[WIP] Implement TCPGen contextual biasing

egs/librispeech/ASR/zipformer/biasing.py

@@ -0,0 +1,284 @@
+# Copyright 2024 Xiaomi Corp. (authors: Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple
+
+import math
+import torch
+import torch.nn as nn
+
+from icefall import ContextGraph
+import logging
+import random
+
+
+class TCPGen(nn.Module):
+ def __init__(
+ self,
+ encoder_dim: int,
+ embed_dim: int,
+ joiner_dim: int,
+ decoder: nn.Module,
+ attn_dim: int = 512,
+ tcpgen_dropout: float = 0.1,
+ ):
+ super().__init__()
+ # embedding for ool (out of biasing list) token
+ self.oolemb = torch.nn.Embedding(1, embed_dim)
+ # project symbol embeddings
+ self.q_proj_sym = torch.nn.Linear(embed_dim, attn_dim)
+ # project encoder embeddings
+ self.q_proj_acoustic = torch.nn.Linear(encoder_dim, attn_dim)
+ # project symbol embeddings (vocabulary + ool)
+ self.k_proj = torch.nn.Linear(embed_dim, attn_dim)
+ # generate tcpgen probability
+ self.tcp_gate = torch.nn.Linear(attn_dim + joiner_dim, 1)
+ self.dropout_tcpgen = torch.nn.Dropout(tcpgen_dropout)
+ self.decoder = decoder
+ self.vocab_size = decoder.vocab_size
+
+ def get_tcpgen_masks(
+ self,
+ targets: torch.Tensor,
+ context_graph: ContextGraph,
+ vocab_size: int,
+ blank_id: int = 0,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ batch_size, sql_len = targets.shape
+ dist_masks = torch.ones((batch_size, sql_len, vocab_size + 1))
+ gen_masks = []
+ yseqs = targets.tolist()
+ for i, yseq in enumerate(yseqs):
+ node = context_graph.root
+ gen_mask = []
+ for j, y in enumerate(yseq):
+ not_matched = False
+ if y == blank_id:
+ node = context_graph.root
+ gen_mask.append(0)
+ elif y in node.next:
+ gen_mask.append(0)
+ node = node.next[y]
+ if node.is_end:
+ node = context_graph.root
+ else:
+ gen_mask.append(1)
+ node = context_graph.root
+ not_matched = True
+ # unmask_index = (
+ # [vocab_size]
+ # if node.token == -1
+ # else list(node.next.keys()) + [vocab_size]
+ # )
+ # logging.info(f"token : {node.token}, keys : {node.next.keys()}")
+ # dist_masks[i, j, unmask_index] = 0
+ if not not_matched:
+ dist_masks[i, j, list(node.next.keys())] = 0
+
+ gen_masks.append(gen_mask + [1] * (sql_len - len(gen_mask)))
+ gen_masks = torch.Tensor(gen_masks).to(targets.device).bool()
+ dist_masks = dist_masks.to(targets.device).bool()
+ if random.random() >= 0.95:
+ logging.info(
+ f"gen_mask nonzero {gen_masks.shape} : {torch.count_nonzero(torch.logical_not(gen_masks), dim=1)}"
+ )
+ logging.info(
+ f"dist_masks nonzero {dist_masks.shape} : {torch.count_nonzero(torch.logical_not(dist_masks), dim=2)}"
+ )
+ return dist_masks, gen_masks
+
+ def get_tcpgen_distribution(
+ self, query: torch.Tensor, dist_masks: torch.Tensor
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ """
+ Args:
+ query:
+ shape : (B, T, s_range, attn_dim)
+ dist_masks:
+ shape : (B, T, s_range, V + 1)
+ """
+ # From original paper, k, v share same embeddings
+ # (V + 1, embed_dim)
+ kv = torch.cat([self.decoder.embedding.weight.data, self.oolemb.weight], dim=0)
+ # (V + 1, attn_dim)
+ kv = self.dropout_tcpgen(self.k_proj(kv))
+ # (B, T, s_range, attn_dim) * (attn_dim, V + 1) -> (B, T, s_range, V + 1)
+ distribution = torch.matmul(query, kv.permute(1, 0)) / math.sqrt(query.size(-1))
+ distribution = distribution.masked_fill(
+ dist_masks, torch.finfo(distribution.dtype).min
+ )
+ distribution = distribution.softmax(dim=-1)
+ # (B, T, s_range, V) * (V, attn_dim) -> (B, T, s_range, attn_dim)
+ # logging.info(f"distribution shape : {distribution.shape}")
+ hptr = torch.matmul(distribution[:, :, :, :-1], kv[:-1, :])
+ hptr = self.dropout_tcpgen(hptr)
+ if random.random() > 0.95:
+ logging.info(
+ f"distribution mean : {torch.mean(distribution, dim=3)}, std: {torch.std(distribution, dim=3)}"
+ )
+ logging.info(
+ f"distribution min : {torch.min(distribution, dim=3)}, max: {torch.max(distribution, dim=3)}"
+ )
+ return hptr, distribution
+
+ def prune_query_and_mask(
+ self,
+ query_sym: torch.Tensor,
+ query_acoustic: torch.Tensor,
+ dist_masks: torch.Tensor,
+ gen_masks: torch.Tensor,
+ ranges: torch.Tensor,
+ ):
+ """Prune the queries from symbols and acoustics with ranges
+ generated by `get_rnnt_prune_ranges` in pruned rnnt loss.
+
+ Args:
+ query_sym:
+ The symbol query, with shape (B, S, attn_dim).
+ query_acoustic:
+ The acoustic query, with shape (B, T, attn_dim).
+ dist_masks:
+ The TCPGen distribution masks, with shape (B, S, V + 1).
+ gen_masks:
+ The TCPGen probability masks, with shape (B, S).
+ ranges:
+ A tensor containing the symbol indexes for each frame that we want to
+ keep. Its shape is (B, T, s_range), see the docs in
+ `get_rnnt_prune_ranges` in rnnt_loss.py for more details of this tensor.
+
+ Returns:
+ Return the pruned query with the shape (B, T, s_range, attn_dim).
+ """
+ assert ranges.shape[0] == query_sym.shape[0], (ranges.shape, query_sym.shape)
+ assert ranges.shape[0] == query_acoustic.shape[0], (
+ ranges.shape,
+ query_acoustic.shape,
+ )
+ assert query_acoustic.shape[1] == ranges.shape[1], (
+ query_acoustic.shape,
+ ranges.shape,
+ )
+ (B, T, s_range) = ranges.shape
+ (B, S, attn_dim) = query_sym.shape
+ assert query_acoustic.shape == (B, T, attn_dim), (
+ query_acoustic.shape,
+ (B, T, attn_dim),
+ )
+ assert dist_masks.shape == (B, S, self.vocab_size + 1), (
+ dist_masks.shape,
+ (B, S, self.vocab_size + 1),
+ )
+ assert gen_masks.shape == (B, S), (gen_masks.shape, (B, S))
+ # (B, T, s_range, attn_dim)
+ query_acoustic_pruned = query_acoustic.unsqueeze(2).expand(
+ (B, T, s_range, attn_dim)
+ )
+ # logging.info(f"query_sym : {query_sym.shape}")
+ # logging.info(f"ranges : {ranges}")
+ # (B, T, s_range, attn_dim)
+ query_sym_pruned = torch.gather(
+ query_sym,
+ dim=1,
+ index=ranges.reshape(B, T * s_range, 1).expand((B, T * s_range, attn_dim)),
+ ).reshape(B, T, s_range, attn_dim)
+ # (B, T, s_range, V + 1)
+ dist_masks_pruned = torch.gather(
+ dist_masks,
+ dim=1,
+ index=ranges.reshape(B, T * s_range, 1).expand(
+ (B, T * s_range, self.vocab_size + 1)
+ ),
+ ).reshape(B, T, s_range, self.vocab_size + 1)
+ # (B, T, s_range)
+ gen_masks_pruned = torch.gather(
+ gen_masks, dim=1, index=ranges.reshape(B, T * s_range)
+ ).reshape(B, T, s_range)
+ return (
+ query_sym_pruned + query_acoustic_pruned,
+ dist_masks_pruned,
+ gen_masks_pruned,
+ )
+
+ def forward(
+ self,
+ targets: torch.Tensor,
+ encoder_embeddings: torch.Tensor,
+ ranges: torch.Tensor,
+ context_graph: ContextGraph,
+ ):
+ """
+ Args:
+ target:
+ The training targets in token ids (padded with blanks). shape : (B, S)
+ encoder_embeddings:
+ The encoder outputs. shape: (B, T, attn_dim)
+ ranges:
+ The prune ranges from pruned rnnt. shape: (B, T, s_range)
+ context_graphs:
+ The context_graphs for each utterance. B == len(context_graphs).
+
+ Return:
+ returns tcpgen embedding with shape (B, T, s_range, attn_dim) and
+ tcpgen distribution with shape (B, T, s_range, V + 1).
+ """
+ query_sym = self.decoder.embedding(targets)
+
+ query_sym = self.q_proj_sym(query_sym) # (B, S, attn_dim)
+ query_acoustic = self.q_proj_acoustic(encoder_embeddings) # (B , T, attn_dim)
+
+ # dist_masks : (B, S, V + 1)
+ # gen_masks : (B, S)
+ dist_masks, gen_masks = self.get_tcpgen_masks(
+ targets=targets, context_graph=context_graph, vocab_size=self.vocab_size
+ )
+ # query : (B, T, s_range, attn_dim)
+ # dist_masks : (B, T, s_range, V + 1)
+ query, dist_masks, gen_masks = self.prune_query_and_mask(
+ query_sym=query_sym,
+ query_acoustic=query_acoustic,
+ dist_masks=dist_masks,
+ gen_masks=gen_masks,
+ ranges=ranges,
+ )
+
+ if random.random() >= 0.95:
+ logging.info(
+ f"pruned gen_mask nonzero {gen_masks.shape} : {torch.count_nonzero(torch.logical_not(gen_masks), dim=1)}"
+ )
+ logging.info(
+ f"pruned dist_masks nonzero {dist_masks.shape} : {torch.count_nonzero(torch.logical_not(dist_masks), dim=3)}"
+ )
+
+ # hptr : (B, T, s_range, attn_dim)
+ # tcpgen_dist : (B, T, s_range, V + 1)
+ hptr, tcpgen_dist = self.get_tcpgen_distribution(query, dist_masks)
+ return hptr, tcpgen_dist, gen_masks
+
+ def generator_prob(
+ self, hptr: torch.Tensor, h_joiner: torch.Tensor, gen_masks: torch.Tensor
+ ) -> torch.Tensor:
+ # tcpgen_prob : (B, T, s_range, 1)
+ tcpgen_prob = self.tcp_gate(torch.cat((h_joiner, hptr), dim=-1))
+ tcpgen_prob = torch.sigmoid(tcpgen_prob)
+ tcpgen_prob = tcpgen_prob.masked_fill(gen_masks.unsqueeze(-1), 0)
+ if random.random() >= 0.95:
+ logging.info(
+ f"tcpgen_prob mean : {torch.mean(tcpgen_prob.squeeze(-1), dim=(1,2))}, std : {torch.std(tcpgen_prob.squeeze(-1), dim=(1, 2))}"
+ )
+ logging.info(
+ f"tcpgen_prob min : {torch.min(tcpgen_prob.squeeze(-1), dim=1)}, max : {torch.max(tcpgen_prob.squeeze(-1), dim=1)}"
+ )
+ return tcpgen_prob

egs/librispeech/ASR/zipformer/joiner.py

@@ -17,6 +17,7 @@
 import torch
 import torch.nn as nn
 from scaling import ScaledLinear
+from typing import Optional
 
 
 class Joiner(nn.Module):
@@ -37,6 +38,7 @@ def forward(
  self,
  encoder_out: torch.Tensor,
  decoder_out: torch.Tensor,
+ tcpgen_hptr: Optional[torch.Tensor] = None,
  project_input: bool = True,
  ) -> torch.Tensor:
  """
@@ -62,6 +64,11 @@ def forward(
  else:
  logit = encoder_out + decoder_out
 
- logit = self.output_linear(torch.tanh(logit))
+ if tcpgen_hptr is not None:
+ logit += tcpgen_hptr
 
- return logit
+ activations = torch.tanh(logit)
+
+ logit = self.output_linear(activations)
+
+ return logit if tcpgen_hptr is None else (logit, activations)