run.sh

# Runs all experiments in the paper
# "Cross-domain Neural Pitch and Periodicity Estimation"

# Args
# $1 - index of GPU to use

# Download datasets
python -m penn.data.download

# Setup and run 16 kHz experiments
python -m penn.data.preprocess --config config/crepe.py
python -m penn.partition
python -m penn.train --config config/crepe.py --gpu $1
python -m penn.train --config config/deepf0.py --gpu $1

# Evaluate baselines at 16 kHz
python -m penn.evaluate --gpu $1 --config config/dio.py
python -m penn.evaluate --gpu $1 --config config/pyin.py
python -m penn.evaluate \
    --gpu $1 \
    --method torchcrepe \
    --config config/torchcrepe.py

# Setup 8 kHz data
python -m penn.data.preprocess

# Run 8 kHz experiments
python -m penn.train --config config/crepe++.py --gpu $1
python -m penn.train --config config/deepf0++.py --gpu $1
python -m penn.train --config config/fcnf0.py --gpu $1
python -m penn.train --config config/fcnf0++.py --gpu $1

# Train on individual datasets
python -m penn.train \
    --config config/fcnf0++-mdb.py \
    --datasets mdb \
    --gpu $1
python -m penn.train \
    --config config/fcnf0++-ptdb.py \
    --datasets ptdb \
    --gpu $1

# Train ablations
python -m penn.train --config config/fcnf0++-ablate-batchsize.py --gpu $1
python -m penn.train --config config/fcnf0++-ablate-earlystop.py --gpu $1
python -m penn.train --config config/fcnf0++-ablate-inputnorm.py --gpu $1
python -m penn.train --config config/fcnf0++-ablate-layernorm.py --gpu $1
python -m penn.train --config config/fcnf0++-ablate-loss.py --gpu $1
python -m penn.train --config config/fcnf0++-ablate-quantization.py --gpu $1
python -m penn.train --config config/fcnf0++-ablate-unvoiced.py --gpu $1

# Evaluate locally normal decoding
python -m penn.evaluate \
    --config config/fcnf0++-ablate-decoder.py \
    --checkpoint runs/fcnf0++/00250000.pt \
    --gpu $1

# Plot data and inference distributions
python -m penn.plot.density \
    --datasets mdb \
    --output_file results/mdb_on_mdb.pdf \
    --checkpoint runs/fcnf0++-mdb/00250000.pt \
    --gpu $1
python -m penn.plot.density \
    --datasets ptdb \
    --output_file results/mdb_on_ptdb.pdf \
    --checkpoint runs/fcnf0++-mdb/00250000.pt \
    --gpu $1
python -m penn.plot.density \
    --datasets ptdb \
    --output_file results/ptdb_on_ptdb.pdf \
    --checkpoint runs/fcnf0++-ptdb/00250000.pt \
    --gpu $1
python -m penn.plot.density \
    --datasets mdb \
    --output_file results/ptdb_on_mdb.pdf \
    --checkpoint runs/fcnf0++-ptdb/00250000.pt \
    --gpu $1
python -m penn.plot.density \
    --datasets mdb \
    --output_file results/both_on_mdb.pdf \
    --checkpoint runs/fcnf0++/00250000.pt \
    --gpu $1
python -m penn.plot.density \
    --datasets ptdb \
    --output_file results/both_on_ptdb.pdf \
    --checkpoint runs/fcnf0++/00250000.pt \
    --gpu $1

# Plot voiced/unvoiced threshold landscape
python -m penn.plot.threshold \
    --names FCNF0++ "FCNF0++ (voiced only)" \
    --evaluations fcnf0++-ablate-decoder fcnf0++-ablate-unvoiced \
    --output_file results/threshold.pdf

# Plot pitch posteriorgram figures
python -m penn.plot.logits \
    --config config/fcnf0++.py \
    --audio_file test/assets/gershwin.wav \
    --output_file results/fcnf0++-gershwin.pdf \
    --checkpoint runs/fcnf0++/00250000.pt \
    --gpu $1
# Note - You will need to replace this checkpoint with the final checkpoint
#        that was produced during fcnf0 training
python -m penn.plot.logits \
    --config config/fcnf0.py \
    --audio_file test/assets/gershwin.wav \
    --output_file results/fcnf0-gershwin.pdf \
    --checkpoint runs/fcnf0/00071000.pt \
    --gpu $1