Add modified beam search.

egs/librispeech/ASR/transducer_stateless/beam_search.py

@@ -17,7 +17,6 @@
 from dataclasses import dataclass
 from typing import Dict, List, Optional
 
-import numpy as np
 import torch
 from model import Transducer
 
@@ -108,8 +107,11 @@ class Hypothesis:
     # Newly predicted tokens are appended to `ys`.
     ys: List[int]
 
-    # The log prob of ys
+    # The log prob of ys.
-    log_prob: float
+    # It contains only one entry.
+    # TODO(fangjun): It was a float before. We need to change its usage
+    # in greedy_search and beam_search.
+    log_prob: torch.Tensor
 
     @property
     def key(self) -> str:
@@ -145,8 +147,10 @@ class HypothesisList(object):
         """
         key = hyp.key
         if key in self:
-            old_hyp = self._data[key]
+            old_hyp = self._data[key]  # shallow copy
-            old_hyp.log_prob = np.logaddexp(old_hyp.log_prob, hyp.log_prob)
+            torch.logaddexp(
+                old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob
+            )
         else:
             self._data[key] = hyp
 
@@ -348,47 +352,70 @@ def modified_beam_search(
     T = encoder_out.size(1)
 
     B = HypothesisList()
-    B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0))
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
 
     encoder_out_len = torch.tensor([1])
     decoder_out_len = torch.tensor([1])
 
-    decoder_cache: Dict[str, torch.Tensor] = {}
     for t in range(T):
         # fmt: off
         current_encoder_out = encoder_out[:, t:t+1, :]
+        # current_encoder_out is of shape (1, 1, encoder_out_dim)
         # fmt: on
-        A = B
+        A = list(B)
         B = HypothesisList()
 
-        joint_cache: Dict[str, torch.Tensor] = {}
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
 
-        for hyp in A:
+        decoder_input = torch.tensor(
-            decoder_out = run_decoder(
+            [hyp.ys[-context_size:] for hyp in A],
-                ys=hyp.ys, model=model, decoder_cache=decoder_cache
+            device=device,
         )
-            key = "_".join(map(str, hyp.ys[-context_size:]))
+        # decoder_input is of shape (num_hyps, context_size)
-            key += f"-t-{t}"
-            log_prob = run_joiner(
-                key=key,
-                model=model,
-                encoder_out=current_encoder_out,
-                decoder_out=decoder_out,
-                encoder_out_len=encoder_out_len,
-                decoder_out_len=decoder_out_len,
-                joint_cache=joint_cache,
-            )
-            log_prob = log_prob.cpu().tolist()
 
-            for i, v in enumerate(log_prob):
+        decoder_out = model.decoder(decoder_input, need_pad=False)
-                if i == blank_id:
+        # decoder_output is of shape (num_hyps, 1, decoder_output_dim)
-                    # Use [:] to make a copy
+
+        current_encoder_out = current_encoder_out.expand(
+            decoder_out.size(0), 1, -1
+        )
+
+        logits = model.joiner(
+            current_encoder_out,
+            decoder_out,
+            encoder_out_len.expand(decoder_out.size(0)),
+            decoder_out_len.expand(decoder_out.size(0)),
+        )
+        # logits is of shape (num_hyps, vocab_size)
+        log_probs = logits.log_softmax(dim=-1)
+
+        log_probs.add_(ys_log_probs)
+
+        log_probs = log_probs.reshape(-1)
+        topk_log_probs, topk_indexes = log_probs.topk(beam)
+
+        # topk_hyp_indexes are indexes into `A`
+        topk_hyp_indexes = topk_indexes // logits.size(-1)
+        topk_token_indexes = topk_indexes % logits.size(-1)
+
+        topk_hyp_indexes = topk_hyp_indexes.tolist()
+        topk_token_indexes = topk_token_indexes.tolist()
+
+        for i in range(len(topk_hyp_indexes)):
+            hyp = A[topk_hyp_indexes[i]]
             new_ys = hyp.ys[:]
-                else:
+            new_token = topk_token_indexes[i]
-                    new_ys = hyp.ys + [i]
+            if new_token != blank_id:
-                new_hyp = Hypothesis(ys=new_ys, log_prob=hyp.log_prob + v)
+                new_ys.append(new_token)
+            new_log_prob = topk_log_probs[i]
+            new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
             B.add(new_hyp)
-        B = B.topk(beam)
 
     best_hyp = B.get_most_probable(length_norm=True)
     ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks

egs/librispeech/ASR/transducer_stateless/decoder.py

@@ -75,24 +75,24 @@ class Decoder(nn.Module):
         """
         Args:
           y:
-            A 2-D tensor of shape (N, U) with blank prepended.
+            A 2-D tensor of shape (N, U).
           need_pad:
             True to left pad the input. Should be True during training.
             False to not pad the input. Should be False during inference.
         Returns:
           Return a tensor of shape (N, U, embedding_dim).
         """
-        embeding_out = self.embedding(y)
+        embedding_out = self.embedding(y)
         if self.context_size > 1:
-            embeding_out = embeding_out.permute(0, 2, 1)
+            embedding_out = embedding_out.permute(0, 2, 1)
             if need_pad is True:
-                embeding_out = F.pad(
+                embedding_out = F.pad(
-                    embeding_out, pad=(self.context_size - 1, 0)
+                    embedding_out, pad=(self.context_size - 1, 0)
                 )
             else:
                 # During inference time, there is no need to do extra padding
                 # as we only need one output
-                assert embeding_out.size(-1) == self.context_size
+                assert embedding_out.size(-1) == self.context_size
-            embeding_out = self.conv(embeding_out)
+            embedding_out = self.conv(embedding_out)
-            embeding_out = embeding_out.permute(0, 2, 1)
+            embedding_out = embedding_out.permute(0, 2, 1)
-        return embeding_out
+        return embedding_out