Add tail padding to Emformer model. (#384)

egs/librispeech/ASR/transducer_emformer/decode.py

@@ -58,6 +58,7 @@ Usage:
 
 import argparse
 import logging
+import math
 from collections import defaultdict
 from pathlib import Path
 from typing import Dict, List, Optional, Tuple
@@ -69,7 +70,7 @@ import torch.nn as nn
 from asr_datamodule import LibriSpeechAsrDataModule
 from beam_search import (
     beam_search,
-    fast_beam_search,
+    fast_beam_search_one_best,
     greedy_search,
     greedy_search_batch,
     modified_beam_search,
@@ -88,6 +89,8 @@ from icefall.utils import (
     write_error_stats,
 )
 
+LOG_EPS = math.log(1e-10)
+
 
 def get_parser():
     parser = argparse.ArgumentParser(
@@ -247,13 +250,20 @@ def decode_one_batch(
     supervisions = batch["supervisions"]
     feature_lens = supervisions["num_frames"].to(device)
 
+    feature_lens += params.left_context_length
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.left_context_length),
+        value=LOG_EPS,
+    )
+
     encoder_out, encoder_out_lens = model.encoder(
         x=feature, x_lens=feature_lens
     )
     hyps = []
 
     if params.decoding_method == "fast_beam_search":
-        hyp_tokens = fast_beam_search(
+        hyp_tokens = fast_beam_search_one_best(
             model=model,
             decoding_graph=decoding_graph,
             encoder_out=encoder_out,
@@ -271,6 +281,7 @@ def decode_one_batch(
         hyp_tokens = greedy_search_batch(
             model=model,
             encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
@@ -278,6 +289,7 @@ def decode_one_batch(
         hyp_tokens = modified_beam_search(
             model=model,
             encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
             beam=params.beam_size,
         )
         for hyp in sp.decode(hyp_tokens):
@@ -357,9 +369,9 @@ def decode_dataset(
         num_batches = "?"
 
     if params.decoding_method == "greedy_search":
-        log_interval = 100
+        log_interval = 50
     else:
-        log_interval = 2
+        log_interval = 20
 
     results = defaultdict(list)
     for batch_idx, batch in enumerate(dl):
@@ -478,8 +490,9 @@ def main():
     sp = spm.SentencePieceProcessor()
     sp.load(params.bpe_model)
 
-    # <blk> is defined in local/train_bpe_model.py
+    # <blk> and <unk> are defined in local/train_bpe_model.py
     params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
     params.vocab_size = sp.get_piece_size()
 
     logging.info(params)

egs/librispeech/ASR/transducer_emformer/export.py

@@ -24,7 +24,11 @@ Usage:
   --exp-dir ./transducer_emformer/exp \
   --bpe-model data/lang_bpe_500/bpe.model \
   --epoch 20 \
-  --avg 10
+  --avg 10 \
+  --num-encoder-layers 18 \
+  --left-context-length 128 \
+  --segment-length 8 \
+  --right-context-length 4
 
 It will generate a file exp_dir/pretrained.pt
 
@@ -39,7 +43,7 @@ you can do:
         --exp-dir ./transducer_emformer/exp \
         --epoch 9999 \
         --avg 1 \
-        --max-duration 1000 \
+        --max-duration 100 \
         --bpe-model data/lang_bpe_500/bpe.model
 """
 
@@ -51,7 +55,11 @@ import sentencepiece as spm
 import torch
 from train import add_model_arguments, get_params, get_transducer_model
 
-from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.checkpoint import (
+    average_checkpoints,
+    find_checkpoints,
+    load_checkpoint,
+)
 from icefall.utils import str2bool
 
 
@@ -64,8 +72,19 @@ def get_parser():
         "--epoch",
         type=int,
         default=28,
-        help="It specifies the checkpoint to use for decoding."
+        help="""It specifies the checkpoint to use for averaging.
-        "Note: Epoch counts from 0.",
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
     )
 
     parser.add_argument(
@@ -74,13 +93,13 @@ def get_parser():
         default=15,
         help="Number of checkpoints to average. Automatically select "
         "consecutive checkpoints before the checkpoint specified by "
-        "'--epoch'. ",
+        "'--epoch' and '--iter'",
     )
 
     parser.add_argument(
         "--exp-dir",
         type=str,
-        default="pruned_transducer_stateless/exp",
+        default="transducer_emformer/exp",
         help="""It specifies the directory where all training related
         files, e.g., checkpoints, log, etc, are saved
         """,
@@ -118,8 +137,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -143,7 +160,24 @@ def main():
 
     model.to(device)
 
-    if params.avg == 1:
+    if params.iter > 0:
+        filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+            : params.avg
+        ]
+        if len(filenames) == 0:
+            raise ValueError(
+                f"No checkpoints found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        elif len(filenames) < params.avg:
+            raise ValueError(
+                f"Not enough checkpoints ({len(filenames)}) found for"
+                f" --iter {params.iter}, --avg {params.avg}"
+            )
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+    elif params.avg == 1:
         load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
     else:
         start = params.epoch - params.avg + 1

egs/librispeech/ASR/transducer_emformer/pretrained.py

@@ -0,0 +1,352 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# 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.
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method greedy_search \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method modified_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless2/pretrained.py \
+    --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+    --bpe-model ./data/lang_bpe_500/bpe.model \
+    --method fast_beam_search \
+    --beam-size 4 \
+    /path/to/foo.wav \
+    /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by
+./pruned_transducer_stateless2/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import k2
+import kaldifeat
+import sentencepiece as spm
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import get_params, get_transducer_model, add_model_arguments
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "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,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+
+    params.update(vars(args))
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features, batch_first=True, padding_value=math.log(1e-10)
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features, x_lens=feature_lengths
+    )
+
+    num_waves = encoder_out.size(0)
+    hyps = []
+    msg = f"Using {params.method}"
+    if params.method == "beam_search":
+        msg += f" with beam size {params.beam_size}"
+    logging.info(msg)
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(f"Unsupported method: {params.method}")
+
+            hyps.append(sp.decode(hyp).split())
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/transducer_emformer/train.py

@@ -19,15 +19,23 @@
 """
 Usage:
 
-export CUDA_VISIBLE_DEVICES="0,1,2,3"
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
 
 ./transducer_emformer/train.py \
-  --world-size 4 \
+  --world-size 8 \
-  --num-epochs 30 \
+  --num-epochs 65 \
-  --start-epoch 0 \
+  --start-epoch 60 \
-  --exp-dir transducer_emformer/exp \
+  --exp-dir transducer_emformer/exp-full \
   --full-libri 1 \
-  --max-duration 300
+  --max-duration 200 \
+  --prune-range 5 \
+  --lr-factor 5 \
+  --lm-scale 0.25 \
+  --master-port 12358 \
+  --num-encoder-layers 18 \
+  --left-context-length 128 \
+  --segment-length 8 \
+  --right-context-length 4
 """