adds an option to include EOS in transducer model training;

rewrites masked_copy_cached_state() to make it clearer and more general;
code adaptation/changes according to the commits on Nov 2, 2022

Author: freewym

espresso/criterions/ctc_loss.py

@@ -12,10 +12,11 @@
 import torch.nn.functional as F
 from omegaconf import II
 
-from fairseq import metrics, utils
+from fairseq import utils
 from fairseq.criterions import FairseqCriterion, register_criterion
 from fairseq.data import data_utils
 from fairseq.dataclass import FairseqDataclass
+from fairseq.logging import metrics
 from fairseq.tasks import FairseqTask
 
 logger = logging.getLogger(__name__)

espresso/criterions/transducer_loss.py

@@ -11,10 +11,11 @@
 import torch
 from omegaconf import II
 
-from fairseq import metrics, utils
+from fairseq import utils
 from fairseq.criterions import FairseqCriterion, register_criterion
 from fairseq.data import data_utils
 from fairseq.dataclass import ChoiceEnum, FairseqDataclass
+from fairseq.logging import metrics
 from fairseq.tasks import FairseqTask
 
 logger = logging.getLogger(__name__)
@@ -36,6 +37,7 @@ class TransducerLossCriterionConfig(FairseqDataclass):
         default="torchaudio",
         metadata={"help": "choice of loss backend (native or torchaudio)"},
     )
+    include_eos: bool = II("task.include_eos_in_transducer_loss")
 
 
 @register_criterion("transducer_loss", dataclass=TransducerLossCriterionConfig)
@@ -64,6 +66,7 @@ def __init__(self, cfg: TransducerLossCriterionConfig, task: FairseqTask):
                 )
             self.rnnt_loss = rnnt_loss
 
+        self.include_eos = cfg.include_eos
         self.dictionary = task.target_dictionary
         self.prev_num_updates = -1
 
@@ -73,13 +76,15 @@ def forward(self, model, sample, reduce=True):
         )  # B x T x U x V, B
 
         if "target_lengths" in sample:
-            target_lengths = (
-                sample["target_lengths"].int() - 1
-            )  # Note: ensure EOS is excluded
+            target_lengths = sample["target_lengths"].int()
+            if not self.include_eos:
+                target_lengths -= 1  # excludes EOS
         else:
             target_lengths = (
                 (
                     (sample["target"] != self.pad_idx)
+                    if self.include_eos
+                    else (sample["target"] != self.pad_idx)
                     & (sample["target"] != self.eos_idx)
                 )
                 .sum(-1)
@@ -124,7 +129,9 @@ def forward(self, model, sample, reduce=True):
 
         loss = self.rnnt_loss(
             net_output,
-            sample["target"][:, :-1].int().contiguous(),  # exclude the last EOS column
+            (sample["target"] if self.include_eos else sample["target"][:, :-1])
+            .int()
+            .contiguous(),
             encoder_out_lengths.int(),
             target_lengths,
             blank=self.blank_idx,

espresso/data/asr_dataset.py

@@ -21,6 +21,7 @@ def collate(
     left_pad_source=True,
     left_pad_target=False,
     input_feeding=True,
+    maybe_bos_idx=None,
     pad_to_length=None,
     pad_to_multiple=1,
     src_bucketed=False,
@@ -89,11 +90,16 @@ def merge(key, left_pad, move_eos_to_beginning=False, pad_to_length=None):
             prev_output_tokens = merge(
                 "target",
                 left_pad=left_pad_target,
-                move_eos_to_beginning=True,
+                move_eos_to_beginning=(maybe_bos_idx is None),
                 pad_to_length=pad_to_length["target"]
                 if pad_to_length is not None
                 else None,
             )
+            if maybe_bos_idx is not None:
+                all_bos_vec = prev_output_tokens.new_full((1, 1), maybe_bos_idx).expand(
+                    len(samples), 1
+                )
+                prev_output_tokens = torch.cat([all_bos_vec, prev_output_tokens], dim=1)
     else:
         ntokens = src_lengths.sum().item()
 
@@ -148,6 +154,10 @@ class AsrDataset(FairseqDataset):
             (default: True).
         input_feeding (bool, optional): create a shifted version of the targets
             to be passed into the model for teacher forcing (default: True).
+        prepend_bos_as_input_feeding (bool, optional): target prepended with BOS symbol
+            (instead of moving EOS to the beginning of that) as input feeding. This is
+            currently only for a transducer model training setting where EOS is retained
+            in target when evaluating the loss (default: False).
         constraints (Tensor, optional): 2d tensor with a concatenated, zero-
             delimited list of constraints for each sentence.
         num_buckets (int, optional): if set to a value greater than 0, then
@@ -176,6 +186,7 @@ def __init__(
         left_pad_target=False,
         shuffle=True,
         input_feeding=True,
+        prepend_bos_as_input_feeding=False,
         constraints=None,
         num_buckets=0,
         src_lang_id=None,
@@ -193,6 +204,7 @@ def __init__(
         self.left_pad_target = left_pad_target
         self.shuffle = shuffle
         self.input_feeding = input_feeding
+        self.prepend_bos_as_input_feeding = prepend_bos_as_input_feeding
         self.constraints = constraints
         self.src_lang_id = src_lang_id
         self.tgt_lang_id = tgt_lang_id
@@ -334,6 +346,9 @@ def collater(self, samples, pad_to_length=None):
             left_pad_source=self.left_pad_source,
             left_pad_target=self.left_pad_target,
             input_feeding=self.input_feeding,
+            maybe_bos_idx=self.dictionary.bos()
+            if self.prepend_bos_as_input_feeding
+            else None,
             pad_to_length=pad_to_length,
             pad_to_multiple=self.pad_to_multiple,
             src_bucketed=(self.buckets is not None),

espresso/data/feat_text_dataset.py

@@ -57,7 +57,7 @@ def __init__(
     ):
         super().__init__()
         assert len(utt_ids) == len(rxfiles)
-        self.dtype = np.float
+        self.dtype = float
         self.utt_ids = utt_ids
         self.rxfiles = rxfiles
         self.size = len(utt_ids)  # number of utterances
@@ -338,7 +338,6 @@ def __init__(
         self, utt_ids: List[str], texts: List[str], dictionary=None, append_eos=True
     ):
         super().__init__()
-        self.dtype = np.float
         self.dictionary = dictionary
         self.append_eos = append_eos
         self.read_text(utt_ids, texts, dictionary)

espresso/models/speech_lstm.py

@@ -1017,28 +1017,16 @@ def masked_copy_cached_state(
             src_cached_state[2],
         )
 
-        def masked_copy_state(state: Optional[Tensor], src_state: Optional[Tensor]):
-            if state is None:
-                assert src_state is None
-                return None
-            else:
-                assert (
-                    state.size(0) == mask.size(0)
-                    and src_state is not None
-                    and state.size() == src_state.size()
-                )
-                state[mask, ...] = src_state[mask, ...]
-                return state
-
-        prev_hiddens = [
-            masked_copy_state(p, src_p)
-            for (p, src_p) in zip(prev_hiddens, src_prev_hiddens)
-        ]
-        prev_cells = [
-            masked_copy_state(p, src_p)
-            for (p, src_p) in zip(prev_cells, src_prev_cells)
-        ]
-        input_feed = masked_copy_state(input_feed, src_input_feed)
+        mask = mask.unsqueeze(1)
+        prev_hiddens = speech_utils.apply_to_sample_pair(
+            lambda x, y, z=mask: torch.where(z, x, y), src_prev_hiddens, prev_hiddens
+        )
+        prev_cells = speech_utils.apply_to_sample_pair(
+            lambda x, y, z=mask: torch.where(z, x, y), src_prev_cells, prev_cells
+        )
+        input_feed = speech_utils.apply_to_sample_pair(
+            lambda x, y, z=mask: torch.where(z, x, y), src_input_feed, input_feed
+        )
 
         cached_state_new = torch.jit.annotate(
             Dict[str, Optional[Tensor]],

espresso/models/transformer/speech_transformer_decoder.py

@@ -12,6 +12,7 @@
 import torch.nn.functional as F
 from torch import Tensor
 
+import espresso.tools.utils as speech_utils
 from espresso.models.transformer import SpeechTransformerConfig
 from espresso.modules import (
     RelativePositionalEmbedding,
@@ -430,34 +431,23 @@ def masked_copy_cached_state(
                 F.pad(src_p, (0, 1)) for src_p in src_prev_key_padding_mask
             ]
 
-        def masked_copy_state(state: Optional[Tensor], src_state: Optional[Tensor]):
-            if state is None:
-                assert src_state is None
-                return None
-            else:
-                assert (
-                    state.size(0) == mask.size(0)
-                    and src_state is not None
-                    and state.size() == src_state.size()
-                )
-                state[mask, ...] = src_state[mask, ...]
-                return state
-
-        prev_key = [
-            masked_copy_state(p, src_p) for (p, src_p) in zip(prev_key, src_prev_key)
-        ]
-        prev_value = [
-            masked_copy_state(p, src_p)
-            for (p, src_p) in zip(prev_value, src_prev_value)
-        ]
+        kv_mask = mask.unsqueeze(1).unsqueeze(2).unsqueeze(3)
+        prev_key = speech_utils.apply_to_sample_pair(
+            lambda x, y, z=kv_mask: torch.where(z, x, y), src_prev_key, prev_key
+        )
+        prev_value = speech_utils.apply_to_sample_pair(
+            lambda x, y, z=kv_mask: torch.where(z, x, y), src_prev_value, prev_value
+        )
         if prev_key_padding_mask is None:
             prev_key_padding_mask = src_prev_key_padding_mask
         else:
             assert src_prev_key_padding_mask is not None
-            prev_key_padding_mask = [
-                masked_copy_state(p, src_p)
-                for (p, src_p) in zip(prev_key_padding_mask, src_prev_key_padding_mask)
-            ]
+            pad_mask = mask.unsqueeze(1)
+            prev_key_padding_mask = speech_utils.apply_to_sample_pair(
+                lambda x, y, z=pad_mask: torch.where(z, x, y),
+                src_prev_key_padding_mask,
+                prev_key_padding_mask,
+            )
 
         cached_state = torch.jit.annotate(
             Dict[str, Optional[Tensor]],

espresso/models/transformer/speech_transformer_encoder.py

@@ -9,6 +9,7 @@
 
 import torch
 import torch.nn as nn
+from torch import Tensor
 
 import espresso.tools.utils as speech_utils
 from espresso.models.transformer import SpeechTransformerConfig
@@ -314,7 +315,9 @@ def forward_scriptable(
                 src_tokens,
                 ~speech_utils.sequence_mask(src_lengths, src_tokens.size(1)),
             )
-        has_pads = src_tokens.device.type == "xla" or encoder_padding_mask.any()
+        has_pads: Tensor = (
+            torch.tensor(src_tokens.device.type == "xla") or encoder_padding_mask.any()
+        )
 
         if self.fc0 is not None:
             x = self.dropout_module(x)
@@ -330,8 +333,9 @@ def forward_scriptable(
             x = self.quant_noise(x)
 
         # account for padding while computing the representation
-        if has_pads:
-            x = x * (1 - encoder_padding_mask.unsqueeze(-1).type_as(x))
+        x = x * (
+            1 - encoder_padding_mask.unsqueeze(-1).type_as(x) * has_pads.type_as(x)
+        )
 
         # B x T x C -> T x B x C
         x = x.transpose(0, 1)