Add modified beam search in batch mode.

egs/librispeech/ASR/pruned_transducer_stateless/pretrained.py

@@ -127,7 +127,7 @@ def get_parser():
     parser.add_argument(
         "--max-sym-per-frame",
         type=int,
-        default=3,
+        default=1,
         help="""Maximum number of symbols per frame. Used only when
         --method is greedy_search.
         """,

egs/librispeech/ASR/transducer_stateless/beam_search.py

@@ -17,6 +17,7 @@
 from dataclasses import dataclass
 from typing import Dict, List, Optional
 
+import k2
 import torch
 from model import Transducer
 
@@ -385,13 +386,158 @@ def run_joiner(
     return log_prob
 
 
+def _get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape:
+    """Return a ragged shape with axes [utt][num_hyps].
+
+    Args:
+      hyps:
+        len(hyps) == batch_size. It contains the current hypothesis for
+        each utterance in the batch.
+    Returns:
+      Return a ragged shape with 2 axes [utt][num_hyps]. Note that
+      the shape is on CPU.
+    """
+    num_hyps = [len(h) for h in hyps]
+
+    # torch.cumsum() is inclusive sum, so we put a 0 at the beginning
+    # to get exclusive sum later.
+    num_hyps.insert(0, 0)
+
+    num_hyps = torch.tensor(num_hyps)
+    row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32)
+    ans = k2.ragged.create_ragged_shape2(
+        row_splits=row_splits, cached_tot_size=row_splits[-1].item()
+    )
+    return ans
+
+
 def modified_beam_search(
     model: Transducer,
     encoder_out: torch.Tensor,
     beam: int = 4,
+) -> List[List[int]]:
+    """Beam search in batch mode with --max-sym-per-frame=1 being hardcodded.
+
+    Args:
+      model:
+        The transducer model.
+      encoder_out:
+        Output from the encoder. Its shape is (N, T, C).
+      beam:
+        Number of active paths during the beam search.
+    Returns:
+      Return a list-of-list of token IDs. ans[i] is the decoding results
+      for the i-th utterance.
+    """
+    assert encoder_out.ndim == 3, encoder_out.shape
+
+    batch_size = encoder_out.size(0)
+    T = encoder_out.size(1)
+
+    blank_id = model.decoder.blank_id
+    context_size = model.decoder.context_size
+    device = model.device
+    B = [HypothesisList() for _ in range(batch_size)]
+    for i in range(batch_size):
+        B[i].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])
+    for t in range(T):
+        current_encoder_out = encoder_out[:, t : t + 1, :]  # noqa
+        # current_encoder_out's shape is: (batch_size, 1, encoder_out_dim)
+
+        hyps_shape = _get_hyps_shape(B).to(device)
+
+        A = [list(b) for b in B]
+        B = [HypothesisList() for _ in range(batch_size)]
+
+        ys_log_probs = torch.cat(
+            [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps]
+        )  # (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyps in A for hyp in hyps],
+            device=device,
+            dtype=torch.int64,
+        )  # (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False)
+        # decoder_output is of shape (num_hyps, 1, decoder_output_dim)
+
+        # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor
+        # as index, so we use `to(torch.int64)` below.
+        current_encoder_out = torch.index_select(
+            current_encoder_out,
+            dim=0,
+            index=hyps_shape.row_ids(1).to(torch.int64),
+        )  # (num_hyps, 1, encoder_out_dim)
+
+        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)  # (num_hyps, vocab_size)
+
+        log_probs.add_(ys_log_probs)
+
+        vocab_size = log_probs.size(-1)
+
+        log_probs = log_probs.reshape(-1)
+
+        row_splits = hyps_shape.row_splits(1) * vocab_size
+        log_probs_shape = k2.ragged.create_ragged_shape2(
+            row_splits=row_splits, cached_tot_size=log_probs.numel()
+        )
+        ragged_log_probs = k2.RaggedTensor(
+            shape=log_probs_shape, value=log_probs
+        )
+
+        for i in range(batch_size):
+            topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
+
+            topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
+            topk_token_indexes = (topk_indexes % vocab_size).tolist()
+
+            for k in range(len(topk_hyp_indexes)):
+                hyp_idx = topk_hyp_indexes[k]
+                hyp = A[i][hyp_idx]
+
+                new_ys = hyp.ys[:]
+                new_token = topk_token_indexes[k]
+                if new_token != blank_id:
+                    new_ys.append(new_token)
+
+                new_log_prob = topk_log_probs[k]
+                new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob)
+                B[i].add(new_hyp)
+
+    best_hyps = [b.get_most_probable(length_norm=True) for b in B]
+    ans = [h.ys[context_size:] for h in best_hyps]
+
+    return ans
+
+
+def _deprecated_modified_beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
 ) -> List[int]:
     """It limits the maximum number of symbols per frame to 1.
 
+    It decodes only one utterance at a time. We keep it only for reference.
+    The function :func:`modified_beam_search` should be preferred as it
+    supports batch decoding.
+
     Args:
       model:
         An instance of `Transducer`.

egs/librispeech/ASR/transducer_stateless/decode.py

@@ -198,6 +198,12 @@ def decode_one_batch(
             model=model,
             encoder_out=encoder_out,
         )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
     else:
         batch_size = encoder_out.size(0)
         for i in range(batch_size):
@@ -216,12 +222,6 @@ def decode_one_batch(
                     encoder_out=encoder_out_i,
                     beam=params.beam_size,
                 )
-            elif params.decoding_method == "modified_beam_search":
-                hyp = modified_beam_search(
-                    model=model,
-                    encoder_out=encoder_out_i,
-                    beam=params.beam_size,
-                )
             else:
                 raise ValueError(
                     f"Unsupported decoding method: {params.decoding_method}"

egs/librispeech/ASR/transducer_stateless/pretrained.py

@@ -61,14 +61,15 @@ import sentencepiece as spm
 import torch
 import torch.nn as nn
 import torchaudio
-from beam_search import beam_search, greedy_search, modified_beam_search
+from beam_search import (
-from conformer import Conformer
+    beam_search,
-from decoder import Decoder
+    greedy_search,
-from joiner import Joiner
+    greedy_search_batch,
-from model import Transducer
+    modified_beam_search,
+)
 from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
 
-from icefall.env import get_env_info
 from icefall.utils import AttributeDict
 
 
@@ -115,6 +116,13 @@ def get_parser():
         "The sample rate has to be 16kHz.",
     )
 
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
     parser.add_argument(
         "--beam-size",
         type=int,
@@ -132,7 +140,7 @@ def get_parser():
     parser.add_argument(
         "--max-sym-per-frame",
         type=int,
-        default=3,
+        default=1,
         help="""Maximum number of symbols per frame. Used only when
         --method is greedy_search.
         """,
@@ -141,70 +149,6 @@ def get_parser():
     return parser
 
 
-def get_params() -> AttributeDict:
-    params = AttributeDict(
-        {
-            "sample_rate": 16000,
-            # parameters for conformer
-            "feature_dim": 80,
-            "encoder_out_dim": 512,
-            "subsampling_factor": 4,
-            "attention_dim": 512,
-            "nhead": 8,
-            "dim_feedforward": 2048,
-            "num_encoder_layers": 12,
-            "vgg_frontend": False,
-            "env_info": get_env_info(),
-        }
-    )
-    return params
-
-
-def get_encoder_model(params: AttributeDict) -> nn.Module:
-    encoder = Conformer(
-        num_features=params.feature_dim,
-        output_dim=params.encoder_out_dim,
-        subsampling_factor=params.subsampling_factor,
-        d_model=params.attention_dim,
-        nhead=params.nhead,
-        dim_feedforward=params.dim_feedforward,
-        num_encoder_layers=params.num_encoder_layers,
-        vgg_frontend=params.vgg_frontend,
-    )
-    return encoder
-
-
-def get_decoder_model(params: AttributeDict) -> nn.Module:
-    decoder = Decoder(
-        vocab_size=params.vocab_size,
-        embedding_dim=params.encoder_out_dim,
-        blank_id=params.blank_id,
-        context_size=params.context_size,
-    )
-    return decoder
-
-
-def get_joiner_model(params: AttributeDict) -> nn.Module:
-    joiner = Joiner(
-        input_dim=params.encoder_out_dim,
-        output_dim=params.vocab_size,
-    )
-    return joiner
-
-
-def get_transducer_model(params: AttributeDict) -> nn.Module:
-    encoder = get_encoder_model(params)
-    decoder = get_decoder_model(params)
-    joiner = get_joiner_model(params)
-
-    model = Transducer(
-        encoder=encoder,
-        decoder=decoder,
-        joiner=joiner,
-    )
-    return model
-
-
 def read_sound_files(
     filenames: List[str], expected_sample_rate: float
 ) -> List[torch.Tensor]:
@@ -294,11 +238,23 @@ def main():
         )
 
     num_waves = encoder_out.size(0)
-    hyps = []
+    hyp_list = []
     msg = f"Using {params.method}"
     if params.method == "beam_search":
         msg += f" with beam size {params.beam_size}"
     logging.info(msg)
+
+    if params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
     for i in range(num_waves):
         # fmt: off
         encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
@@ -313,14 +269,11 @@ def main():
             hyp = beam_search(
                 model=model, encoder_out=encoder_out_i, beam=params.beam_size
             )
-        elif params.method == "modified_beam_search":
-            hyp = modified_beam_search(
-                model=model, encoder_out=encoder_out_i, beam=params.beam_size
-            )
         else:
             raise ValueError(f"Unsupported method: {params.method}")
+        hyp_list.append(hyp)
 
-        hyps.append(sp.decode(hyp).split())
+    hyps = [sp.decode(hyp).split() for hyp in hyp_list]
 
     s = "\n"
     for filename, hyp in zip(params.sound_files, hyps):

egs/librispeech/ASR/transducer_stateless_multi_datasets/decode.py

@@ -46,15 +46,16 @@ import sentencepiece as spm
 import torch
 import torch.nn as nn
 from asr_datamodule import AsrDataModule
-from beam_search import beam_search, greedy_search, modified_beam_search
+from beam_search import (
-from conformer import Conformer
+    beam_search,
-from decoder import Decoder
+    greedy_search,
-from joiner import Joiner
+    greedy_search_batch,
+    modified_beam_search,
+)
 from librispeech import LibriSpeech
-from model import Transducer
+from train import get_params, get_transducer_model
 
 from icefall.checkpoint import average_checkpoints, load_checkpoint
-from icefall.env import get_env_info
 from icefall.utils import (
     AttributeDict,
     setup_logger,
@@ -127,7 +128,7 @@ def get_parser():
     parser.add_argument(
         "--max-sym-per-frame",
         type=int,
-        default=3,
+        default=1,
         help="""Maximum number of symbols per frame.
         Used only when --decoding_method is greedy_search""",
     )
@@ -135,71 +136,6 @@ def get_parser():
     return parser
 
 
-def get_params() -> AttributeDict:
-    params = AttributeDict(
-        {
-            # parameters for conformer
-            "feature_dim": 80,
-            "encoder_out_dim": 512,
-            "subsampling_factor": 4,
-            "attention_dim": 512,
-            "nhead": 8,
-            "dim_feedforward": 2048,
-            "num_encoder_layers": 12,
-            "vgg_frontend": False,
-            "env_info": get_env_info(),
-        }
-    )
-    return params
-
-
-def get_encoder_model(params: AttributeDict):
-    # TODO: We can add an option to switch between Conformer and Transformer
-    encoder = Conformer(
-        num_features=params.feature_dim,
-        output_dim=params.encoder_out_dim,
-        subsampling_factor=params.subsampling_factor,
-        d_model=params.attention_dim,
-        nhead=params.nhead,
-        dim_feedforward=params.dim_feedforward,
-        num_encoder_layers=params.num_encoder_layers,
-        vgg_frontend=params.vgg_frontend,
-    )
-    return encoder
-
-
-def get_decoder_model(params: AttributeDict):
-    decoder = Decoder(
-        vocab_size=params.vocab_size,
-        embedding_dim=params.encoder_out_dim,
-        blank_id=params.blank_id,
-        context_size=params.context_size,
-    )
-    return decoder
-
-
-def get_joiner_model(params: AttributeDict):
-    joiner = Joiner(
-        input_dim=params.encoder_out_dim,
-        output_dim=params.vocab_size,
-    )
-    return joiner
-
-
-def get_transducer_model(params: AttributeDict):
-    encoder = get_encoder_model(params)
-    decoder = get_decoder_model(params)
-    joiner = get_joiner_model(params)
-
-    model = Transducer(
-        encoder=encoder,
-        decoder=decoder,
-        joiner=joiner,
-    )
-
-    return model
-
-
 def decode_one_batch(
     params: AttributeDict,
     model: nn.Module,
@@ -244,9 +180,24 @@ def decode_one_batch(
     encoder_out, encoder_out_lens = model.encoder(
         x=feature, x_lens=feature_lens
     )
-    hyps = []
+    hyp_list = []
     batch_size = encoder_out.size(0)
 
+    if (
+        params.decoding_method == "greedy_search"
+        and params.max_sym_per_frame == 1
+    ):
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+    else:
         for i in range(batch_size):
             # fmt: off
             encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
@@ -259,17 +210,23 @@ def decode_one_batch(
                 )
             elif params.decoding_method == "beam_search":
                 hyp = beam_search(
-                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
                 )
             elif params.decoding_method == "modified_beam_search":
                 hyp = modified_beam_search(
-                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
                 )
             else:
                 raise ValueError(
                     f"Unsupported decoding method: {params.decoding_method}"
                 )
-        hyps.append(sp.decode(hyp).split())
+            hyp_list.append(sp.decode(hyp).split())
+
+    hyps = [sp.decode(hyp).split() for hyp in hyp_list]
 
     if params.decoding_method == "greedy_search":
         return {"greedy_search": hyps}
@@ -483,8 +440,5 @@ def main():
     logging.info("Done!")
 
 
-torch.set_num_threads(1)
-torch.set_num_interop_threads(1)
-
 if __name__ == "__main__":
     main()

egs/librispeech/ASR/transducer_stateless_multi_datasets/pretrained.py

@@ -61,14 +61,15 @@ import sentencepiece as spm
 import torch
 import torch.nn as nn
 import torchaudio
-from beam_search import beam_search, greedy_search, modified_beam_search
+from beam_search import (
-from conformer import Conformer
+    beam_search,
-from decoder import Decoder
+    greedy_search,
-from joiner import Joiner
+    greedy_search_batch,
-from model import Transducer
+    modified_beam_search,
+)
 from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model
 
-from icefall.env import get_env_info
 from icefall.utils import AttributeDict
 
 
@@ -115,6 +116,13 @@ def get_parser():
         "The sample rate has to be 16kHz.",
     )
 
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
     parser.add_argument(
         "--beam-size",
         type=int,
@@ -132,7 +140,7 @@ def get_parser():
     parser.add_argument(
         "--max-sym-per-frame",
         type=int,
-        default=3,
+        default=1,
         help="""Maximum number of symbols per frame. Used only when
         --method is greedy_search.
         """,
@@ -141,70 +149,6 @@ def get_parser():
     return parser
 
 
-def get_params() -> AttributeDict:
-    params = AttributeDict(
-        {
-            "sample_rate": 16000,
-            # parameters for conformer
-            "feature_dim": 80,
-            "encoder_out_dim": 512,
-            "subsampling_factor": 4,
-            "attention_dim": 512,
-            "nhead": 8,
-            "dim_feedforward": 2048,
-            "num_encoder_layers": 12,
-            "vgg_frontend": False,
-            "env_info": get_env_info(),
-        }
-    )
-    return params
-
-
-def get_encoder_model(params: AttributeDict) -> nn.Module:
-    encoder = Conformer(
-        num_features=params.feature_dim,
-        output_dim=params.encoder_out_dim,
-        subsampling_factor=params.subsampling_factor,
-        d_model=params.attention_dim,
-        nhead=params.nhead,
-        dim_feedforward=params.dim_feedforward,
-        num_encoder_layers=params.num_encoder_layers,
-        vgg_frontend=params.vgg_frontend,
-    )
-    return encoder
-
-
-def get_decoder_model(params: AttributeDict) -> nn.Module:
-    decoder = Decoder(
-        vocab_size=params.vocab_size,
-        embedding_dim=params.encoder_out_dim,
-        blank_id=params.blank_id,
-        context_size=params.context_size,
-    )
-    return decoder
-
-
-def get_joiner_model(params: AttributeDict) -> nn.Module:
-    joiner = Joiner(
-        input_dim=params.encoder_out_dim,
-        output_dim=params.vocab_size,
-    )
-    return joiner
-
-
-def get_transducer_model(params: AttributeDict) -> nn.Module:
-    encoder = get_encoder_model(params)
-    decoder = get_decoder_model(params)
-    joiner = get_joiner_model(params)
-
-    model = Transducer(
-        encoder=encoder,
-        decoder=decoder,
-        joiner=joiner,
-    )
-    return model
-
-
 def read_sound_files(
     filenames: List[str], expected_sample_rate: float
 ) -> List[torch.Tensor]:
@@ -294,11 +238,25 @@ def main():
         )
 
     num_waves = encoder_out.size(0)
-    hyps = []
+    hyp_list = []
     msg = f"Using {params.method}"
     if params.method == "beam_search":
         msg += f" with beam size {params.beam_size}"
     logging.info(msg)
+
+    if params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            beam=params.beam_size,
+        )
+
+    else:
         for i in range(num_waves):
             # fmt: off
             encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
@@ -311,16 +269,15 @@ def main():
                 )
             elif params.method == "beam_search":
                 hyp = beam_search(
-                model=model, encoder_out=encoder_out_i, beam=params.beam_size
+                    model=model,
-            )
+                    encoder_out=encoder_out_i,
-        elif params.method == "modified_beam_search":
+                    beam=params.beam_size,
-            hyp = modified_beam_search(
-                model=model, encoder_out=encoder_out_i, beam=params.beam_size
                 )
             else:
                 raise ValueError(f"Unsupported method: {params.method}")
+            hyp_list.append(hyp)
 
-        hyps.append(sp.decode(hyp).split())
+    hyps = [sp.decode(hyp).split() for hyp in hyp_list]
 
     s = "\n"
     for filename, hyp in zip(params.sound_files, hyps):