumping parts are the same as the standard
 # nnet3 setup, and you can skip them by setting "--stage 8" if you have already
 # run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
+local/nnet3/run_ivector_common.sh --stage $stage --exp-root $exp_root \
                                   --speed-perturb true \
                                   --train-set $train_set \
-                                  --unsup-train-set $unsupervised_set \
-                                  --semisup-train-set $semisup_train_set \
+                                  --ivector-train-set "$ivector_train_set" \
                                   --nnet3-affix "$nnet3_affix" || exit 1
 
-if [ $stage -le 9 ]; then
+if [ "$train_set" != "$ivector_train_set" ]; then
+  if [ $stage -le 9 ]; then
+    # We extract iVectors on all the ${train_set} data, which will be what we
+    # train the system on.
+    # having a larger number of speakers is helpful for generalization, and to
+    # handle per-utterance decoding well (iVector starts at zero).
+    utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+      data/${train_set}_sp_hires data/${train_set}_sp_max2_hires
+
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+      data/${train_set}_sp_max2_hires $exp_root/nnet3${nnet3_affix}/extractor \
+      $exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires || exit 1;
+  fi
+fi
+
+if [ $stage -le 10 ]; then
   # Get the alignments as lattices (gives the chain training more freedom).
   # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" \
     --generate-ali-from-lats true data/${train_set}_sp \
     data/lang_unk $gmm_dir $lat_dir || exit 1;
   rm $lat_dir/fsts.*.gz # save space
 fi
 
-if [ $stage -le 10 ]; then
+if [ $stage -le 11 ]; then
   # Create a version of the lang/ directory that has one state per phone in the
   # topo file. [note, it really has two states.. the first one is only repeated
   # once, the second one has zero or more repeats.]
@@ -97,15 +120,18 @@ if [ $stage -le 10 ]; then
   steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
 fi
 
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+if [ -z "$common_treedir" ]; then
+  if [ $stage -le 12 ]; then
+    # Build a tree using our new topology.
+    steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+        --context-opts "--context-width=2 --central-position=1" \
+        --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+  fi
+else
+  treedir=$common_treedir
 fi
 
-if [ $stage -le 12 ]; then
+if [ $stage -le 13 ]; then
   echo "$0: creating neural net configs using the xconfig parser";
 
   num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
@@ -137,7 +163,7 @@ if [ $stage -le 12 ]; then
   fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
 
   ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
+  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
 
   # adding the layers for xent branch
   # This block prints the configs for a separate output that will be
@@ -149,12 +175,12 @@ if [ $stage -le 12 ]; then
   # constant; and the 0.5 was tuned so as to make the relative progress
   # similar in the xent and regular final layers.
   output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
 
+EOF
   steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
 fi
 
-if [ $stage -le 13 ]; then
+if [ $stage -le 14 ]; then
   if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
     utils/create_split_dir.pl \
      /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
@@ -164,6 +190,7 @@ if [ $stage -le 13 ]; then
   touch $dir/egs/.nodelete # keep egs around when that run dies.
 
   steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$common_egs_dir" \
     --cmd "$decode_cmd" \
     --feat.online-ivector-dir $train_ivector_dir \
     --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
@@ -191,7 +218,6 @@ if [ $stage -le 13 ]; then
     --egs.chunk-right-context $chunk_right_context \
     --egs.chunk-left-context-initial 0 \
     --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
     --cleanup.remove-egs $remove_egs \
     --feat-dir $train_data_dir \
     --tree-dir $treedir \
@@ -200,15 +226,15 @@ if [ $stage -le 13 ]; then
 fi
 
 graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
+if [ $stage -le 15 ]; then
   # Note: it might appear that this $lang directory is mismatched, and it is as
   # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
   # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test_poco_unk $dir $graph_dir
 fi
 
 decode_suff=
-if [ $stage -le 15 ]; then
+if [ $stage -le 16 ]; then
   iter_opts=
   if [ ! -z $decode_iter ]; then
     iter_opts=" --iter $decode_iter "
@@ -223,8 +249,8 @@ if [ $stage -le 15 ]; then
           --extra-left-context-initial 0 \
           --extra-right-context-final 0 \
           --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
+          --online-ivector-dir $exp_root/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
+          $graph_dir data/${decode_set}_hires $dir/decode_poco_unk_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
       ) &
   done
 fi
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_g.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1b.sh
similarity index 74%
rename from egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_g.sh
rename to egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1b.sh
index ff4e8d55efc..2632cb45fee 100755
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_g.sh
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1b.sh
@@ -1,21 +1,33 @@
 #!/bin/bash
+
+# Copyright 2017-18  Vimal Manohar
+# Apache 2.0
+
 set -e
+set -o pipefail
 
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
+# This is fisher chain recipe for training a model on a subset of around
+# 100-300 hours of supervised data.
+# This system uses phone LM to model UNK.
+# local/semisup/run_50k.sh and local/semisup/run_100k.sh show how to call this.
 
 # configs for 'chain'
 stage=0
-tdnn_affix=7g
 train_stage=-10
 get_egs_stage=-10
-decode_iter=
+exp_root=exp/semisup_100k
+
+nj=30
+tdnn_affix=_1b
 train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_e
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
+ivector_train_set=   # dataset for training i-vector extractor
+
+nnet3_affix=  # affix for nnet3 dir -- relates to i-vector used
+chain_affix=  # affix for chain dir
+tree_affix=bi_a
+gmm=tri4a  # Expect GMM model in $exp/$gmm for alignment
+
+# Neural network opts
 hidden_dim=1024
 cell_dim=1024
 projection_dim=256
@@ -34,13 +46,16 @@ extra_left_context=50
 extra_right_context=0
 
 remove_egs=false
-common_egs_dir=
+common_egs_dir=   # if provided, will skip egs generation
+common_treedir=   # if provided, will skip the tree building stage
+
+decode_iter=
 
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
 . ./cmd.sh
-. ./path.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
 . ./utils/parse_options.sh
 
 if ! cuda-compiled; then
@@ -51,33 +66,48 @@ where "nvcc" is installed.
 EOF
 fi
 
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_unk_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
+gmm_dir=$exp_root/$gmm   # used to get training lattices (for chain supervision)
+treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+lat_dir=$exp_root/chain${chain_affix}/${gmm}_${train_set}_sp_unk_lats  # training lattices directory
+dir=$exp_root/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
 train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
+train_ivector_dir=$exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
 lang=data/lang_chain_unk
 
 # The iVector-extraction and feature-dumping parts are the same as the standard
 # nnet3 setup, and you can skip them by setting "--stage 8" if you have already
 # run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
+local/nnet3/run_ivector_common.sh --stage $stage --exp-root $exp_root \
                                   --speed-perturb true \
                                   --train-set $train_set \
+                                  --ivector-train-set "$ivector_train_set" \
                                   --nnet3-affix "$nnet3_affix" || exit 1
 
-if [ $stage -le 9 ]; then
+if [ "$train_set" != "$ivector_train_set" ]; then
+  if [ $stage -le 9 ]; then
+    # We extract iVectors on all the ${train_set} data, which will be what we
+    # train the system on.
+    # having a larger number of speakers is helpful for generalization, and to
+    # handle per-utterance decoding well (iVector starts at zero).
+    utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+      data/${train_set}_sp_hires data/${train_set}_sp_max2_hires
+
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+      data/${train_set}_sp_max2_hires $exp_root/nnet3${nnet3_affix}/extractor \
+      $exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires || exit 1;
+  fi
+fi
+
+if [ $stage -le 10 ]; then
   # Get the alignments as lattices (gives the chain training more freedom).
   # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" \
     --generate-ali-from-lats true data/${train_set}_sp \
     data/lang_unk $gmm_dir $lat_dir || exit 1;
   rm $lat_dir/fsts.*.gz # save space
 fi
 
-if [ $stage -le 10 ]; then
+if [ $stage -le 11 ]; then
   # Create a version of the lang/ directory that has one state per phone in the
   # topo file. [note, it really has two states.. the first one is only repeated
   # once, the second one has zero or more repeats.]
@@ -90,15 +120,18 @@ if [ $stage -le 10 ]; then
   steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
 fi
 
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+if [ -z "$common_treedir" ]; then
+  if [ $stage -le 12 ]; then
+    # Build a tree using our new topology.
+    steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+        --context-opts "--context-width=2 --central-position=1" \
+        --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+  fi
+else
+  treedir=$common_treedir
 fi
 
-if [ $stage -le 12 ]; then
+if [ $stage -le 13 ]; then
   echo "$0: creating neural net configs using the xconfig parser";
 
   num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
@@ -150,7 +183,7 @@ EOF
   steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
 fi
 
-if [ $stage -le 13 ]; then
+if [ $stage -le 14 ]; then
   if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
     utils/create_split_dir.pl \
      /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
@@ -160,6 +193,7 @@ if [ $stage -le 13 ]; then
   touch $dir/egs/.nodelete # keep egs around when that run dies.
 
   steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$common_egs_dir" \
     --cmd "$decode_cmd" \
     --feat.online-ivector-dir $train_ivector_dir \
     --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
@@ -181,13 +215,12 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.momentum 0.0 \
     --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
-    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true --constrained false" \
+    --egs.opts "--frames-overlap-per-eg 0 --generate-egs-scp true" \
     --egs.chunk-width 160,140,110,80 \
     --egs.chunk-left-context $chunk_left_context \
     --egs.chunk-right-context $chunk_right_context \
     --egs.chunk-left-context-initial 0 \
     --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
     --cleanup.remove-egs $remove_egs \
     --feat-dir $train_data_dir \
     --tree-dir $treedir \
@@ -196,15 +229,15 @@ if [ $stage -le 13 ]; then
 fi
 
 graph_dir=$dir/graph_poco_unk
-if [ $stage -le 14 ]; then
+if [ $stage -le 15 ]; then
   # Note: it might appear that this $lang directory is mismatched, and it is as
   # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
   # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test_unk $dir $graph_dir
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test_poco_unk $dir $graph_dir
 fi
 
 decode_suff=
-if [ $stage -le 15 ]; then
+if [ $stage -le 16 ]; then
   iter_opts=
   if [ ! -z $decode_iter ]; then
     iter_opts=" --iter $decode_iter "
@@ -219,8 +252,8 @@ if [ $stage -le 15 ]; then
           --extra-left-context-initial 0 \
           --extra-right-context-final 0 \
           --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
+          --online-ivector-dir $exp_root/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
+          $graph_dir data/${decode_set}_hires $dir/decode_poco_unk_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
       ) &
   done
 fi
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1c.sh
similarity index 73%
rename from egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_b.sh
rename to egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1c.sh
index aa0387cc1d4..34a0ace9878 100755
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_b.sh
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1c.sh
@@ -1,35 +1,44 @@
 #!/bin/bash
+
+# Copyright 2017-18  Vimal Manohar
+# Apache 2.0
+
 set -e
+set -o pipefail
 
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
-# This is similar to _d, but uses a phone LM UNK model
+# This is fisher chain recipe for training a model on a subset of around
+# 15-50 hours of supervised data.
+# This system uses phone LM to model UNK.
+# local/semisup/run_50k.sh and local/semisup/run_100k.sh show how to call this.
 
 # configs for 'chain'
 stage=0
-tdnn_affix=7b
 train_stage=-10
 get_egs_stage=-10
-decode_iter=
+exp_root=exp/semisup_50k
+
+nj=30
+tdnn_affix=_1c
 train_set=train_sup50k
-unsupervised_set=train_unsup100_250k
-semisup_train_set=semisup50k_250k
-tree_affix=bi_j
-nnet3_affix=_semi50k_250k
-chain_affix=_semi50k_250k
-exp=exp/semisup_50k
-gmm=tri4a
+ivector_train_set=   # dataset for training i-vector extractor
+
+nnet3_affix=  # affix for nnet3 dir -- relates to i-vector used
+chain_affix=  # affix for chain dir
+tree_affix=bi_c
+gmm=tri4a  # Expect GMM model in $exp/$gmm for alignment
+
+# Neural network opts
 hidden_dim=1024
 cell_dim=1024
 projection_dim=256
 
 # training options
-num_epochs=10
+num_epochs=4
 minibatch_size=64,32
 chunk_left_context=40
 chunk_right_context=0
 dropout_schedule='0,0@0.20,0.3@0.50,0'
 xent_regularize=0.025
-self_repair_scale=0.00001
 label_delay=5
 
 # decode options
@@ -37,15 +46,16 @@ extra_left_context=50
 extra_right_context=0
 
 remove_egs=false
-common_egs_dir=
+common_egs_dir=   # if provided, will skip egs generation
+common_treedir=   # if provided, will skip the tree building stage
 
-# End configuration section.
+decode_iter=
 
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
 . ./cmd.sh
-. ./path.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
 . ./utils/parse_options.sh
 
 if ! cuda-compiled; then
@@ -56,35 +66,48 @@ where "nvcc" is installed.
 EOF
 fi
 
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
+gmm_dir=$exp_root/$gmm   # used to get training lattices (for chain supervision)
+treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+lat_dir=$exp_root/chain${chain_affix}/${gmm}_${train_set}_sp_lats  # training lattices directory
+dir=$exp_root/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
 train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires
+train_ivector_dir=$exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
 lang=data/lang_chain
 
 # The iVector-extraction and feature-dumping parts are the same as the standard
 # nnet3 setup, and you can skip them by setting "--stage 8" if you have already
 # run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
+local/nnet3/run_ivector_common.sh --stage $stage --exp-root $exp_root \
                                   --speed-perturb true \
                                   --train-set $train_set \
-                                  --unsup-train-set $unsupervised_set \
-                                  --semisup-train-set $semisup_train_set \
+                                  --ivector-train-set "$ivector_train_set" \
                                   --nnet3-affix "$nnet3_affix" || exit 1
 
-if [ $stage -le 9 ]; then
+if [ "$train_set" != "$ivector_train_set" ]; then
+  if [ $stage -le 9 ]; then
+    # We extract iVectors on all the ${train_set} data, which will be what we
+    # train the system on.
+    # having a larger number of speakers is helpful for generalization, and to
+    # handle per-utterance decoding well (iVector starts at zero).
+    utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+      data/${train_set}_sp_hires data/${train_set}_sp_max2_hires
+
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+      data/${train_set}_sp_max2_hires $exp_root/nnet3${nnet3_affix}/extractor \
+      $exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires || exit 1;
+  fi
+fi
+
+if [ $stage -le 10 ]; then
   # Get the alignments as lattices (gives the chain training more freedom).
   # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" \
     --generate-ali-from-lats true data/${train_set}_sp \
     data/lang $gmm_dir $lat_dir || exit 1;
   rm $lat_dir/fsts.*.gz # save space
 fi
 
-if [ $stage -le 10 ]; then
+if [ $stage -le 11 ]; then
   # Create a version of the lang/ directory that has one state per phone in the
   # topo file. [note, it really has two states.. the first one is only repeated
   # once, the second one has zero or more repeats.]
@@ -97,15 +120,18 @@ if [ $stage -le 10 ]; then
   steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
 fi
 
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+if [ -z "$common_treedir" ]; then
+  if [ $stage -le 12 ]; then
+    # Build a tree using our new topology.
+    steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+        --context-opts "--context-width=2 --central-position=1" \
+        --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+  fi
+else
+  treedir=$common_treedir
 fi
 
-if [ $stage -le 12 ]; then
+if [ $stage -le 13 ]; then
   echo "$0: creating neural net configs using the xconfig parser";
 
   num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
@@ -137,7 +163,7 @@ if [ $stage -le 12 ]; then
   fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
 
   ## adding the layers for chain branch
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
+  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
 
   # adding the layers for xent branch
   # This block prints the configs for a separate output that will be
@@ -149,12 +175,12 @@ if [ $stage -le 12 ]; then
   # constant; and the 0.5 was tuned so as to make the relative progress
   # similar in the xent and regular final layers.
   output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-EOF
 
+EOF
   steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
 fi
 
-if [ $stage -le 13 ]; then
+if [ $stage -le 14 ]; then
   if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
     utils/create_split_dir.pl \
      /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
@@ -164,6 +190,7 @@ if [ $stage -le 13 ]; then
   touch $dir/egs/.nodelete # keep egs around when that run dies.
 
   steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$common_egs_dir" \
     --cmd "$decode_cmd" \
     --feat.online-ivector-dir $train_ivector_dir \
     --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
@@ -191,7 +218,6 @@ if [ $stage -le 13 ]; then
     --egs.chunk-right-context $chunk_right_context \
     --egs.chunk-left-context-initial 0 \
     --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
     --cleanup.remove-egs $remove_egs \
     --feat-dir $train_data_dir \
     --tree-dir $treedir \
@@ -200,15 +226,15 @@ if [ $stage -le 13 ]; then
 fi
 
 graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
+if [ $stage -le 15 ]; then
   # Note: it might appear that this $lang directory is mismatched, and it is as
   # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
   # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test_poco $dir $graph_dir
 fi
 
 decode_suff=
-if [ $stage -le 15 ]; then
+if [ $stage -le 16 ]; then
   iter_opts=
   if [ ! -z $decode_iter ]; then
     iter_opts=" --iter $decode_iter "
@@ -223,8 +249,8 @@ if [ $stage -le 15 ]; then
           --extra-left-context-initial 0 \
           --extra-right-context-final 0 \
           --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
+          --online-ivector-dir $exp_root/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
+          $graph_dir data/${decode_set}_hires $dir/decode_poco_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
       ) &
   done
 fi
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_e.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1d.sh
similarity index 74%
rename from egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_e.sh
rename to egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1d.sh
index 05fe3a017e3..67a0026b076 100755
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_100k_e.sh
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_1d.sh
@@ -1,21 +1,33 @@
 #!/bin/bash
+
+# Copyright 2017-18  Vimal Manohar
+# Apache 2.0
+
 set -e
+set -o pipefail
 
-# This is fisher chain recipe for training a model on a subset of around 15 hours.
+# This is fisher chain recipe for training a model on a subset of around
+# 100-300 hours of supervised data.
+# This system uses phone LM to model UNK.
+# local/semisup/run_50k.sh and local/semisup/run_100k.sh show how to call this.
 
 # configs for 'chain'
 stage=0
-tdnn_affix=7e
 train_stage=-10
 get_egs_stage=-10
-decode_iter=
+exp_root=exp/semisup_100k
+
+nj=30
+tdnn_affix=_1d
 train_set=train_sup
-ivector_train_set=train_sup
-tree_affix=bi_e
-nnet3_affix=
-chain_affix=
-exp=exp/semisup_100k
-gmm=tri4a
+ivector_train_set=   # dataset for training i-vector extractor
+
+nnet3_affix=  # affix for nnet3 dir -- relates to i-vector used
+chain_affix=  # affix for chain dir
+tree_affix=bi_c
+gmm=tri4a  # Expect GMM model in $exp/$gmm for alignment
+
+# Neural network opts
 hidden_dim=1024
 cell_dim=1024
 projection_dim=256
@@ -34,13 +46,16 @@ extra_left_context=50
 extra_right_context=0
 
 remove_egs=false
-common_egs_dir=
+common_egs_dir=   # if provided, will skip egs generation
+common_treedir=   # if provided, will skip the tree building stage
+
+decode_iter=
 
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
 . ./cmd.sh
-. ./path.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
 . ./utils/parse_options.sh
 
 if ! cuda-compiled; then
@@ -51,33 +66,48 @@ where "nvcc" is installed.
 EOF
 fi
 
-gmm_dir=$exp/$gmm   # used to get training lattices (for chain supervision)
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-lat_dir=$exp/chain${chain_affix}/$(basename $gmm_dir)_${train_set}_sp_lats  # training lattices directory
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
+gmm_dir=$exp_root/$gmm   # used to get training lattices (for chain supervision)
+treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+lat_dir=$exp_root/chain${chain_affix}/${gmm}_${train_set}_sp_lats  # training lattices directory
+dir=$exp_root/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
 train_data_dir=data/${train_set}_sp_hires
-train_ivector_dir=$exp/nnet3${nnet3_affix}/ivectors_${ivector_train_set}_sp_hires
+train_ivector_dir=$exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
 lang=data/lang_chain
 
 # The iVector-extraction and feature-dumping parts are the same as the standard
 # nnet3 setup, and you can skip them by setting "--stage 8" if you have already
 # run those things.
-
-local/semisup/nnet3/run_ivector_common.sh --stage $stage --exp $exp \
+local/nnet3/run_ivector_common.sh --stage $stage --exp-root $exp_root \
                                   --speed-perturb true \
                                   --train-set $train_set \
+                                  --ivector-train-set "$ivector_train_set" \
                                   --nnet3-affix "$nnet3_affix" || exit 1
 
-if [ $stage -le 9 ]; then
+if [ "$train_set" != "$ivector_train_set" ]; then
+  if [ $stage -le 9 ]; then
+    # We extract iVectors on all the ${train_set} data, which will be what we
+    # train the system on.
+    # having a larger number of speakers is helpful for generalization, and to
+    # handle per-utterance decoding well (iVector starts at zero).
+    utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+      data/${train_set}_sp_hires data/${train_set}_sp_max2_hires
+
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+      data/${train_set}_sp_max2_hires $exp_root/nnet3${nnet3_affix}/extractor \
+      $exp_root/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires || exit 1;
+  fi
+fi
+
+if [ $stage -le 10 ]; then
   # Get the alignments as lattices (gives the chain training more freedom).
   # use the same num-jobs as the alignments
-  steps/align_fmllr_lats.sh --nj 30 --cmd "$train_cmd" \
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" \
     --generate-ali-from-lats true data/${train_set}_sp \
     data/lang $gmm_dir $lat_dir || exit 1;
   rm $lat_dir/fsts.*.gz # save space
 fi
 
-if [ $stage -le 10 ]; then
+if [ $stage -le 11 ]; then
   # Create a version of the lang/ directory that has one state per phone in the
   # topo file. [note, it really has two states.. the first one is only repeated
   # once, the second one has zero or more repeats.]
@@ -90,15 +120,18 @@ if [ $stage -le 10 ]; then
   steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
 fi
 
-if [ $stage -le 11 ]; then
-  # Build a tree using our new topology.
-  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
-      --leftmost-questions-truncate -1 \
-      --context-opts "--context-width=2 --central-position=1" \
-      --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+if [ -z "$common_treedir" ]; then
+  if [ $stage -le 12 ]; then
+    # Build a tree using our new topology.
+    steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+        --context-opts "--context-width=2 --central-position=1" \
+        --cmd "$train_cmd" 7000 data/${train_set}_sp $lang $lat_dir $treedir || exit 1
+  fi
+else
+  treedir=$common_treedir
 fi
 
-if [ $stage -le 12 ]; then
+if [ $stage -le 13 ]; then
   echo "$0: creating neural net configs using the xconfig parser";
 
   num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
@@ -150,7 +183,7 @@ EOF
   steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
 fi
 
-if [ $stage -le 13 ]; then
+if [ $stage -le 14 ]; then
   if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
     utils/create_split_dir.pl \
      /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
@@ -160,6 +193,7 @@ if [ $stage -le 13 ]; then
   touch $dir/egs/.nodelete # keep egs around when that run dies.
 
   steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$common_egs_dir" \
     --cmd "$decode_cmd" \
     --feat.online-ivector-dir $train_ivector_dir \
     --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
@@ -187,7 +221,6 @@ if [ $stage -le 13 ]; then
     --egs.chunk-right-context $chunk_right_context \
     --egs.chunk-left-context-initial 0 \
     --egs.chunk-right-context-final 0 \
-    --egs.dir "$common_egs_dir" \
     --cleanup.remove-egs $remove_egs \
     --feat-dir $train_data_dir \
     --tree-dir $treedir \
@@ -196,15 +229,15 @@ if [ $stage -le 13 ]; then
 fi
 
 graph_dir=$dir/graph_poco
-if [ $stage -le 14 ]; then
+if [ $stage -le 15 ]; then
   # Note: it might appear that this $lang directory is mismatched, and it is as
   # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
   # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_poco_test $dir $graph_dir
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test_poco $dir $graph_dir
 fi
 
 decode_suff=
-if [ $stage -le 15 ]; then
+if [ $stage -le 16 ]; then
   iter_opts=
   if [ ! -z $decode_iter ]; then
     iter_opts=" --iter $decode_iter "
@@ -219,8 +252,8 @@ if [ $stage -le 15 ]; then
           --extra-left-context-initial 0 \
           --extra-right-context-final 0 \
           --frames-per-chunk 160 \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
+          --online-ivector-dir $exp_root/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
+          $graph_dir data/${decode_set}_hires $dir/decode_poco_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
       ) &
   done
 fi
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_1a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_1a.sh
new file mode 100644
index 00000000000..c149a856515
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_1a.sh
@@ -0,0 +1,468 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script is semi-supervised recipe with around 50 hours of supervised data
+# and 250 hours unsupervised data with naive splitting.
+# Based on "Semi-Supervised Training of Acoustic Models using Lattice-Free MMI",
+# Vimal Manohar, Hossein Hadian, Daniel Povey, Sanjeev Khudanpur, ICASSP 2018
+# http://www.danielpovey.com/files/2018_icassp_semisupervised_mmi.pdf
+# local/semisup/run_50k.sh shows how to call this.
+
+# We use the combined data for i-vector extractor training.
+# We use 4-gram LM trained on 1250 hours of data excluding the 250 hours
+# unsupervised data to create LM for decoding. Rescoring is done with
+# a larger 4-gram LM.
+# This differs from the case in run_tdnn_100k_semisupervised.sh.
+
+# This script uses phone LM to model UNK.
+# This script uses the same tree as that for the seed model.
+# See the comments in the script about how to change these.
+
+# Unsupervised set: train_unsup100k_250k (250 hour subset of Fisher excluding 100 hours for supervised)
+# unsup_frames_per_eg=150
+# Deriv weights: Lattice posterior of best path pdf
+# Unsupervised weight: 1.0
+# Weights for phone LM (supervised, unsupervised): 3,2
+# LM for decoding unsupervised data: 4gram
+# Supervision: Naive split lattices
+
+set -u -e -o pipefail
+
+stage=0
+train_stage=-100
+nj=80
+test_nj=50
+
+# The following 3 options decide the output directory for semi-supervised
+# chain system
+# dir=${exp_root}/chain${chain_affix}/tdnn_lstm${tdnn_affix}
+
+exp_root=exp/semisup_50k
+chain_affix=_semi50k_100k_250k    # affix for chain dir
+                                  # 50 hour subset out of 100 hours of supervised data
+                                  # 250 hour subset out of (1500-100=1400) hours of unsupervised data 
+tdnn_affix=_semisup_1a
+
+# Datasets -- Expects data/$supervised_set and data/$unsupervised_set to be
+# present
+supervised_set=train_sup50k
+unsupervised_set=train_unsup100k_250k
+
+# Input seed system
+sup_chain_dir=exp/semisup_50k/chain_semi50k_100k_250k/tdnn_lstm_1a_sp  # supervised chain system
+sup_lat_dir=exp/semisup_50k/chain_semi50k_100k_250k/tri4a_train_sup50k_unk_lats  # lattices for supervised set
+sup_tree_dir=exp/semisup_50k/chain_semi50k_100k_250k/tree_bi_a  # tree directory for supervised chain system
+ivector_root_dir=exp/semisup_50k/nnet3_semi50k_100k_250k  # i-vector extractor root directory
+
+# Semi-supervised options
+supervision_weights=1.0,1.0   # Weights for supervised, unsupervised data egs.
+                              # Can be used to scale down the effect of unsupervised data
+                              # by using a smaller scale for it e.g. 1.0,0.3
+lm_weights=3,2  # Weights on phone counts from supervised, unsupervised data for denominator FST creation
+
+sup_egs_dir=   # Supply this to skip supervised egs creation
+unsup_egs_dir=  # Supply this to skip unsupervised egs creation
+unsup_egs_opts=  # Extra options to pass to unsupervised egs creation
+
+# Neural network opts
+hidden_dim=1024
+cell_dim=1024
+projection_dim=256
+
+# training options
+num_epochs=2
+minibatch_size=64,32
+chunk_left_context=40
+chunk_right_context=0
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=5
+
+# decode options
+extra_left_context=50
+extra_right_context=0
+frames_per_chunk_decoding=160
+
+decode_iter=  # Iteration to decode with
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
+. ./utils/parse_options.sh
+
+# The following can be replaced with the versions that do not model
+# UNK using phone LM. $sup_lat_dir should also ideally be changed.
+unsup_decode_lang=data/lang_test_poco_ex250k_unk
+unsup_decode_graph_affix=_poco_ex250k_unk
+test_lang=data/lang_test_poco_unk
+test_graph_affix=_poco_unk
+
+unsup_rescore_lang=${unsup_decode_lang}_big
+
+dir=$exp_root/chain${chain_affix}/tdnn_lst${tdnn_affix}
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+supervised_set_perturbed=${supervised_set}_sp
+unsupervised_set_perturbed=${unsupervised_set}_sp
+
+sup_ivector_dir=$ivector_root_dir/ivectors_${supervised_set_perturbed}_hires
+
+graphdir=$sup_chain_dir/graph${unsup_decode_graph_affix}
+
+for f in data/${supervised_set_perturbed}/feats.scp \
+  data/${supervised_set_perturbed}_hires/feats.scp \
+  $ivector_root_dir/extractor/final.ie $sup_ivector_dir/ivector_online.scp \
+  $sup_lat_dir/lat.1.gz $sup_tree_dir/ali.1.gz \
+  $unsup_decode_lang/G.fst; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find file $f"
+    exit 1
+  fi
+done
+
+if [ ! -f $graphdir/HCLG.fst ]; then
+  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $sup_chain_dir $graphdir
+fi
+
+if [ $stage -le 2 ]; then
+  utils/data/perturb_data_dir_speed_3way.sh data/${unsupervised_set} \
+    data/${unsupervised_set_perturbed}_hires
+  utils/data/perturb_data_dir_volume.sh \
+    data/${unsupervised_set_perturbed}_hires
+
+  steps/make_mfcc.sh --cmd "$train_cmd" --nj $nj \
+    --mfcc-config conf/mfcc_hires.conf \
+    data/${unsupervised_set_perturbed}_hires
+  steps/compute_cmvn_stats.sh data/${unsupervised_set_perturbed}_hires
+  utils/fix_data_dir.sh data/${unsupervised_set_perturbed}_hires
+fi
+
+# Extract i-vectors for the unsupervised data
+if [ $stage -le 3 ]; then
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+    data/${unsupervised_set_perturbed}_hires data/${unsupervised_set_perturbed}_max2_hires
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+    data/${unsupervised_set_perturbed}_max2_hires $ivector_root_dir/extractor \
+    $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires || exit 1
+fi
+
+# Decode unsupervised data and write lattices in non-compact
+# undeterminized format
+# Set --skip-scoring to false in order to score the unsupervised data
+if [ $stage -le 4 ]; then
+  echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
+  steps/nnet3/decode_semisup.sh --num-threads 4 --nj $nj --cmd "$decode_cmd" \
+            --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
+            --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+            --frames-per-chunk 160 \
+            --extra-left-context $extra_left_context \
+            --extra-right-context $extra_right_context \
+            --extra-left-context-initial 0 --extra-right-context-final 0 \
+            --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
+            $graphdir data/${unsupervised_set_perturbed}_hires $sup_chain_dir/decode_${unsupervised_set_perturbed}
+fi
+
+# Rescore undeterminized lattices with larger LM
+if [ $stage -le 5 ]; then
+  steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
+    --acwt 0.1 --beam 8.0 --skip-scoring true \
+    $unsup_decode_lang $unsup_rescore_lang \
+    data/${unsupervised_set_perturbed}_hires \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed} \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed}_big
+  ln -sf ../final.mdl $sup_chain_dir/decode_${unsupervised_set_perturbed}_big/final.mdl
+fi
+
+# Get best path alignment and lattice posterior of best path alignment to be
+# used as frame-weights in lattice-based training
+if [ $stage -le 8 ]; then
+  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
+    data/${unsupervised_set_perturbed}_hires \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed}_big \
+    $sup_chain_dir/best_path_${unsupervised_set_perturbed}_big
+fi
+
+frame_subsampling_factor=1
+if [ -f $sup_chain_dir/frame_subsampling_factor ]; then
+  frame_subsampling_factor=$(cat $sup_chain_dir/frame_subsampling_factor)
+fi
+cmvn_opts=$(cat $sup_chain_dir/cmvn_opts) || exit 1
+
+diff $sup_tree_dir/tree $sup_chain_dir/tree || { echo "$0: $sup_tree_dir/tree and $sup_chain_dir/tree differ"; exit 1; }
+
+# Uncomment the following lines if you need to build new tree using both
+# supervised and unsupervised data. This may help if amount of
+# supervised data used to train the seed system tree is very small.
+# unsupervised data
+
+# tree_affix=bi_semisup_a
+# treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+# if [ -f $treedir/final.mdl ]; then
+#   echo "$0: $treedir/final.mdl exists. Remove it and run again."
+#   exit 1
+# fi
+#
+# if [ $stage -le 9 ]; then
+#   # This is usually 3 for chain systems.
+#   echo $frame_subsampling_factor > \
+#     $sup_chain_dir/best_path_${unsupervised_set_perturbed}_big/frame_subsampling_factor
+#
+#   # This should be 1 if using a different source for supervised data alignments.
+#   # However alignments in seed tree directory have already been sub-sampled.
+#   echo $frame_subsampling_factor > \
+#     $sup_tree_dir/frame_subsampling_factor
+#
+#   # Build a new tree using stats from both supervised and unsupervised data
+#   steps/nnet3/chain/build_tree_multiple_sources.sh \
+#     --use-fmllr false --context-opts "--context-width=2 --central-position=1" \
+#     --frame-subsampling-factor $frame_subsampling_factor \
+#     7000 $lang \
+#     data/${supervised_set_perturbed} \
+#     ${sup_tree_dir} \
+#     data/${unsupervised_set_perturbed} \
+#     $chaindir/best_path_${unsupervised_set_perturbed} \
+#     $treedir || exit 1
+# fi
+#
+# sup_tree_dir=$treedir   # Use the new tree dir for further steps
+
+# Train denominator FST using phone alignments from
+# supervised and unsupervised data
+if [ $stage -le 10 ]; then
+  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
+    ${sup_tree_dir} ${sup_chain_dir}/best_path_${unsupervised_set_perturbed}_big \
+    $dir
+fi
+
+if [ $stage -le 11 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $sup_tree_dir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=40"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
+
+  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+
+  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+  # We use separate outputs for supervised and unsupervised data
+  # so we can properly track the train and valid objectives.
+
+  output name=output-0 input=output.affine
+  output name=output-1 input=output.affine
+
+  output name=output-0-xent input=output-xent.log-softmax
+  output name=output-1-xent input=output-xent.log-softmax
+EOF
+
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+. $dir/configs/vars
+
+left_context=$[model_left_context + chunk_left_context]
+right_context=$[model_right_context + chunk_right_context]
+left_context_initial=$model_left_context
+right_context_final=$model_right_context
+
+egs_left_context=$(perl -e "print int($left_context + $frame_subsampling_factor / 2)")
+egs_right_context=$(perl -e "print int($right_context + $frame_subsampling_factor / 2)")
+egs_left_context_initial=$(perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)")
+egs_right_context_final=$(perl -e "print int($right_context_final + $frame_subsampling_factor / 2)")
+
+if [ -z "$sup_egs_dir" ]; then
+  sup_egs_dir=$dir/egs_${supervised_set_perturbed}
+  frames_per_eg=$(cat $sup_chain_dir/egs/info/frames_per_eg)
+
+  if [ $stage -le 12 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
+    fi
+    mkdir -p $sup_egs_dir/
+    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the supervised data"
+    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
+               --left-context $egs_left_context --right-context $egs_right_context \
+               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+               --frame-subsampling-factor $frame_subsampling_factor \
+               --alignment-subsampling-factor $frame_subsampling_factor \
+               --frames-per-eg $frames_per_eg \
+               --frames-per-iter 1500000 \
+               --cmvn-opts "$cmvn_opts" \
+               --online-ivector-dir $sup_ivector_dir \
+               --generate-egs-scp true \
+               data/${supervised_set_perturbed}_hires $dir \
+               $sup_lat_dir $sup_egs_dir
+  fi
+else
+  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
+fi
+
+unsup_frames_per_eg=150  # Using a frames-per-eg of 150 for unsupervised data
+                         # was found to be better than allowing smaller chunks
+                         # (160,140,110,80) like for supervised system
+lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices when
+                      # creating numerator supervision
+lattice_prune_beam=4.0  # beam for pruning the lattices prior to getting egs
+                        # for unsupervised data
+tolerance=1   # frame-tolerance for chain training
+
+unsup_lat_dir=${sup_chain_dir}/decode_${unsupervised_set_perturbed}_big
+if [ -z "$unsup_egs_dir" ]; then
+  unsup_egs_dir=$dir/egs_${unsupervised_set_perturbed}
+
+  if [ $stage -le 13 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
+    fi
+    mkdir -p $unsup_egs_dir
+    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the unsupervised data"
+    steps/nnet3/chain/get_egs.sh \
+      --cmd "$decode_cmd" --alignment-subsampling-factor 1 \
+      --left-tolerance $tolerance --right-tolerance $tolerance \
+      --left-context $egs_left_context --right-context $egs_right_context \
+      --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+      --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
+      --frame-subsampling-factor $frame_subsampling_factor \
+      --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
+      --lattice-prune-beam "$lattice_prune_beam" \
+      --deriv-weights-scp $sup_chain_dir/best_path_${unsupervised_set_perturbed}_big/weights.scp \
+      --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+      --generate-egs-scp true $unsup_egs_opts \
+      data/${unsupervised_set_perturbed}_hires $dir \
+      $unsup_lat_dir $unsup_egs_dir
+  fi
+fi
+
+comb_egs_dir=$dir/comb_egs
+if [ $stage -le 14 ]; then
+  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
+    --block-size 128 \
+    --lang2weight $supervision_weights 2 \
+    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
+  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
+fi
+
+if [ $train_stage -le -4 ]; then
+  # This is to skip stages of den-fst creation, which was already done.
+  train_stage=-4
+fi
+
+if [ $stage -le 15 ]; then
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$comb_egs_dir" \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $sup_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights true \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.chunk-width $frames_per_eg \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.chunk-left-context-initial 0 \
+    --egs.chunk-right-context-final 0 \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.num-chunk-per-minibatch "$minibatch_size" \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.max-param-change 2.0 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --cleanup.remove-egs false \
+    --feat-dir data/${supervised_set_perturbed}_hires \
+    --tree-dir $sup_tree_dir \
+    --lat-dir $sup_lat_dir \
+    --dir $dir || exit 1;
+fi
+
+test_graph_dir=$dir/graph${test_graph_affix}
+if [ $stage -le 17 ]; then
+  # Note: it might appear that this $lang directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --self-loop-scale 1.0 ${test_lang} $dir $test_graph_dir
+fi
+
+if [ $stage -le 18 ]; then
+  rm -f $dir/.error
+  for decode_set in dev test; do
+    (
+      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
+      if [ $num_jobs -gt $test_nj ]; then num_jobs=$test_nj; fi
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+        --nj $num_jobs --cmd "$decode_cmd" ${decode_iter:+--iter $decode_iter} \
+        --online-ivector-dir $ivector_root_dir/ivectors_${decode_set}_hires \
+        --frames-per-chunk $frames_per_chunk_decoding \
+        --extra-left-context $extra_left_context \
+        --extra-right-context $extra_right_context \
+        --extra-left-context-initial 0 --extra-right-context-final 0 \
+        $test_graph_dir data/${decode_set}_hires \
+        $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || touch $dir/.error
+    ) &
+  done
+  wait;
+  if [ -f $dir/.error ]; then
+    echo "$0: Decoding failed. See $dir/decode${test_graph_affix}_*/log/*"
+    exit 1
+  fi
+fi
+
+exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_1c.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_1c.sh
new file mode 100644
index 00000000000..7de9103668f
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_1c.sh
@@ -0,0 +1,468 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script is semi-supervised recipe with around 50 hours of supervised data
+# and 250 hours unsupervised data with naive splitting.
+# Based on "Semi-Supervised Training of Acoustic Models using Lattice-Free MMI",
+# Vimal Manohar, Hossein Hadian, Daniel Povey, Sanjeev Khudanpur, ICASSP 2018
+# http://www.danielpovey.com/files/2018_icassp_semisupervised_mmi.pdf
+# local/semisup/run_50k.sh shows how to call this.
+
+# We use the combined data for i-vector extractor training.
+# We use 4-gram LM trained on 1250 hours of data excluding the 250 hours
+# unsupervised data to create LM for decoding. Rescoring is done with
+# a larger 4-gram LM.
+# This differs from the case in run_tdnn_100k_semisupervised.sh.
+
+# This script uses phone LM to model UNK.
+# This script uses the same tree as that for the seed model.
+# See the comments in the script about how to change these.
+
+# Unsupervised set: train_unsup100k_250k (250 hour subset of Fisher excluding 100 hours for supervised)
+# unsup_frames_per_eg=150
+# Deriv weights: Lattice posterior of best path pdf
+# Unsupervised weight: 1.0
+# Weights for phone LM (supervised, unsupervised): 3,2
+# LM for decoding unsupervised data: 4gram
+# Supervision: Naive split lattices
+
+set -u -e -o pipefail
+
+stage=0
+train_stage=-100
+nj=80
+test_nj=50
+
+# The following 3 options decide the output directory for semi-supervised
+# chain system
+# dir=${exp_root}/chain${chain_affix}/tdnn_lstm${tdnn_affix}
+
+exp_root=exp/semisup_50k
+chain_affix=_semi50k_100k_250k    # affix for chain dir
+                                  # 50 hour subset out of 100 hours of supervised data
+                                  # 250 hour subset out of (1500-100=1400) hours of unsupervised data 
+tdnn_affix=_semisup_1c
+
+# Datasets -- Expects data/$supervised_set and data/$unsupervised_set to be
+# present
+supervised_set=train_sup50k
+unsupervised_set=train_unsup100k_250k
+
+# Input seed system
+sup_chain_dir=exp/semisup_50k/chain_semi50k_100k_250k/tdnn_lstm_1c_sp  # supervised chain system
+sup_lat_dir=exp/semisup_50k/chain_semi50k_100k_250k/tri4a_train_sup50k_lats  # lattices for supervised set
+sup_tree_dir=exp/semisup_50k/chain_semi50k_100k_250k/tree_bi_c  # tree directory for supervised chain system
+ivector_root_dir=exp/semisup_50k/nnet3_semi50k_100k_250k  # i-vector extractor root directory
+
+# Semi-supervised options
+supervision_weights=1.0,1.0   # Weights for supervised, unsupervised data egs.
+                              # Can be used to scale down the effect of unsupervised data
+                              # by using a smaller scale for it e.g. 1.0,0.3
+lm_weights=3,2  # Weights on phone counts from supervised, unsupervised data for denominator FST creation
+
+sup_egs_dir=   # Supply this to skip supervised egs creation
+unsup_egs_dir=  # Supply this to skip unsupervised egs creation
+unsup_egs_opts=  # Extra options to pass to unsupervised egs creation
+
+# Neural network opts
+hidden_dim=1024
+cell_dim=1024
+projection_dim=256
+
+# training options
+num_epochs=2
+minibatch_size=64,32
+chunk_left_context=40
+chunk_right_context=0
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=5
+
+# decode options
+extra_left_context=50
+extra_right_context=0
+frames_per_chunk_decoding=160
+
+decode_iter=  # Iteration to decode with
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+if [ -f ./path.sh ]; then . ./path.sh; fi
+. ./utils/parse_options.sh
+
+# The following can be replaced with the versions that do not model
+# UNK using phone LM. $sup_lat_dir should also ideally be changed.
+unsup_decode_lang=data/lang_test_poco_ex250k
+unsup_decode_graph_affix=_poco_ex250k
+test_lang=data/lang_test_poco
+test_graph_affix=_poco
+
+unsup_rescore_lang=${unsup_decode_lang}_big
+
+dir=$exp_root/chain${chain_affix}/tdnn_lst${tdnn_affix}
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+supervised_set_perturbed=${supervised_set}_sp
+unsupervised_set_perturbed=${unsupervised_set}_sp
+
+sup_ivector_dir=$ivector_root_dir/ivectors_${supervised_set_perturbed}_hires
+
+graphdir=$sup_chain_dir/graph${unsup_decode_graph_affix}
+
+for f in data/${supervised_set_perturbed}/feats.scp \
+  data/${supervised_set_perturbed}_hires/feats.scp \
+  $ivector_root_dir/extractor/final.ie $sup_ivector_dir/ivector_online.scp \
+  $sup_lat_dir/lat.1.gz $sup_tree_dir/ali.1.gz \
+  $unsup_decode_lang/G.fst; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find file $f"
+    exit 1
+  fi
+done
+
+if [ ! -f $graphdir/HCLG.fst ]; then
+  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $sup_chain_dir $graphdir
+fi
+
+if [ $stage -le 2 ]; then
+  utils/data/perturb_data_dir_speed_3way.sh data/${unsupervised_set} \
+    data/${unsupervised_set_perturbed}_hires
+  utils/data/perturb_data_dir_volume.sh \
+    data/${unsupervised_set_perturbed}_hires
+
+  steps/make_mfcc.sh --cmd "$train_cmd" --nj $nj \
+    --mfcc-config conf/mfcc_hires.conf \
+    data/${unsupervised_set_perturbed}_hires
+  steps/compute_cmvn_stats.sh data/${unsupervised_set_perturbed}_hires
+  utils/fix_data_dir.sh data/${unsupervised_set_perturbed}_hires
+fi
+
+# Extract i-vectors for the unsupervised data
+if [ $stage -le 3 ]; then
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+    data/${unsupervised_set_perturbed}_hires data/${unsupervised_set_perturbed}_max2_hires
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+    data/${unsupervised_set_perturbed}_max2_hires $ivector_root_dir/extractor \
+    $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires || exit 1
+fi
+
+# Decode unsupervised data and write lattices in non-compact
+# undeterminized format
+# Set --skip-scoring to false in order to score the unsupervised data
+if [ $stage -le 4 ]; then
+  echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
+  steps/nnet3/decode_semisup.sh --num-threads 4 --nj $nj --cmd "$decode_cmd" \
+            --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
+            --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+            --frames-per-chunk 160 \
+            --extra-left-context $extra_left_context \
+            --extra-right-context $extra_right_context \
+            --extra-left-context-initial 0 --extra-right-context-final 0 \
+            --scoring-opts "--min-lmwt 10 --max-lmwt 10" --word-determinize false \
+            $graphdir data/${unsupervised_set_perturbed}_hires $sup_chain_dir/decode_${unsupervised_set_perturbed}
+fi
+
+# Rescore undeterminized lattices with larger LM
+if [ $stage -le 5 ]; then
+  steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
+    --acwt 0.1 --beam 8.0 --skip-scoring true \
+    $unsup_decode_lang $unsup_rescore_lang \
+    data/${unsupervised_set_perturbed}_hires \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed} \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed}_big
+  ln -sf ../final.mdl $sup_chain_dir/decode_${unsupervised_set_perturbed}_big/final.mdl
+fi
+
+# Get best path alignment and lattice posterior of best path alignment to be
+# used as frame-weights in lattice-based training
+if [ $stage -le 8 ]; then
+  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
+    data/${unsupervised_set_perturbed}_hires \
+    $sup_chain_dir/decode_${unsupervised_set_perturbed}_big \
+    $sup_chain_dir/best_path_${unsupervised_set_perturbed}_big
+fi
+
+frame_subsampling_factor=1
+if [ -f $sup_chain_dir/frame_subsampling_factor ]; then
+  frame_subsampling_factor=$(cat $sup_chain_dir/frame_subsampling_factor)
+fi
+cmvn_opts=$(cat $sup_chain_dir/cmvn_opts) || exit 1
+
+diff $sup_tree_dir/tree $sup_chain_dir/tree || { echo "$0: $sup_tree_dir/tree and $sup_chain_dir/tree differ"; exit 1; }
+
+# Uncomment the following lines if you need to build new tree using both
+# supervised and unsupervised data. This may help if amount of
+# supervised data used to train the seed system tree is very small.
+# unsupervised data
+
+# tree_affix=bi_semisup_a
+# treedir=$exp_root/chain${chain_affix}/tree_${tree_affix}
+# if [ -f $treedir/final.mdl ]; then
+#   echo "$0: $treedir/final.mdl exists. Remove it and run again."
+#   exit 1
+# fi
+#
+# if [ $stage -le 9 ]; then
+#   # This is usually 3 for chain systems.
+#   echo $frame_subsampling_factor > \
+#     $sup_chain_dir/best_path_${unsupervised_set_perturbed}_big/frame_subsampling_factor
+#
+#   # This should be 1 if using a different source for supervised data alignments.
+#   # However alignments in seed tree directory have already been sub-sampled.
+#   echo $frame_subsampling_factor > \
+#     $sup_tree_dir/frame_subsampling_factor
+#
+#   # Build a new tree using stats from both supervised and unsupervised data
+#   steps/nnet3/chain/build_tree_multiple_sources.sh \
+#     --use-fmllr false --context-opts "--context-width=2 --central-position=1" \
+#     --frame-subsampling-factor $frame_subsampling_factor \
+#     7000 $lang \
+#     data/${supervised_set_perturbed} \
+#     ${sup_tree_dir} \
+#     data/${unsupervised_set_perturbed} \
+#     $chaindir/best_path_${unsupervised_set_perturbed} \
+#     $treedir || exit 1
+# fi
+#
+# sup_tree_dir=$treedir   # Use the new tree dir for further steps
+
+# Train denominator FST using phone alignments from
+# supervised and unsupervised data
+if [ $stage -le 10 ]; then
+  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
+    ${sup_tree_dir} ${sup_chain_dir}/best_path_${unsupervised_set_perturbed}_big \
+    $dir
+fi
+
+if [ $stage -le 11 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $sup_tree_dir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=40"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
+
+  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
+  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
+
+  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+  # We use separate outputs for supervised and unsupervised data
+  # so we can properly track the train and valid objectives.
+
+  output name=output-0 input=output.affine
+  output name=output-1 input=output.affine
+
+  output name=output-0-xent input=output-xent.log-softmax
+  output name=output-1-xent input=output-xent.log-softmax
+EOF
+
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+. $dir/configs/vars
+
+left_context=$[model_left_context + chunk_left_context]
+right_context=$[model_right_context + chunk_right_context]
+left_context_initial=$model_left_context
+right_context_final=$model_right_context
+
+egs_left_context=$(perl -e "print int($left_context + $frame_subsampling_factor / 2)")
+egs_right_context=$(perl -e "print int($right_context + $frame_subsampling_factor / 2)")
+egs_left_context_initial=$(perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)")
+egs_right_context_final=$(perl -e "print int($right_context_final + $frame_subsampling_factor / 2)")
+
+if [ -z "$sup_egs_dir" ]; then
+  sup_egs_dir=$dir/egs_${supervised_set_perturbed}
+  frames_per_eg=$(cat $sup_chain_dir/egs/info/frames_per_eg)
+
+  if [ $stage -le 12 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
+    fi
+    mkdir -p $sup_egs_dir/
+    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the supervised data"
+    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
+               --left-context $egs_left_context --right-context $egs_right_context \
+               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+               --frame-subsampling-factor $frame_subsampling_factor \
+               --alignment-subsampling-factor $frame_subsampling_factor \
+               --frames-per-eg $frames_per_eg \
+               --frames-per-iter 1500000 \
+               --cmvn-opts "$cmvn_opts" \
+               --online-ivector-dir $sup_ivector_dir \
+               --generate-egs-scp true \
+               data/${supervised_set_perturbed}_hires $dir \
+               $sup_lat_dir $sup_egs_dir
+  fi
+else
+  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
+fi
+
+unsup_frames_per_eg=150  # Using a frames-per-eg of 150 for unsupervised data
+                         # was found to be better than allowing smaller chunks
+                         # (160,140,110,80) like for supervised system
+lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices when
+                      # creating numerator supervision
+lattice_prune_beam=4.0  # beam for pruning the lattices prior to getting egs
+                        # for unsupervised data
+tolerance=1   # frame-tolerance for chain training
+
+unsup_lat_dir=${sup_chain_dir}/decode_${unsupervised_set_perturbed}_big
+if [ -z "$unsup_egs_dir" ]; then
+  unsup_egs_dir=$dir/egs_${unsupervised_set_perturbed}
+
+  if [ $stage -le 13 ]; then
+    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
+      utils/create_split_dir.pl \
+       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
+    fi
+    mkdir -p $unsup_egs_dir
+    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
+
+    echo "$0: generating egs from the unsupervised data"
+    steps/nnet3/chain/get_egs.sh \
+      --cmd "$decode_cmd" --alignment-subsampling-factor 1 \
+      --left-tolerance $tolerance --right-tolerance $tolerance \
+      --left-context $egs_left_context --right-context $egs_right_context \
+      --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
+      --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
+      --frame-subsampling-factor $frame_subsampling_factor \
+      --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
+      --lattice-prune-beam "$lattice_prune_beam" \
+      --deriv-weights-scp $sup_chain_dir/best_path_${unsupervised_set_perturbed}_big/weights.scp \
+      --online-ivector-dir $ivector_root_dir/ivectors_${unsupervised_set_perturbed}_hires \
+      --generate-egs-scp true $unsup_egs_opts \
+      data/${unsupervised_set_perturbed}_hires $dir \
+      $unsup_lat_dir $unsup_egs_dir
+  fi
+fi
+
+comb_egs_dir=$dir/comb_egs
+if [ $stage -le 14 ]; then
+  steps/nnet3/chain/multilingual/combine_egs.sh --cmd "$train_cmd" \
+    --block-size 128 \
+    --lang2weight $supervision_weights 2 \
+    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
+  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
+fi
+
+if [ $train_stage -le -4 ]; then
+  # This is to skip stages of den-fst creation, which was already done.
+  train_stage=-4
+fi
+
+if [ $stage -le 15 ]; then
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$comb_egs_dir" \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $sup_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights true \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.chunk-width $frames_per_eg \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.chunk-left-context-initial 0 \
+    --egs.chunk-right-context-final 0 \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.num-chunk-per-minibatch "$minibatch_size" \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.max-param-change 2.0 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --cleanup.remove-egs false \
+    --feat-dir data/${supervised_set_perturbed}_hires \
+    --tree-dir $sup_tree_dir \
+    --lat-dir $sup_lat_dir \
+    --dir $dir || exit 1;
+fi
+
+test_graph_dir=$dir/graph${test_graph_affix}
+if [ $stage -le 17 ]; then
+  # Note: it might appear that this $lang directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --self-loop-scale 1.0 ${test_lang} $dir $test_graph_dir
+fi
+
+if [ $stage -le 18 ]; then
+  rm -f $dir/.error
+  for decode_set in dev test; do
+    (
+      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
+      if [ $num_jobs -gt $test_nj ]; then num_jobs=$test_nj; fi
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+        --nj $num_jobs --cmd "$decode_cmd" ${decode_iter:+--iter $decode_iter} \
+        --online-ivector-dir $ivector_root_dir/ivectors_${decode_set}_hires \
+        --frames-per-chunk $frames_per_chunk_decoding \
+        --extra-left-context $extra_left_context \
+        --extra-right-context $extra_right_context \
+        --extra-left-context-initial 0 --extra-right-context-final 0 \
+        $test_graph_dir data/${decode_set}_hires \
+        $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || touch $dir/.error
+    ) &
+  done
+  wait;
+  if [ -f $dir/.error ]; then
+    echo "$0: Decoding failed. See $dir/decode${test_graph_affix}_*/log/*"
+    exit 1
+  fi
+fi
+
+exit 0;
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_a.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_a.sh
deleted file mode 100644
index bab9e69bbf3..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_a.sh
+++ /dev/null
@@ -1,511 +0,0 @@
-#!/bin/bash
-
-# This script uses phone LM to model UNK.
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=40
-exp=exp/semisup_50k
-
-unsupervised_set=train_unsup100k_250k  # set this to your choice of unsupervised data
-supervised_set=train_sup50k
-semisup_train_set=semisup50k_250k
-
-tdnn_affix=7a  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_i
-
-gmm=tri4a
-
-nnet3_affix=_semi50k_250k    # affix for nnet3 and chain dir -- relates to i-vector used
-chain_affix=_semi50k_250k
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb1b  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=5,2
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024 
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_50k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain_unk
-unsup_decode_lang=data/lang_poco_test_ex250k_unk
-unsup_rescore_lang=${unsup_decode_lang}_big
-unsup_decode_graph_affix=_poco_ex250k_unk
-
-test_lang=data/lang_poco_test_unk
-test_graph_affix=_poco_unk
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-done
-
-decode_affix=${decode_affix}_big
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-fi
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/$gmm
-
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${chain_affix}/${gmm}_${supervised_set}_unk_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. 2 \
-    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs 4 \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_b.sh b/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_b.sh
deleted file mode 100644
index ebd6c090267..00000000000
--- a/egs/fisher_english/s5/local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_conf_b.sh
+++ /dev/null
@@ -1,511 +0,0 @@
-#!/bin/bash
-
-# This script uses phone LM to model UNK.
-
-# Unsupervised set: train_unsup100k_250k
-# unsup_frames_per_eg=150
-# Deriv weights: Lattice posterior of best path pdf
-# Unsupervised weight: 1.0
-# Weights for phone LM (supervised, unsupervises): 3,2
-# LM for decoding unsupervised data: 4gram
-
-set -u -e -o pipefail
-
-stage=-2
-train_stage=-100
-nj=40
-decode_nj=40
-exp=exp/semisup_50k
-
-unsupervised_set=train_unsup100k_250k  # set this to your choice of unsupervised data
-supervised_set=train_sup50k
-semisup_train_set=semisup50k_250k
-
-tdnn_affix=7b  # affix for the supervised chain-model directory
-train_supervised_opts="--stage -10 --train-stage -10"
-tree_affix=bi_j
-
-gmm=tri4a
-
-nnet3_affix=_semi50k_250k    # affix for nnet3 and chain dir -- relates to i-vector used
-chain_affix=_semi50k_250k
-
-# Unsupervised options
-decode_affix=
-egs_affix=  # affix for the egs that are generated from unsupervised data and for the comined egs dir
-unsup_frames_per_eg=150  # if empty will be equal to the supervised model's config -- you will need to change minibatch_size for comb training accordingly
-lattice_lm_scale=0.5  # lm-scale for using the weights from unsupervised lattices
-lattice_prune_beam=4.0  # If supplied will prune the lattices prior to getting egs for unsupervised data
-tolerance=1
-phone_insertion_penalty=
-
-# Semi-supervised options
-comb_affix=comb1b  # affix for new chain-model directory trained on the combined supervised+unsupervised subsets
-supervision_weights=1.0,1.0
-lm_weights=5,2
-sup_egs_dir=
-unsup_egs_dir=
-unsup_egs_opts=
-
-remove_egs=false
-common_egs_dir=
-
-hidden_dim=1024 
-cell_dim=1024
-projection_dim=256
-
-apply_deriv_weights=true
-use_smart_splitting=true
-
-# training options
-num_epochs=2
-minibatch_size=64,32
-chunk_left_context=40
-chunk_right_context=0
-dropout_schedule='0,0@0.20,0.3@0.50,0'
-xent_regularize=0.025
-self_repair_scale=0.00001
-label_delay=5
-
-# decode options
-extra_left_context=50
-extra_right_context=0
-
-decode_iter=
-
-do_finetuning=false
-
-finetune_stage=-2
-finetune_suffix=_finetune
-finetune_iter=final
-num_epochs_finetune=1
-finetune_xent_regularize=0.1
-finetune_opts="--chain.mmi-factor-schedule=0.05,0.05 --chain.smbr-factor-schedule=0.05,0.05"
-
-# End configuration section.
-echo "$0 $@"  # Print the command line for logging
-
-. cmd.sh
-. ./path.sh
-. ./utils/parse_options.sh
-
-egs_affix=${egs_affix}_prun${lattice_prune_beam}_lmwt${lattice_lm_scale}_tol${tolerance}
-if $use_smart_splitting; then
-  comb_affix=${comb_affix:+${comb_affix}_smart}
-else
-  comb_affix=${comb_affix:+${comb_affix}_naive}
-fi
-
-RANDOM=0
-
-if ! cuda-compiled; then
-  cat <<EOF && exit 1
-This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
-If you want to use GPUs (and have them), go to src/, and configure and make on a machine
-where "nvcc" is installed.
-EOF
-fi
-
-if false && [ $stage -le 1 ]; then
-  echo "$0: chain training on the supervised subset data/${supervised_set}"
-  local/semisup/chain/run_tdnn_lstm_50k.sh $train_supervised_opts --remove-egs false \
-                          --train-set $supervised_set --ivector-train-set $supervised_set \
-                          --nnet3-affix "$nnet3_affix" --tdnn-affix "$tdnn_affix" --exp $exp
-fi
-
-lang=data/lang_chain
-unsup_decode_lang=data/lang_poco_test_ex250k
-unsup_rescore_lang=${unsup_decode_lang}_big
-unsup_decode_graph_affix=_poco_ex250k
-
-test_lang=data/lang_poco_test
-test_graph_affix=_poco
-
-decode_affix=${decode_affix}${unsup_decode_graph_affix}
-
-extractor=$exp/nnet3${nnet3_affix}/extractor
-chaindir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}_sp
-graphdir=$chaindir/graph${unsup_decode_graph_affix}
-if [ ! -f $graphdir/HCLG.fst ]; then
-  utils/mkgraph.sh --self-loop-scale 1.0 $unsup_decode_lang $chaindir $graphdir
-fi
-
-if [ $stage -le 2 ]; then
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set} 10000 data/${unsupervised_set}_10k
-  utils/subset_data_dir.sh --speakers data/${unsupervised_set}_10k 5000 data/${unsupervised_set}_10k_calib_train
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k_calib_train/utt2spk data/${unsupervised_set}_10k/utt2spk) \
-    data/${unsupervised_set}_10k data/${unsupervised_set}_10k_calib_dev
-  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/${unsupervised_set}_10k/utt2spk data/${unsupervised_set}/utt2spk) \
-    data/${unsupervised_set} data/${unsupervised_set}_n10k
-fi
-
-unsupervised_set=${unsupervised_set}_n10k
-
-for dset in $unsupervised_set; do
-  if [ $stage -le 3 ]; then
-    if [ -f data/${dset}_sp_hires/feats.scp ]; then
-      echo "$0: data/${dset}_sp_hires/feats.scp exists. Remove it or re-run from next stage"
-      exit 1
-    fi
-
-    utils/data/perturb_data_dir_speed_3way.sh data/$dset data/${dset}_sp_hires
-    utils/data/perturb_data_dir_volume.sh data/${dset}_sp_hires
-
-    steps/make_mfcc.sh --nj $decode_nj --cmd "$train_cmd" \
-      --mfcc-config conf/mfcc_hires.conf data/${dset}_sp_hires || exit 1
-  fi
-
-  if [ $stage -le 5 ]; then
-    echo "$0: getting the decoding lattices for the unsupervised subset using the chain model at: $chaindir"
-    steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
-              --acwt 1.0 --post-decode-acwt 10.0 --write-compact false --skip-scoring true \
-              --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-              --frames-per-chunk 160 \
-              --extra-left-context $extra_left_context \
-              --extra-right-context $extra_right_context \
-              --extra-left-context-initial 0 --extra-right-context-final 0 \
-              --scoring-opts "--min-lmwt 10 --max-lmwt 10" --determinize-opts "--word-determinize=false" \
-              $graphdir data/${dset}_sp_hires $chaindir/decode_${dset}_sp${decode_affix}
-  fi
-
-  if [ $stage -le 6 ]; then
-    steps/lmrescore_const_arpa_undeterminized.sh --cmd "$decode_cmd" \
-      --write-compact false --acwt 0.1 --beam 8.0 --skip-scoring true \
-      ${unsup_decode_lang} ${unsup_rescore_lang} \
-      data/${dset}_sp_hires \
-      $chaindir/decode_${dset}_sp${decode_affix} \
-      $chaindir/decode_${dset}_sp${decode_affix}_big
-
-    ln -sf ../final.mdl $chaindir/decode_${dset}_sp${decode_affix}_big/ || true
-  fi
-done
-
-decode_affix=${decode_affix}_big
-if [ $stage -le 8 ]; then
-  steps/best_path_weights.sh --cmd "${train_cmd}" --acwt 0.1 \
-    data/${unsupervised_set}_sp_hires $lang \
-    $chaindir/decode_${unsupervised_set}_sp${decode_affix} \
-    $chaindir/best_path_${unsupervised_set}_sp${decode_affix}
-fi
-
-frame_subsampling_factor=1
-if [ -f $chaindir/frame_subsampling_factor ]; then
-  frame_subsampling_factor=`cat $chaindir/frame_subsampling_factor`
-fi
-cmvn_opts=`cat $chaindir/cmvn_opts` || exit 1
-
-sup_ali_dir=$exp/$gmm
-
-treedir=$exp/chain${chain_affix}/tree_${tree_affix}
-if [ ! -f $treedir/final.mdl ]; then
-  echo "$0: $treedir/final.mdl does not exist."
-  exit 1
-fi
-
-diff $treedir/tree $chaindir/tree || { echo "$0: $treedir/tree and $chaindir/tree differ"; exit 1; }
-
-dir=$exp/chain${chain_affix}/tdnn_lstm${tdnn_affix}${decode_affix}${egs_affix}${comb_affix:+_$comb_affix}
-
-
-if [ $stage -le 10 ]; then
-  steps/nnet3/chain/make_weighted_den_fst.sh --num-repeats $lm_weights --cmd "$train_cmd" \
-    ${treedir} ${chaindir}/best_path_${unsupervised_set}_sp${decode_affix} \
-    $dir
-fi
-
-if [ $stage -le 11 ]; then
-  echo "$0: creating neural net configs using the xconfig parser";
-
-  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-
-  lstm_opts="decay-time=40"
-
-  mkdir -p $dir/configs
-  cat <<EOF > $dir/configs/network.xconfig
-  input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
-  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
-
-  # the first splicing is moved before the lda layer, so no splicing here
-  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$hidden_dim
-
-  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$hidden_dim
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$hidden_dim
-  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 dropout-proportion=0.0 $lstm_opts
-
-  output-layer name=output input=lstm3 output-delay=$label_delay dim=$num_targets include-log-softmax=false max-change=1.5
-
-  # adding the layers for xent branch
-  # This block prints the configs for a separate output that will be
-  # trained with a cross-entropy objective in the 'chain' models... this
-  # has the effect of regularizing the hidden parts of the model.  we use
-  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
-  # 0.5 / args.xent_regularize is suitable as it means the xent
-  # final-layer learns at a rate independent of the regularization
-  # constant; and the 0.5 was tuned so as to make the relative progress
-  # similar in the xent and regular final layers.
-  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
-
-  output name=output-0 input=output.affine@$label_delay skip-in-init=true
-  output name=output-1 input=output.affine@$label_delay skip-in-init=true
-
-  output name=output-0-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-  output name=output-1-xent input=output-xent.log-softmax@$label_delay skip-in-init=true
-EOF
-
-  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
-fi
-
-. $dir/configs/vars
-
-left_context=$[model_left_context + chunk_left_context]
-right_context=$[model_right_context + chunk_right_context]
-left_context_initial=$model_left_context
-right_context_final=$model_right_context
-
-egs_left_context=`perl -e "print int($left_context + $frame_subsampling_factor / 2)"`
-egs_right_context=`perl -e "print int($right_context + $frame_subsampling_factor / 2)"`
-egs_left_context_initial=`perl -e "print int($left_context_initial + $frame_subsampling_factor / 2)"`
-egs_right_context_final=`perl -e "print int($right_context_final + $frame_subsampling_factor / 2)"`
-
-supervised_set=${supervised_set}_sp
-sup_lat_dir=$exp/chain${chain_affix}/${gmm}_${supervised_set}_lats
-if [ -z "$sup_egs_dir" ]; then
-  sup_egs_dir=$dir/egs_${supervised_set}
-  frames_per_eg=$(cat $chaindir/egs/info/frames_per_eg)
-
-  if [ $stage -le 12 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $sup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$sup_egs_dir/storage $sup_egs_dir/storage
-    fi
-    mkdir -p $sup_egs_dir/
-    touch $sup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the supervised data"
-    steps/nnet3/chain/get_egs.sh --cmd "$decode_cmd" \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --alignment-subsampling-factor 3 \
-               --frames-per-eg $frames_per_eg \
-               --frames-per-iter 1500000 \
-               --cmvn-opts "$cmvn_opts" \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true \
-               data/${supervised_set}_hires $dir \
-               $sup_lat_dir $sup_egs_dir
-  fi
-else
-  frames_per_eg=$(cat $sup_egs_dir/info/frames_per_eg)
-fi
-
-unsupervised_set=${unsupervised_set}_sp
-unsup_lat_dir=${chaindir}/decode_${unsupervised_set}${decode_affix} 
-
-if [ -z "$unsup_egs_dir" ]; then
-  [ -z $unsup_frames_per_eg ] && [ ! -z "$frames_per_eg" ] && unsup_frames_per_eg=$frames_per_eg
-  unsup_egs_dir=$dir/egs_${unsupervised_set}${decode_affix}${egs_affix}
-
-  if [ $stage -le 13 ]; then
-    if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $unsup_egs_dir/storage ]; then
-      utils/create_split_dir.pl \
-       /export/b0{5,6,7,8}/$USER/kaldi-data/egs/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$unsup_egs_dir/storage $unsup_egs_dir/storage
-    fi
-    mkdir -p $unsup_egs_dir
-    touch $unsup_egs_dir/.nodelete # keep egs around when that run dies.
-
-    echo "$0: generating egs from the unsupervised data"
-    if $use_smart_splitting; then
-      get_egs_script=steps/nnet3/chain/get_egs_split.sh
-    else
-      get_egs_script=steps/nnet3/chain/get_egs.sh
-    fi
-
-    $get_egs_script --cmd "$decode_cmd --h-rt 100:00:00" --alignment-subsampling-factor 1 \
-               --left-tolerance $tolerance --right-tolerance $tolerance \
-               --left-context $egs_left_context --right-context $egs_right_context \
-               --left-context-initial $egs_left_context_initial --right-context-final $egs_right_context_final \
-               --frames-per-eg $unsup_frames_per_eg --frames-per-iter 1500000 \
-               --frame-subsampling-factor $frame_subsampling_factor \
-               --cmvn-opts "$cmvn_opts" --lattice-lm-scale $lattice_lm_scale \
-               --lattice-prune-beam "$lattice_prune_beam" \
-               --phone-insertion-penalty "$phone_insertion_penalty" \
-               --deriv-weights-scp $chaindir/best_path_${unsupervised_set}${decode_affix}/weights.scp \
-               --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-               --generate-egs-scp true $unsup_egs_opts \
-               data/${unsupervised_set}_hires $dir \
-               $unsup_lat_dir $unsup_egs_dir
-  fi
-fi
-
-comb_egs_dir=$dir/${comb_affix}_egs${decode_affix}${egs_affix}_multi
-
-if [ $stage -le 14 ]; then
-  steps/nnet3/multilingual/combine_egs.sh --cmd "$train_cmd" \
-    --minibatch-size 64 --frames-per-iter 1500000 \
-    --lang2weight $supervision_weights --egs-prefix cegs. 2 \
-    $sup_egs_dir $unsup_egs_dir $comb_egs_dir
-  touch $comb_egs_dir/.nodelete # keep egs around when that run dies.
-fi
-
-if [ $train_stage -le -4 ]; then
-  train_stage=-4
-fi
-
-if [ $stage -le 15 ]; then
-  steps/nnet3/chain/train.py --stage $train_stage \
-    --egs.dir "$comb_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${semisup_train_set}_sp_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize $xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights $apply_deriv_weights \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width $frames_per_eg \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.dropout-schedule $dropout_schedule \
-    --trainer.num-chunk-per-minibatch "$minibatch_size" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.max-param-change 2.0 \
-    --trainer.num-epochs 4 \
-    --trainer.optimization.shrink-value 0.99 \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.001 \
-    --trainer.optimization.final-effective-lrate 0.0001 \
-    --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 8 \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir $dir  || exit 1;
-fi
-
-graph_dir=$dir/graph${test_graph_affix}
-if [ $stage -le 17 ]; then
-  # Note: it might appear that this $lang directory is mismatched, and it is as
-  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
-  # the lang directory.
-  utils/mkgraph.sh --self-loop-scale 1.0 $test_lang $dir $graph_dir
-fi
-
-if [ $stage -le 18 ]; then
-  iter_opts=
-  if [ ! -z $decode_iter ]; then
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/${decode_iter}.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/${decode_iter}.mdl $dir/${decode_iter}-output.mdl || exit 1
-    iter_opts=" --iter ${decode_iter}-output "
-  else
-    nnet3-copy --edits="remove-output-nodes name=output;rename-node old-name=output-0 new-name=output" $dir/final.mdl - | \
-      nnet3-am-copy --set-raw-nnet=- $dir/final.mdl $dir/final-output.mdl || exit 1
-    iter_opts=" --iter final-output "
-  fi
-
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 160 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode${test_graph_affix}_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-if ! $do_finetuning; then
-  wait
-  exit 0
-fi
-
-if [ $stage -le 19 ]; then
-  mkdir -p ${dir}${finetune_suffix}
-  
-  for f in phone_lm.fst normalization.fst den.fst tree 0.trans_mdl cmvn_opts; do
-    cp ${dir}/$f ${dir}${finetune_suffix} || exit 1
-  done
-  cp -r ${dir}/configs ${dir}${finetune_suffix} || exit 1
-
-  nnet3-copy --edits="remove-output-nodes name=output;remove-output-nodes name=output-xent;rename-node old-name=output-0 new-name=output;rename-node old-name=output-0-xent new-name=output-xent" \
-    $dir/${finetune_iter}.mdl ${dir}${finetune_suffix}/init.raw
-
-  if [ $finetune_stage -le -1 ]; then
-    finetune_stage=-1
-  fi
-
-  steps/nnet3/chain/train.py --stage $finetune_stage \
-    --trainer.input-model ${dir}${finetune_suffix}/init.raw \
-    --egs.dir "$sup_egs_dir" \
-    --cmd "$decode_cmd" \
-    --feat.online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${supervised_set}_hires \
-    --feat.cmvn-opts "--norm-means=false --norm-vars=false" $finetune_opts \
-    --chain.xent-regularize $finetune_xent_regularize \
-    --chain.leaky-hmm-coefficient 0.1 \
-    --chain.l2-regularize 0.00005 \
-    --chain.apply-deriv-weights true \
-    --chain.lm-opts="--num-extra-lm-states=2000" \
-    --egs.opts "--frames-overlap-per-eg 0" \
-    --egs.chunk-width 150 \
-    --egs.chunk-left-context $chunk_left_context \
-    --egs.chunk-right-context $chunk_right_context \
-    --egs.chunk-left-context-initial 0 \
-    --egs.chunk-right-context-final 0 \
-    --trainer.num-chunk-per-minibatch "150=64/300=32" \
-    --trainer.frames-per-iter 1500000 \
-    --trainer.num-epochs $num_epochs_finetune \
-    --trainer.optimization.num-jobs-initial 3 \
-    --trainer.optimization.num-jobs-final 16 \
-    --trainer.optimization.initial-effective-lrate 0.0001 \
-    --trainer.optimization.final-effective-lrate 0.00001 \
-    --trainer.max-param-change 2.0 \
-    --trainer.optimization.do-final-combination false \
-    --cleanup.remove-egs false \
-    --feat-dir data/${supervised_set}_hires \
-    --tree-dir $treedir \
-    --lat-dir $sup_lat_dir \
-    --dir ${dir}${finetune_suffix} || exit 1;
-fi
-
-dir=${dir}${finetune_suffix}
-
-if [ $stage -le 20 ]; then
-  for decode_set in dev test; do
-      (
-      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
-      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $num_jobs --cmd "$decode_cmd" \
-          --online-ivector-dir $exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
-          --frames-per-chunk 150 \
-          --extra-left-context $extra_left_context \
-          --extra-right-context $extra_right_context \
-          --extra-left-context-initial 0 --extra-right-context-final 0 \
-          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_iter$decode_iter} || exit 1;
-      ) &
-  done
-fi
-
-wait;
-exit 0;
-
diff --git a/egs/fisher_english/s5/local/semisup/run_lstm_100k.sh b/egs/fisher_english/s5/local/semisup/run_lstm_100k.sh
new file mode 100644
index 00000000000..318cc55542a
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/run_lstm_100k.sh
@@ -0,0 +1,89 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script demonstrates semi-supervised training using 100 hours of 
+# supervised data and 250 hours of unsupervised data.
+# We assume the supervised data is in data/train_sup and unsupervised data
+# is in data/train_unsup100k_250k. 
+# For LM training, we only use the supervised set corresponding to 100 hours as 
+# opposed to the case in run_50k.sh, where we included part of the 
+# transcripts in data/train/text.
+# This uses only 100 hours supervised set for i-vector extractor training, 
+# which is different from run_50k.sh, which uses combined supervised + 
+# unsupervised set.
+# This script is expected to be run after stage 8 of run_100k.sh.
+
+. ./cmd.sh
+. ./path.sh 
+
+set -o pipefail
+exp_root=exp/semisup_100k
+
+stage=0
+
+. utils/parse_options.sh
+
+for f in data/train_sup/utt2spk data/train_unsup100k_250k/utt2spk \
+  data/train_sup/text \
+  data/lang_test_poco_sup100k/G.fst \
+  data/lang_test_poco_sup100k_unk/G.fst
+  ; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find $f"
+    exit 1
+  fi
+done
+
+
+###############################################################################
+# Train seed chain system using 100 hours supervised data.
+# Here we train i-vector extractor on only the supervised set.
+###############################################################################
+
+if [ $stage -le 1 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1b.sh \
+    --train-set train_sup \
+    --ivector-train-set "" \
+    --nnet3-affix "" --chain-affix "" \
+    --tdnn-affix _1b --tree-affix bi_a \
+    --gmm tri4a --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Semi-supervised training using 100 hours supervised data and 
+# 250 hours unsupervised data. We use i-vector extractor, tree, lattices 
+# and seed chain system from the previous stage.
+###############################################################################
+
+if [ $stage -le 2 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1b.sh \
+    --supervised-set train_sup \
+    --unsupervised-set train_unsup100k_250k \
+    --sup-chain-dir $exp_root/chain/tdnn_lstm_1b_sp \
+    --sup-lat-dir $exp_root/chain/tri4a_train_sup_unk_lats \
+    --sup-tree-dir $exp_root/chain/tree_bi_a \
+    --ivector-root-dir $exp_root/nnet3 \
+    --chain-affix "" \
+    --tdnn-affix _semisup100k_250k_1b \
+    --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Oracle system trained on combined 350 hours including both supervised and 
+# unsupervised sets. We use i-vector extractor, tree, and GMM trained
+# on only the supervised for fair comparison to semi-supervised experiments.
+###############################################################################
+
+if [ $stage -le 3 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1b.sh \
+    --hidden-dim 1536 --cell-dim 1536 --projection-dim 384 \
+    --train-set semisup100k_250k \
+    --nnet3-affix "" --chain-affix "" \
+    --common-treedir $exp_root/chain/tree_bi_a \
+    --tdnn-affix _oracle100k_250k_1b --nj 100 \
+    --gmm tri4a --exp $exp_root \
+    --stage 9 || exit 1
+fi
+
diff --git a/egs/fisher_english/s5/local/semisup/run_lstm_100k_500k.sh b/egs/fisher_english/s5/local/semisup/run_lstm_100k_500k.sh
new file mode 100644
index 00000000000..9584a0296be
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/run_lstm_100k_500k.sh
@@ -0,0 +1,96 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script demonstrates semi-supervised training using 100 hours of 
+# supervised data and 250 hours of unsupervised data.
+# We assume the supervised data is in data/train_sup and unsupervised data
+# is in data/train_unsup100k_250k. 
+# For LM training, we only use the supervised set corresponding to 100 hours as 
+# opposed to the case in run_50k.sh, where we included part of the 
+# transcripts in data/train/text.
+# This uses only 100 hours supervised set for i-vector extractor training, 
+# which is different from run_50k.sh, which uses combined supervised + 
+# unsupervised set.
+# This script is expected to be run after stage 8 of run_100k.sh.
+
+. ./cmd.sh
+. ./path.sh 
+
+set -o pipefail
+exp_root=exp/semisup_100k
+
+stage=0
+
+. utils/parse_options.sh
+
+for f in data/train_sup/utt2spk data/train/utt2spk \
+  data/train_sup/text \
+  data/lang_test_poco_sup100k/G.fst \
+  data/lang_test_poco_sup100k_unk/G.fst
+  ; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find $f"
+    exit 1
+  fi
+done
+
+if [ $stage -le 0 ]; then
+  utils/subset_data_dir.sh --utt-list <(utils/filter_scp.pl --exclude data/train_sup/utt2spk data/train/utt2spk) data/train data/train_unsup100k
+  utils/subset_data_dir.sh --speakers data/train_unsup100k 500000 data/train_unsup100k_500k
+  utils/combine_data.sh data/semisup100k_500k data/train_sup data/train_unsup100k_500k
+fi
+
+###############################################################################
+# Train seed chain system using 100 hours supervised data.
+# Here we train i-vector extractor on only the supervised set.
+###############################################################################
+
+if [ $stage -le 1 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1b.sh \
+    --train-set train_sup \
+    --ivector-train-set "" \
+    --nnet3-affix "" --chain-affix "" \
+    --tdnn-affix _1b --tree-affix bi_a \
+    --gmm tri4a --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Semi-supervised training using 100 hours supervised data and 
+# 250 hours unsupervised data. We use i-vector extractor, tree, lattices 
+# and seed chain system from the previous stage.
+###############################################################################
+
+if [ $stage -le 2 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_100k_semisupervised_1b.sh \
+    --hidden-dim 1536 --cell-dim 1536 --projection-dim 384 \
+    --supervised-set train_sup \
+    --unsupervised-set train_unsup100k_500k \
+    --lm-weights 3,1 \
+    --sup-chain-dir $exp_root/chain/tdnn_lstm_1b_sp \
+    --sup-lat-dir $exp_root/chain/tri4a_train_sup_unk_lats \
+    --sup-tree-dir $exp_root/chain/tree_bi_a \
+    --ivector-root-dir $exp_root/nnet3 \
+    --chain-affix "" \
+    --tdnn-affix _semisup100k_500k_1b \
+    --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Oracle system trained on combined 350 hours including both supervised and 
+# unsupervised sets. We use i-vector extractor, tree, and GMM trained
+# on only the supervised for fair comparison to semi-supervised experiments.
+###############################################################################
+
+if [ $stage -le 3 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1b.sh \
+    --hidden-dim 1536 --cell-dim 1536 --projection-dim 384 \
+    --train-set semisup100k_500k \
+    --nnet3-affix "" --chain-affix "" \
+    --common-treedir $exp_root/chain/tree_bi_a \
+    --tdnn-affix _oracle100k_500k_1b --nj 100 \
+    --gmm tri4a --exp $exp_root \
+    --stage 9 || exit 1
+fi
+
diff --git a/egs/fisher_english/s5/local/semisup/run_lstm_100k_ex250k.sh b/egs/fisher_english/s5/local/semisup/run_lstm_100k_ex250k.sh
new file mode 100644
index 00000000000..1f9ee00dc8b
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/run_lstm_100k_ex250k.sh
@@ -0,0 +1,90 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script demonstrates semi-supervised training using 100 hours of 
+# supervised data and 250 hours of unsupervised data.
+# We assume the supervised data is in data/train_sup and unsupervised data
+# is in data/train_unsup100k_250k. 
+# For LM training, we only use the supervised set corresponding to 100 hours as 
+# opposed to the case in run_50k.sh, where we included part of the 
+# transcripts in data/train/text.
+# This uses only 100 hours supervised set for i-vector extractor training, 
+# which is different from run_50k.sh, which uses combined supervised + 
+# unsupervised set.
+# This script is expected to be run after stage 8 of run_100k.sh.
+
+. ./cmd.sh
+. ./path.sh 
+
+set -o pipefail
+exp_root=exp/semisup_100k
+
+stage=0
+
+. utils/parse_options.sh
+
+for f in data/train_sup/utt2spk data/train_unsup100k_250k/utt2spk \
+  data/train_sup/text \
+  data/lang_test_poco_ex250k/G.fst \
+  data/lang_test_poco_ex250k_unk/G.fst \
+  data/lang_test_poco_ex250k_big/G.carpa \
+  data/lang_test_poco_ex250k_unk_big/G.carp
+  ; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find $f"
+    exit 1
+  fi
+done
+
+
+###############################################################################
+# Train seed chain system using 100 hours supervised data.
+# Here we train i-vector extractor on only the supervised set.
+###############################################################################
+
+if [ $stage -le 1 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1b.sh \
+    --train-set train_sup \
+    --ivector-train-set "" \
+    --nnet3-affix "" --chain-affix "" \
+    --tdnn-affix _1b --tree-affix bi_a \
+    --gmm tri4a --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Semi-supervised training using 100 hours supervised data and 
+# 250 hours unsupervised data. We use i-vector extractor, tree, lattices 
+# and seed chain system from the previous stage.
+###############################################################################
+
+if [ $stage -le 2 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_1a.sh \
+    --supervised-set train_sup \
+    --unsupervised-set train_unsup100k_250k \
+    --sup-chain-dir $exp_root/chain/tdnn_lstm_1b_sp \
+    --sup-lat-dir $exp_root/chain/tri4a_train_sup_unk_lats \
+    --sup-tree-dir $exp_root/chain/tree_bi_a \
+    --ivector-root-dir $exp_root/nnet3 \
+    --chain-affix "" \
+    --tdnn-affix _semisup100k_250k_ex250k_1a \
+    --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Oracle system trained on combined 350 hours including both supervised and 
+# unsupervised sets. We use i-vector extractor, tree, and GMM trained
+# on only the supervised for fair comparison to semi-supervised experiments.
+###############################################################################
+
+if [ $stage -le 3 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1b.sh \
+    --train-set semisup100k_250k \
+    --nnet3-affix "" --chain-affix "" \
+    --common-treedir $exp_root/chain/tree_bi_a \
+    --tdnn-affix 1b_oracle --nj 100 \
+    --gmm tri4a --exp $exp_root \
+    --stage 9 || exit 1
+fi
+
diff --git a/egs/fisher_english/s5/local/semisup/run_lstm_50k.sh b/egs/fisher_english/s5/local/semisup/run_lstm_50k.sh
new file mode 100644
index 00000000000..379090ec029
--- /dev/null
+++ b/egs/fisher_english/s5/local/semisup/run_lstm_50k.sh
@@ -0,0 +1,90 @@
+#!/bin/bash
+
+# Copyright 2017  Vimal Manohar
+# Apache 2.0
+
+# This script demonstrates semi-supervised training using 50 hours of 
+# supervised data and 250 hours of unsupervised data.
+# This script is expected to be run after stage 8 of run_50k.sh.
+
+# We assume the supervised data is in data/train_sup and unsupervised data
+# is in data/train_unsup100k_250k. 
+# For LM training, we assume there is data/train/text, from which
+# we will exclude the utterances contained in the unsupervised set.
+# We use all 300 hours of semi-supervised data for i-vector extractor training.
+
+# This differs from run_lstm_100k.sh, which uses only 100 hours supervised data for 
+# both i-vector extractor training and LM training.
+
+. ./cmd.sh
+. ./path.sh 
+
+set -o pipefail
+exp_root=exp/semisup_50k
+
+stage=0
+
+. utils/parse_options.sh
+
+for f in data/train_sup/utt2spk data/train_unsup100k_250k/utt2spk \
+  data/train/text \
+  data/lang_test_poco_ex250k_big/G.carpa \
+  data/lang_test_poco_ex250k/G.fst \
+  data/lang_test_poco_ex250k_unk_big/G.carpa \
+  data/lang_test_poco_ex250k_unk/G.fst; do
+  if [ ! -f $f ]; then
+    echo "$0: Could not find $f"
+    exit 1
+  fi
+done
+
+###############################################################################
+# Train seed chain system using 50 hours supervised data.
+# Here we train i-vector extractor on combined supervised and unsupervised data
+###############################################################################
+
+if [ $stage -le 1 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1a.sh \
+    --train-set train_sup50k \
+    --ivector-train-set semisup50k_100k_250k \
+    --nnet3-affix _semi50k_100k_250k \
+    --chain-affix _semi50k_100k_250k \
+    --tdnn-affix _1a --tree-affix bi_a \
+    --gmm tri4a --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Semi-supervised training using 50 hours supervised data and 
+# 250 hours unsupervised data. We use i-vector extractor, tree, lattices 
+# and seed chain system from the previous stage.
+###############################################################################
+
+if [ $stage -le 2 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_50k_semisupervised_1a.sh \
+    --supervised-set train_sup50k \
+    --unsupervised-set train_unsup100k_250k \
+    --sup-chain-dir $exp_root/chain_semi50k_100k_250k/tdnn_lstm_1a_sp \
+    --sup-lat-dir $exp_root/chain_semi50k_100k_250k/tri4a_train_sup50k_sp_unk_lats \
+    --sup-tree-dir $exp_root/chain_semi50k_100k_250k/tree_bi_a \
+    --ivector-root-dir $exp_root/nnet3_semi50k_100k_250k \
+    --chain-affix _semi50k_100k_250k \
+    --tdnn-affix _semisup_1a \
+    --exp-root $exp_root || exit 1
+fi
+
+###############################################################################
+# Oracle system trained on combined 300 hours including both supervised and 
+# unsupervised sets. We use i-vector extractor, tree, and GMM trained
+# on only the supervised for fair comparison to semi-supervised experiments.
+###############################################################################
+
+if [ $stage -le 3 ]; then
+  local/semisup/chain/tuning/run_tdnn_lstm_1b.sh \
+    --train-set semisup50k_100k_250k \
+    --nnet3-affix _semi50k_100k_250k \
+    --chain-affix _semi50k_100k_250k \
+    --common-treedir $exp_root/chain_semi50k_100k_250k/tree_bi_a \
+    --tdnn-affix _1b_oracle --nj 100 \
+    --gmm tri4a --exp-root $exp_root \
+    --stage 9 || exit 1
+fi