from local

egs/tedlium2/ASR/README.md

@@ -0,0 +1,18 @@
+
+# Introduction
+
+This recipe includes some different ASR models trained with TedLium3.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                  | Encoder   | Decoder            | Comment                     |
+|----------------------------------|-----------|--------------------|-----------------------------|
+| `transducer_stateless`           | Conformer | Embedding + Conv1d |                             |
+| `pruned_transducer_stateless`    | Conformer | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.

egs/tedlium2/ASR/RESULTS.md

@@ -0,0 +1,152 @@
+## Results
+
+### TedLium3 BPE training results (Pruned Transducer)
+
+#### 2022-03-21
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/261.
+
+The WERs are
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 7.27       | 6.69       | --epoch 29, --avg 13, --max-duration 100 |
+|      beam search (beam size 4)     | 6.70       | 6.04       | --epoch 29, --avg 13, --max-duration 100 |
+| modified beam search (beam size 4) | 6.77       | 6.14       | --epoch 29, --avg 13, --max-duration 100 |
+| fast beam search (set as default)  | 7.14       | 6.50       | --epoch 29, --avg 13, --max-duration 1500|
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --max-duration 300
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/VpA8b7SZQ7CEjZs9WZ5HNA/#scalars
+
+The decoding command is:
+```
+epoch=29
+avg=13
+
+## greedy search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100
+
+## beam search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method beam_search \
+  --beam-size 4
+
+## modified beam search
+./pruned_transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+## fast beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_tedlium3_pruned_transducer_stateless>
+
+### TedLium3 BPE training results (Transducer)
+
+#### Conformer encoder + embedding decoder
+
+##### 2022-03-21
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/233
+And the SpecAugment codes from this PR https://github.com/lhotse-speech/lhotse/pull/604
+
+Conformer encoder + non-current decoder. The decoder
+contains only an embedding layer and a Conv1d (with kernel size 2).
+
+The WERs are
+
+|                                    |     dev    |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 7.19       | 6.70       | --epoch 29, --avg 11, --max-duration 100 |
+|      beam search (beam size 4)     | 7.02       | 6.36       | --epoch 29, --avg 11, --max-duration 100 |
+| modified beam search (beam size 4) | 6.91       | 6.33       | --epoch 29, --avg 11, --max-duration 100 |
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --max-duration 300
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/4ks15jYHR4uMyvpW7Nz76Q/#scalars
+
+The decoding command is:
+```
+epoch=29
+avg=11
+
+## greedy search
+./transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100
+
+## beam search
+./transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method beam_search \
+  --beam-size 4
+
+## modified beam search
+./transducer_stateless/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir transducer_stateless/exp \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_tedlium3_transducer_stateless>

egs/tedlium2/ASR/local/compile_hlg.py

@@ -0,0 +1,166 @@
+#!/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.
+
+
+"""
+This script takes as input lang_dir and generates HLG from
+
+    - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt
+    - L, the lexicon, built from lang_dir/L_disambig.pt
+
+        Caution: We use a lexicon that contains disambiguation symbols
+
+    - G, the LM, built from data/lm/G_3_gram.fst.txt
+
+The generated HLG is saved in $lang_dir/HLG.pt
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import torch
+
+from icefall.lexicon import Lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lm",
+        type=str,
+        default="G_3_gram",
+        help="""Stem name for LM used in HLG compiling.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def compile_HLG(lang_dir: str, lm: str = "G_3_gram") -> k2.Fsa:
+    """
+    Args:
+      lang_dir:
+        The language directory, e.g., data/lang_phone or data/lang_bpe_5000.
+      lm:
+        The language stem base name.
+
+    Return:
+      An FSA representing HLG.
+    """
+    lexicon = Lexicon(lang_dir)
+    max_token_id = max(lexicon.tokens)
+    logging.info(f"Building ctc_topo. max_token_id: {max_token_id}")
+    H = k2.ctc_topo(max_token_id)
+    L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt"))
+
+    if Path(f"data/lm/{lm}.pt").is_file():
+        logging.info(f"Loading pre-compiled {lm}")
+        d = torch.load(f"data/lm/{lm}.pt")
+        G = k2.Fsa.from_dict(d)
+    else:
+        logging.info(f"Loading {lm}.fst.txt")
+        with open(f"data/lm/{lm}.fst.txt") as f:
+            G = k2.Fsa.from_openfst(f.read(), acceptor=False)
+            torch.save(G.as_dict(), f"data/lm/{lm}.pt")
+
+    first_token_disambig_id = lexicon.token_table["#0"]
+    first_word_disambig_id = lexicon.word_table["#0"]
+
+    L = k2.arc_sort(L)
+    G = k2.arc_sort(G)
+
+    logging.info("Intersecting L and G")
+    LG = k2.compose(L, G)
+    logging.info(f"LG shape: {LG.shape}")
+
+    logging.info("Connecting LG")
+    LG = k2.connect(LG)
+    logging.info(f"LG shape after k2.connect: {LG.shape}")
+
+    logging.info(type(LG.aux_labels))
+    logging.info("Determinizing LG")
+
+    LG = k2.determinize(LG)
+    logging.info(type(LG.aux_labels))
+
+    logging.info("Connecting LG after k2.determinize")
+    LG = k2.connect(LG)
+
+    logging.info("Removing disambiguation symbols on LG")
+
+    LG.labels[LG.labels >= first_token_disambig_id] = 0
+    # See https://github.com/k2-fsa/k2/issues/874
+    # for why we need to set LG.properties to None
+    LG.__dict__["_properties"] = None
+
+    assert isinstance(LG.aux_labels, k2.RaggedTensor)
+    LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0
+
+    LG = k2.remove_epsilon(LG)
+    logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}")
+
+    LG = k2.connect(LG)
+    LG.aux_labels = LG.aux_labels.remove_values_eq(0)
+
+    logging.info("Arc sorting LG")
+    LG = k2.arc_sort(LG)
+
+    logging.info("Composing H and LG")
+    # CAUTION: The name of the inner_labels is fixed
+    # to `tokens`. If you want to change it, please
+    # also change other places in icefall that are using
+    # it.
+    HLG = k2.compose(H, LG, inner_labels="tokens")
+
+    logging.info("Connecting LG")
+    HLG = k2.connect(HLG)
+
+    logging.info("Arc sorting LG")
+    HLG = k2.arc_sort(HLG)
+    logging.info(f"HLG.shape: {HLG.shape}")
+
+    return HLG
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    if (lang_dir / "HLG.pt").is_file():
+        logging.info(f"{lang_dir}/HLG.pt already exists - skipping")
+        return
+
+    logging.info(f"Processing {lang_dir}")
+
+    HLG = compile_HLG(lang_dir, args.lm)
+    logging.info(f"Saving HLG.pt to {lang_dir}")
+    torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/tedlium2/ASR/local/compute_fbank_musan.py

@@ -0,0 +1,105 @@
+#!/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.
+
+
+"""
+This file computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter, combine
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_musan():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    prefix = "musan"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    musan_cuts_path = output_dir / "musan_cuts.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        # create chunks of Musan with duration 5 - 10 seconds
+        musan_cuts = (
+            CutSet.from_manifests(
+                recordings=combine(part["recordings"] for part in manifests.values())
+            )
+            .cut_into_windows(10.0)
+            .filter(lambda c: c.duration > 5)
+            .compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/musan_feats",
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        musan_cuts.to_file(musan_cuts_path)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan()

egs/tedlium2/ASR/local/compute_fbank_tedlium.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 	       2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# 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.
+"""
+This file computes fbank features of the TedLium3 dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_tedlium():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    dataset_parts = (
+        "train",
+        "dev",
+        "test",
+    )
+
+    prefix = "tedlium"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        prefix=prefix,
+        suffix=suffix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+                )
+            cur_num_jobs = num_jobs if ex is None else 80
+            cur_num_jobs = min(cur_num_jobs, len(cut_set))
+
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/{prefix}_feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=cur_num_jobs,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+            # Split long cuts into many short and un-overlapping cuts
+            cut_set = cut_set.trim_to_supervisions(keep_overlapping=False)
+            cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_tedlium()

egs/tedlium2/ASR/local/convert_transcript_words_to_bpe_ids.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+
+# Copyright    2022 Xiaomi Corporation  (Author: Mingshuang Luo)
+"""
+Convert a transcript based on words to a list of BPE ids.
+
+For example, if we use 2 as the encoding id of <unk>:
+
+texts = ['this is a <unk> day']
+spm_ids = [[38, 33, 6, 2, 316]]
+
+texts = ['<unk> this is a sunny day']
+spm_ids = [[2, 38, 33, 6, 118, 11, 11, 21, 316]]
+
+texts = ['<unk>']
+spm_ids = [[2]]
+"""
+
+import argparse
+import logging
+from typing import List
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--texts", type=List[str], help="The input transcripts list.")
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    return parser.parse_args()
+
+
+def convert_texts_into_ids(
+    texts: List[str],
+    unk_id: int,
+    sp: spm.SentencePieceProcessor,
+) -> List[List[int]]:
+    """
+    Args:
+      texts:
+        A string list of transcripts, such as ['Today is Monday', 'It's sunny'].
+      unk_id:
+        A number id for the token '<unk>'.
+    Returns:
+      Return an integer list of bpe ids.
+    """
+    y = []
+    for text in texts:
+        y_ids = []
+        if "<unk>" in text:
+            text_segments = text.split("<unk>")
+            id_segments = sp.encode(text_segments, out_type=int)
+            for i in range(len(id_segments)):
+                if i != len(id_segments) - 1:
+                    y_ids.extend(id_segments[i] + [unk_id])
+                else:
+                    y_ids.extend(id_segments[i])
+        else:
+            y_ids = sp.encode(text, out_type=int)
+        y.append(y_ids)
+
+    return y
+
+
+def main():
+    args = get_args()
+    texts = args.texts
+    bpe_model = args.bpe_model
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(bpe_model)
+    unk_id = sp.piece_to_id("<unk>")
+
+    y = convert_texts_into_ids(
+        texts=texts,
+        unk_id=unk_id,
+        sp=sp,
+    )
+    logging.info(f"The input texts: {texts}")
+    logging.info(f"The encoding ids: {y}")
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/local/convert_transcript_words_to_tokens.py

@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+"""
+Convert a transcript file containing words to a corpus file containing tokens
+for LM training with the help of a lexicon.
+
+If the lexicon contains phones, the resulting LM will be a phone LM; If the
+lexicon contains word pieces, the resulting LM will be a word piece LM.
+
+If a word has multiple pronunciations, the one that appears first in the lexicon
+is kept; others are removed.
+
+If the input transcript is:
+
+    hello zoo world hello
+    world zoo
+    foo zoo world hellO
+
+and if the lexicon is
+
+    <UNK> SPN
+    hello h e l l o 2
+    hello h e l l o
+    world w o r l d
+    zoo z o o
+
+Then the output is
+
+    h e l l o 2 z o o w o r l d h e l l o 2
+    w o r l d z o o
+    SPN z o o w o r l d SPN
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List
+
+from generate_unique_lexicon import filter_multiple_pronunications
+
+from icefall.lexicon import read_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="The input transcript file."
+        "We assume that the transcript file consists of "
+        "lines. Each line consists of space separated words.",
+    )
+    parser.add_argument("--lexicon", type=str, help="The input lexicon file.")
+    parser.add_argument("--oov", type=str, default="<UNK>", help="The OOV word.")
+
+    return parser.parse_args()
+
+
+def process_line(lexicon: Dict[str, List[str]], line: str, oov_token: str) -> None:
+    """
+    Args:
+      lexicon:
+        A dict containing pronunciations. Its keys are words and values
+        are pronunciations (i.e., tokens).
+      line:
+        A line of transcript consisting of space(s) separated words.
+      oov_token:
+        The pronunciation of the oov word if a word in `line` is not present
+        in the lexicon.
+    Returns:
+      Return None.
+    """
+    s = ""
+    words = line.strip().split()
+    for i, w in enumerate(words):
+        tokens = lexicon.get(w, oov_token)
+        s += " ".join(tokens)
+        s += " "
+    print(s.strip())
+
+
+def main():
+    args = get_args()
+    assert Path(args.lexicon).is_file()
+    assert Path(args.transcript).is_file()
+    assert len(args.oov) > 0
+
+    # Only the first pronunciation of a word is kept
+    lexicon = filter_multiple_pronunications(read_lexicon(args.lexicon))
+
+    lexicon = dict(lexicon)
+
+    assert args.oov in lexicon
+
+    oov_token = lexicon[args.oov]
+
+    with open(args.transcript) as f:
+        for line in f:
+            process_line(lexicon=lexicon, line=line, oov_token=oov_token)
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/local/display_manifest_statistics.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						   Mingshuang Luo)
+#
+# 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.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest_lazy
+
+
+def main():
+    path = "./data/fbank/tedlium_cuts_train.jsonl.gz"
+    path = "./data/fbank/tedlium_cuts_dev.jsonl.gz"
+    path = "./data/fbank/tedlium_cuts_test.jsonl.gz"
+
+    cuts = load_manifest_lazy(path)
+    cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+## train
+Cuts count: 804789
+Total duration (hours): 1370.6
+Speech duration (hours): 1370.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    6.1
+std     3.1
+min     0.5
+25%     3.7
+50%     6.0
+75%     8.3
+99.5%   14.9
+99.9%   16.6
+max     33.3
+
+## dev
+Cuts count: 507
+Total duration (hours): 1.6
+Speech duration (hours): 1.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    11.3
+std     5.7
+min     0.5
+25%     7.5
+50%     10.6
+75%     14.4
+99.5%   29.8
+99.9%   37.7
+max     39.9
+
+## test
+Cuts count: 1155
+Total duration (hours): 2.6
+Speech duration (hours): 2.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    8.2
+std     4.3
+min     0.3
+25%     4.6
+50%     8.2
+75%     10.9
+99.5%   22.1
+99.9%   26.7
+max     32.5
+"""

egs/tedlium2/ASR/local/generate_unique_lexicon.py

@@ -0,0 +1,98 @@
+#!/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.
+
+"""
+This file takes as input a lexicon.txt and output a new lexicon,
+in which each word has a unique pronunciation.
+
+The way to do this is to keep only the first pronunciation of a word
+in lexicon.txt.
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from typing import List, Tuple
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        This file will generate a new file uniq_lexicon.txt
+        in it.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def filter_multiple_pronunications(
+    lexicon: List[Tuple[str, List[str]]]
+) -> List[Tuple[str, List[str]]]:
+    """Remove multiple pronunciations of words from a lexicon.
+
+    If a word has more than one pronunciation in the lexicon, only
+    the first one is kept, while other pronunciations are removed
+    from the lexicon.
+
+    Args:
+      lexicon:
+        The input lexicon, containing a list of (word, [p1, p2, ..., pn]),
+        where "p1, p2, ..., pn" are the pronunciations of the "word".
+    Returns:
+      Return a new lexicon where each word has a unique pronunciation.
+    """
+    seen = set()
+    ans = []
+
+    for word, tokens in lexicon:
+        if word in seen:
+            continue
+        seen.add(word)
+        ans.append((word, tokens))
+    return ans
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+
+    lexicon_filename = lang_dir / "lexicon.txt"
+
+    in_lexicon = read_lexicon(lexicon_filename)
+
+    out_lexicon = filter_multiple_pronunications(in_lexicon)
+
+    write_lexicon(lang_dir / "uniq_lexicon.txt", out_lexicon)
+
+    logging.info(f"Number of entries in lexicon.txt: {len(in_lexicon)}")
+    logging.info(f"Number of entries in uniq_lexicon.txt: {len(out_lexicon)}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/tedlium2/ASR/local/prepare_lang.py

@@ -0,0 +1,413 @@
+#!/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.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+from icefall.utils import str2bool
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain a file lexicon.txt.
+        Generated files by this script are saved into this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    lexicon_filename = lang_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(lang_dir / "tokens.txt", token2id)
+    write_mapping(lang_dir / "words.txt", word2id)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/local/prepare_lang_bpe.py

@@ -0,0 +1,259 @@
+#!/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.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/bpe.model,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import argparse
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+from icefall.utils import str2bool
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [token2id[i] for i in pieces]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def generate_lexicon(
+    model_file: str, words: List[str]
+) -> Tuple[Lexicon, Dict[str, int]]:
+    """Generate a lexicon from a BPE model.
+
+    Args:
+      model_file:
+        Path to a sentencepiece model.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a tuple with two elements:
+        - A dict whose keys are words and values are the corresponding
+          word pieces.
+        - A dict representing the token symbol, mapping from tokens to IDs.
+    """
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(model_file))
+
+    # Convert word to word piece IDs instead of word piece strings
+    # to avoid OOV tokens.
+    words_pieces_ids: List[List[int]] = sp.encode(words, out_type=int)
+
+    # Now convert word piece IDs back to word piece strings.
+    words_pieces: List[List[str]] = [sp.id_to_piece(ids) for ids in words_pieces_ids]
+
+    lexicon = []
+    for word, pieces in zip(words, words_pieces):
+        lexicon.append((word, pieces))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", [sp.id_to_piece(sp.unk_id())]))
+
+    token2id: Dict[str, int] = dict()
+    for i in range(sp.vocab_size()):
+        token2id[sp.id_to_piece(i)] = i
+
+    return lexicon, token2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        It should contain the bpe.model and words.txt
+        """,
+    )
+
+    parser.add_argument(
+        "--debug",
+        type=str2bool,
+        default=False,
+        help="""True for debugging, which will generate
+        a visualization of the lexicon FST.
+
+        Caution: If your lexicon contains hundreds of thousands
+        of lines, please set it to False!
+
+        See "test/test_bpe_lexicon.py" for usage.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    lang_dir = Path(args.lang_dir)
+    model_file = lang_dir / "bpe.model"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    lexicon, token_sym_table = generate_lexicon(model_file, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+    if args.debug:
+        labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt")
+        aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+        L.labels_sym = labels_sym
+        L.aux_labels_sym = aux_labels_sym
+        L.draw(f"{lang_dir / 'L.svg'}", title="L.pt")
+
+        L_disambig.labels_sym = labels_sym
+        L_disambig.aux_labels_sym = aux_labels_sym
+        L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/local/prepare_lexicon.py

@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+# Copyright    2022  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# 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.
+
+
+"""
+This script takes as input supervisions json dir "data/manifests"
+consisting of supervisions_train.json and does the following:
+
+1. Generate lexicon_words.txt.
+
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import lhotse
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--manifests-dir",
+        type=str,
+        help="""Input directory.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def prepare_lexicon(manifests_dir: str, lang_dir: str):
+    """
+    Args:
+      manifests_dir:
+        The manifests directory, e.g., data/manifests.
+      lang_dir:
+        The language directory, e.g., data/lang_phone.
+
+    Return:
+      The lexicon_words.txt file.
+    """
+    words = set()
+
+    lexicon = Path(lang_dir) / "lexicon_words.txt"
+    sups = lhotse.load_manifest(f"{manifests_dir}/tedlium_supervisions_train.jsonl.gz")
+    for s in sups:
+        # list the words units and filter the empty item
+        words_list = list(filter(None, s.text.split()))
+
+        for word in words_list:
+            if word not in words and word != "<unk>":
+                words.add(word)
+
+    with open(lexicon, "w") as f:
+        for word in sorted(words):
+            f.write(word + "  " + word)
+            f.write("\n")
+
+
+def main():
+    args = get_args()
+    manifests_dir = Path(args.manifests_dir)
+    lang_dir = Path(args.lang_dir)
+
+    logging.info("Generating lexicon_words.txt")
+    prepare_lexicon(manifests_dir, lang_dir)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/tedlium2/ASR/local/prepare_transcripts.py

@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang Luo)
+#
+# 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.
+
+
+"""
+This script takes as input supervisions json dir "data/manifests"
+consisting of supervisions_train.json and does the following:
+
+1. Generate train.text.
+
+"""
+import argparse
+import logging
+from pathlib import Path
+
+import lhotse
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--manifests-dir",
+        type=str,
+        help="""Input directory.
+        """,
+    )
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Output directory.
+        """,
+    )
+
+    return parser.parse_args()
+
+
+def prepare_transcripts(manifests_dir: str, lang_dir: str):
+    """
+    Args:
+      manifests_dir:
+        The manifests directory, e.g., data/manifests.
+      lang_dir:
+        The language directory, e.g., data/lang_phone.
+
+    Return:
+      The train.text in lang_dir.
+    """
+    texts = []
+
+    train_text = Path(lang_dir) / "train.text"
+    sups = lhotse.load_manifest(f"{manifests_dir}/tedlium_supervisions_train.jsonl.gz")
+    for s in sups:
+        texts.append(s.text)
+
+    with open(train_text, "w") as f:
+        for text in texts:
+            f.write(text)
+            f.write("\n")
+
+
+def main():
+    args = get_args()
+    manifests_dir = Path(args.manifests_dir)
+    lang_dir = Path(args.lang_dir)
+
+    logging.info("Generating train.text")
+    prepare_transcripts(manifests_dir, lang_dir)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()

egs/tedlium2/ASR/local/test_prepare_lang.py

@@ -0,0 +1,104 @@
+#!/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.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(lexicon_disambig, phone2id=phone2id, word2id=word2id)
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/local/train_bpe_model.py

@@ -0,0 +1,100 @@
+#!/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.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "unigram"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    character_coverage = 1.0
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/prepare.sh

@@ -0,0 +1,172 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+stage=0
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/tedlium3
+#      You can find data, doc, legacy, LM, etc, inside it.
+#      You can download them from https://www.openslr.org/51
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  5000
+  2000
+  1000
+  500
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/tedlium3,
+  # you can create a symlink
+  #
+  # ln -sfv /path/to/tedlium3 $dl_dir/tedlium3
+  #
+  if [ ! -d $dl_dir/tedlium3 ]; then
+    lhotse download tedlium $dl_dir
+    mv $dl_dir/TEDLIUM_release-3 $dl_dir/tedlium3
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #ln -sfv /path/to/musan $dl_dir/musan
+
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare tedlium3 manifests"
+  if [ ! -f data/manifests/.tedlium3.done ]; then
+    # We assume that you have downloaded the tedlium3 corpus
+    # to $dl_dir/tedlium3
+    mkdir -p data/manifests
+    lhotse prepare tedlium $dl_dir/tedlium3 data/manifests
+    touch data/manifests/.tedlium3.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifests"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -e data/manifests/.musan.done ]; then
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute fbank for tedlium3"
+
+  if [ ! -e data/fbank/.tedlium3.done ]; then
+    mkdir -p data/fbank
+    python3 ./local/compute_fbank_tedlium.py
+    touch data/fbank/.tedlium3.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -e data/fbank/.musan.done ]; then
+    mkdir -p data/fbank
+    python3 ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Prepare phone based lang"
+  lang_dir=data/lang_phone
+  mkdir -p $lang_dir
+
+  if [ ! -f $lang_dir/train.text ]; then
+    ./local/prepare_transcripts.py \
+      --lang-dir $lang_dir \
+      --manifests-dir data/manifests
+  fi
+
+  if [ ! -f $lang_dir/lexicon_words.txt ]; then
+    ./local/prepare_lexicon.py \
+      --lang-dir $lang_dir \
+      --manifests-dir data/manifests
+  fi
+
+  (echo '!SIL SIL'; echo '<UNK> <UNK>'; ) |
+    cat - $lang_dir/lexicon_words.txt |
+    sort | uniq > $lang_dir/lexicon.txt
+
+  if [ ! -f $lang_dir/L_disambig.pt ]; then
+    ./local/prepare_lang.py --lang-dir $lang_dir
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+    # We reuse words.txt from phone based lexicon
+    # so that the two can share G.pt later.
+    cp data/lang_phone/words.txt $lang_dir
+
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      log "Generate data for BPE training"
+      cat data/lang_phone/train.text |
+      cut -d " " -f 2- > $lang_dir/transcript_words.txt
+      # remove the <unk> for transcript_words.txt
+      sed -i 's/ <unk>//g' $lang_dir/transcript_words.txt
+      sed -i 's/<unk> //g' $lang_dir/transcript_words.txt
+      sed -i 's/<unk>//g' $lang_dir/transcript_words.txt
+    fi
+
+    ./local/train_bpe_model.py \
+      --lang-dir $lang_dir \
+      --vocab-size $vocab_size \
+      --transcript $lang_dir/transcript_words.txt
+
+    if [ ! -f $lang_dir/L_disambig.pt ]; then
+      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
+    fi
+  done
+fi

egs/tedlium2/ASR/pruned_transducer_stateless/asr_datamodule.py

@@ -0,0 +1,368 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2021  Xiaomi Corporation (Author: Mingshuang Luo)
+#
+# 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.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class TedLiumAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. TEDLium3 dev
+    and test).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset.",
+        )
+
+    def train_dataloaders(
+        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=10,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                    max_frames_mask_fraction=0.15,
+                    p=0.9,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to get Musan cuts")
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        logging.info("About to create train dataloader")
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts_test: CutSet) -> DataLoader:
+
+        logging.debug("About to create test dataset")
+        if self.args.on_the_fly_feats:
+            test = K2SpeechRecognitionDataset(
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            test = K2SpeechRecognitionDataset(
+                return_cuts=self.args.return_cuts,
+            )
+
+        test_sampler = DynamicBucketingSampler(
+            cuts_test,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=test_sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "tedlium_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_test.jsonl.gz")

egs/tedlium2/ASR/pruned_transducer_stateless/decode.py

@@ -0,0 +1,525 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Mingshuang Luo)
+#
+# 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_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 13 \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+"""
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An interger indicating how many candidates we will keep for each
+        frame. Used only when --decoding-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 --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-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 --decoding_method is greedy_search""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    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_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.decoding_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())
+    elif params.decoding_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())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    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())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            decoding_graph=decoding_graph,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <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)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tedlium = TedLiumAsrDataModule(args)
+    dev_cuts = tedlium.dev_cuts()
+    test_cuts = tedlium.test_cuts()
+
+    dev_dl = tedlium.valid_dataloaders(dev_cuts)
+    test_dl = tedlium.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/pruned_transducer_stateless/decoder.py

@@ -0,0 +1,103 @@
+# 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.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        unk_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          unk_id:
+            The ID of the unk symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+        self.unk_id = unk_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        self.vocab_size = vocab_size
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+        self.output_linear = nn.Linear(embedding_dim, vocab_size)
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U) with blank prepended.
+          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).
+        """
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(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 embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = self.output_linear(F.relu(embedding_out))
+        return embedding_out

egs/tedlium2/ASR/pruned_transducer_stateless/encoder_interface.py

@@ -0,0 +1,43 @@
+# 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.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")

egs/tedlium2/ASR/pruned_transducer_stateless/export.py

@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+# 					                         Mingshuang Luo)
+#
+# 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.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless/export.py \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 29 \
+  --avg 13
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/tedlium3/ASR
+    ./pruned_transducer_stateless/decode.py \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 1 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <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)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/tedlium2/ASR/pruned_transducer_stateless/joiner.py

@@ -0,0 +1,52 @@
+# 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.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, inner_dim: int, output_dim: int):
+        super().__init__()
+
+        self.inner_linear = nn.Linear(input_dim, inner_dim)
+        self.output_linear = nn.Linear(inner_dim, output_dim)
+
+    def forward(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C) during
+            training or (N, C) in case of streaming decoding.
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C) during
+            training or (N, C) in case of streaming decoding.
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim
+        assert encoder_out.ndim in (2, 4)
+        assert encoder_out.shape == decoder_out.shape
+
+        logit = encoder_out + decoder_out
+
+        logit = self.inner_linear(torch.tanh(logit))
+
+        output = self.output_linear(F.relu(logit))
+
+        return output

egs/tedlium2/ASR/pruned_transducer_stateless/model.py

@@ -0,0 +1,175 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, 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 Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        reduction: str = "sum",
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          reduction:
+            "sum" to sum the losses over all utterances in the batch.
+            "none" to return the loss in a 1-D tensor for each utterance
+            in the batch.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert reduction in ("sum", "none"), reduction
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, C]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+            lm=decoder_out,
+            am=encoder_out,
+            symbols=y_padded,
+            termination_symbol=blank_id,
+            lm_only_scale=lm_scale,
+            am_only_scale=am_scale,
+            boundary=boundary,
+            reduction=reduction,
+            return_grad=True,
+        )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, C]
+        # lm_pruned : [B, T, prune_range, C]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=encoder_out, lm=decoder_out, ranges=ranges
+        )
+
+        # logits : [B, T, prune_range, C]
+        logits = self.joiner(am_pruned, lm_pruned)
+
+        pruned_loss = k2.rnnt_loss_pruned(
+            logits=logits,
+            symbols=y_padded,
+            ranges=ranges,
+            termination_symbol=blank_id,
+            boundary=boundary,
+            reduction=reduction,
+        )
+
+        return (simple_loss, pruned_loss)

egs/tedlium2/ASR/pruned_transducer_stateless/pretrained.py

@@ -0,0 +1,352 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		 2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# 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_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/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_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/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_stateless/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless/exp/pretrained.pt is generated by
+./pruned_transducer_stateless/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
+
+
+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.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-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(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search and 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 --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-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.
+        """,
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    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}. 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> 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(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)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    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)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    logging.info(msg)
+
+    if params.decoding_method == "fast_beam_search":
+        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.decoding_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())
+    elif params.decoding_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())
+    else:
+        batch_size = encoder_out.size(0)
+
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    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())
+
+    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/tedlium2/ASR/pruned_transducer_stateless/subsampling.py

@@ -0,0 +1,153 @@
+# 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.
+
+
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x

egs/tedlium2/ASR/pruned_transducer_stateless/test_decoder.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						   Mingshuang Luo)
+#
+# 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.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/tedlium3/ASR
+    python ./pruned_transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    unk_id = 2
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        unk_id=unk_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, vocab_size)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, vocab_size)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/pruned_transducer_stateless/train.py

@@ -0,0 +1,767 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# 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:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless/exp \
+  --max-duration 300
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12350,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    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(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network) part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # 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,
+            # parameters for decoder
+            "embedding_dim": 512,
+            # parameters for Noam
+            "warm_step": 80000,
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.vocab_size,
+        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.embedding_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.vocab_size,
+        inner_dim=params.embedding_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 load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    unk_id = params.unk_id
+    y = convert_texts_into_ids(texts, unk_id, sp=sp)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <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)
+
+    logging.info("About to create 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}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    tedlium = TedLiumAsrDataModule(args)
+
+    train_cuts = tedlium.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 17 seconds
+        return 1.0 <= c.duration <= 17.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = tedlium.train_dataloaders(train_cuts)
+    valid_cuts = tedlium.dev_cuts()
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        sp=sp,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/pruned_transducer_stateless/transformer.py

@@ -0,0 +1,416 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# 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.
+
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+from icefall.utils import make_pad_mask
+
+
+class Transformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          dropout:
+            Dropout in encoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.output_dim = output_dim
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        lengths = ((x_lens - 1) // 2 - 1) // 2
+        assert x.size(0) == lengths.max().item()
+
+        mask = make_pad_mask(lengths)
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)

egs/tedlium2/ASR/shared/make_kn_lm.py

@@ -0,0 +1,444 @@
+#!/usr/bin/env python3
+
+# Copyright 2016  Johns Hopkins University (Author: Daniel Povey)
+#           2018  Ruizhe Huang
+# Apache 2.0.
+
+# This is an implementation of computing Kneser-Ney smoothed language model
+# in the same way as srilm. This is a back-off, unmodified version of
+# Kneser-Ney smoothing, which produces the same results as the following
+# command (as an example) of srilm:
+#
+# $ ngram-count -order 4 -kn-modify-counts-at-end -ukndiscount -gt1min 0 -gt2min 0 -gt3min 0 -gt4min 0 \
+# -text corpus.txt -lm lm.arpa
+#
+# The data structure is based on: kaldi/egs/wsj/s5/utils/lang/make_phone_lm.py
+# The smoothing algorithm is based on: http://www.speech.sri.com/projects/srilm/manpages/ngram-discount.7.html
+
+import argparse
+import io
+import math
+import os
+import re
+import sys
+from collections import Counter, defaultdict
+
+parser = argparse.ArgumentParser(
+    description="""
+    Generate kneser-ney language model as arpa format. By default,
+    it will read the corpus from standard input, and output to standard output.
+    """
+)
+parser.add_argument(
+    "-ngram-order",
+    type=int,
+    default=4,
+    choices=[2, 3, 4, 5, 6, 7],
+    help="Order of n-gram",
+)
+parser.add_argument("-text", type=str, default=None, help="Path to the corpus file")
+parser.add_argument(
+    "-lm", type=str, default=None, help="Path to output arpa file for language models"
+)
+parser.add_argument(
+    "-verbose", type=int, default=0, choices=[0, 1, 2, 3, 4, 5], help="Verbose level"
+)
+args = parser.parse_args()
+
+# For encoding-agnostic scripts, we assume byte stream as input.
+# Need to be very careful about the use of strip() and split()
+# in this case, because there is a latin-1 whitespace character
+# (nbsp) which is part of the unicode encoding range.
+# Ref: kaldi/egs/wsj/s5/utils/lang/bpe/prepend_words.py @ 69cd717
+default_encoding = "latin-1"
+
+strip_chars = " \t\r\n"
+whitespace = re.compile("[ \t]+")
+
+
+class CountsForHistory:
+    # This class (which is more like a struct) stores the counts seen in a
+    # particular history-state.  It is used inside class NgramCounts.
+    # It really does the job of a dict from int to float, but it also
+    # keeps track of the total count.
+    def __init__(self):
+        # The 'lambda: defaultdict(float)' is an anonymous function taking no
+        # arguments that returns a new defaultdict(float).
+        self.word_to_count = defaultdict(int)
+        # using a set to count the number of unique contexts
+        self.word_to_context = defaultdict(set)
+        self.word_to_f = dict()  # discounted probability
+        self.word_to_bow = dict()  # back-off weight
+        self.total_count = 0
+
+    def words(self):
+        return self.word_to_count.keys()
+
+    def __str__(self):
+        # e.g. returns ' total=12: 3->4, 4->6, -1->2'
+        return " total={0}: {1}".format(
+            str(self.total_count),
+            ", ".join(
+                [
+                    "{0} -> {1}".format(word, count)
+                    for word, count in self.word_to_count.items()
+                ]
+            ),
+        )
+
+    def add_count(self, predicted_word, context_word, count):
+        assert count >= 0
+
+        self.total_count += count
+        self.word_to_count[predicted_word] += count
+        if context_word is not None:
+            self.word_to_context[predicted_word].add(context_word)
+
+
+class NgramCounts:
+    # A note on data-structure.  Firstly, all words are represented as
+    # integers.  We store n-gram counts as an array, indexed by (history-length
+    # == n-gram order minus one) (note: python calls arrays "lists") of dicts
+    # from histories to counts, where histories are arrays of integers and
+    # "counts" are dicts from integer to float.  For instance, when
+    # accumulating the 4-gram count for the '8' in the sequence '5 6 7 8', we'd
+    # do as follows: self.counts[3][[5,6,7]][8] += 1.0 where the [3] indexes an
+    # array, the [[5,6,7]] indexes a dict, and the [8] indexes a dict.
+    def __init__(self, ngram_order, bos_symbol="<s>", eos_symbol="</s>"):
+        assert ngram_order >= 2
+
+        self.ngram_order = ngram_order
+        self.bos_symbol = bos_symbol
+        self.eos_symbol = eos_symbol
+
+        self.counts = []
+        for n in range(ngram_order):
+            self.counts.append(defaultdict(lambda: CountsForHistory()))
+
+        self.d = []  # list of discounting factor for each order of ngram
+
+    # adds a raw count (called while processing input data).
+    # Suppose we see the sequence '6 7 8 9' and ngram_order=4, 'history'
+    # would be (6,7,8) and 'predicted_word' would be 9; 'count' would be
+    # 1.
+    def add_count(self, history, predicted_word, context_word, count):
+        self.counts[len(history)][history].add_count(
+            predicted_word, context_word, count
+        )
+
+    # 'line' is a string containing a sequence of integer word-ids.
+    # This function adds the un-smoothed counts from this line of text.
+    def add_raw_counts_from_line(self, line):
+        if line == "":
+            words = [self.bos_symbol, self.eos_symbol]
+        else:
+            words = [self.bos_symbol] + whitespace.split(line) + [self.eos_symbol]
+
+        for i in range(len(words)):
+            for n in range(1, self.ngram_order + 1):
+                if i + n > len(words):
+                    break
+                ngram = words[i : i + n]
+                predicted_word = ngram[-1]
+                history = tuple(ngram[:-1])
+                if i == 0 or n == self.ngram_order:
+                    context_word = None
+                else:
+                    context_word = words[i - 1]
+
+                self.add_count(history, predicted_word, context_word, 1)
+
+    def add_raw_counts_from_standard_input(self):
+        lines_processed = 0
+        # byte stream as input
+        infile = io.TextIOWrapper(sys.stdin.buffer, encoding=default_encoding)
+        for line in infile:
+            line = line.strip(strip_chars)
+            self.add_raw_counts_from_line(line)
+            lines_processed += 1
+        if lines_processed == 0 or args.verbose > 0:
+            print(
+                "make_phone_lm.py: processed {0} lines of input".format(
+                    lines_processed
+                ),
+                file=sys.stderr,
+            )
+
+    def add_raw_counts_from_file(self, filename):
+        lines_processed = 0
+        with open(filename, encoding=default_encoding) as fp:
+            for line in fp:
+                line = line.strip(strip_chars)
+                self.add_raw_counts_from_line(line)
+                lines_processed += 1
+        if lines_processed == 0 or args.verbose > 0:
+            print(
+                "make_phone_lm.py: processed {0} lines of input".format(
+                    lines_processed
+                ),
+                file=sys.stderr,
+            )
+
+    def cal_discounting_constants(self):
+        # For each order N of N-grams, we calculate discounting constant D_N = n1_N / (n1_N + 2 * n2_N),
+        # where n1_N is the number of unique N-grams with count = 1 (counts-of-counts).
+        # This constant is used similarly to absolute discounting.
+        # Return value: d is a list of floats, where d[N+1] = D_N
+
+        # for the lowest order, i.e., 1-gram, we do not need to discount, thus the constant is 0
+        # This is a special case: as we currently assumed having seen all vocabularies in the dictionary,
+        # but perhaps this is not the case for some other scenarios.
+        self.d = [0]
+        for n in range(1, self.ngram_order):
+            this_order_counts = self.counts[n]
+            n1 = 0
+            n2 = 0
+            for hist, counts_for_hist in this_order_counts.items():
+                stat = Counter(counts_for_hist.word_to_count.values())
+                n1 += stat[1]
+                n2 += stat[2]
+            assert n1 + 2 * n2 > 0
+
+            # We are doing this max(0.001, xxx) to avoid zero discounting constant D due to n1=0,
+            # which could happen if the number of symbols is small.
+            # Otherwise, zero discounting constant can cause division by zero in computing BOW.
+            self.d.append(max(0.1, n1 * 1.0) / (n1 + 2 * n2))
+
+    def cal_f(self):
+        # f(a_z) is a probability distribution of word sequence a_z.
+        # Typically f(a_z) is discounted to be less than the ML estimate so we have
+        # some leftover probability for the z words unseen in the context (a_).
+        #
+        # f(a_z) = (c(a_z) - D0) / c(a_)    ;; for highest order N-grams
+        # f(_z)  = (n(*_z) - D1) / n(*_*)	;; for lower order N-grams
+
+        # highest order N-grams
+        n = self.ngram_order - 1
+        this_order_counts = self.counts[n]
+        for hist, counts_for_hist in this_order_counts.items():
+            for w, c in counts_for_hist.word_to_count.items():
+                counts_for_hist.word_to_f[w] = (
+                    max((c - self.d[n]), 0) * 1.0 / counts_for_hist.total_count
+                )
+
+        # lower order N-grams
+        for n in range(0, self.ngram_order - 1):
+            this_order_counts = self.counts[n]
+            for hist, counts_for_hist in this_order_counts.items():
+
+                n_star_star = 0
+                for w in counts_for_hist.word_to_count.keys():
+                    n_star_star += len(counts_for_hist.word_to_context[w])
+
+                if n_star_star != 0:
+                    for w in counts_for_hist.word_to_count.keys():
+                        n_star_z = len(counts_for_hist.word_to_context[w])
+                        counts_for_hist.word_to_f[w] = (
+                            max((n_star_z - self.d[n]), 0) * 1.0 / n_star_star
+                        )
+                else:  # patterns begin with <s>, they do not have "modified count", so use raw count instead
+                    for w in counts_for_hist.word_to_count.keys():
+                        n_star_z = counts_for_hist.word_to_count[w]
+                        counts_for_hist.word_to_f[w] = (
+                            max((n_star_z - self.d[n]), 0)
+                            * 1.0
+                            / counts_for_hist.total_count
+                        )
+
+    def cal_bow(self):
+        # Backoff weights are only necessary for ngrams which form a prefix of a longer ngram.
+        # Thus, two sorts of ngrams do not have a bow:
+        # 1) highest order ngram
+        # 2) ngrams ending in </s>
+        #
+        # bow(a_) = (1 - Sum_Z1 f(a_z)) / (1 - Sum_Z1 f(_z))
+        # Note that Z1 is the set of all words with c(a_z) > 0
+
+        # highest order N-grams
+        n = self.ngram_order - 1
+        this_order_counts = self.counts[n]
+        for hist, counts_for_hist in this_order_counts.items():
+            for w in counts_for_hist.word_to_count.keys():
+                counts_for_hist.word_to_bow[w] = None
+
+        # lower order N-grams
+        for n in range(0, self.ngram_order - 1):
+            this_order_counts = self.counts[n]
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    if w == self.eos_symbol:
+                        counts_for_hist.word_to_bow[w] = None
+                    else:
+                        a_ = hist + (w,)
+
+                        assert len(a_) < self.ngram_order
+                        assert a_ in self.counts[len(a_)].keys()
+
+                        a_counts_for_hist = self.counts[len(a_)][a_]
+
+                        sum_z1_f_a_z = 0
+                        for u in a_counts_for_hist.word_to_count.keys():
+                            sum_z1_f_a_z += a_counts_for_hist.word_to_f[u]
+
+                        sum_z1_f_z = 0
+                        _ = a_[1:]
+                        _counts_for_hist = self.counts[len(_)][_]
+                        # Should be careful here: what is Z1
+                        for u in a_counts_for_hist.word_to_count.keys():
+                            sum_z1_f_z += _counts_for_hist.word_to_f[u]
+
+                        if sum_z1_f_z < 1:
+                            # assert sum_z1_f_a_z < 1
+                            counts_for_hist.word_to_bow[w] = (1.0 - sum_z1_f_a_z) / (
+                                1.0 - sum_z1_f_z
+                            )
+                        else:
+                            counts_for_hist.word_to_bow[w] = None
+
+    def print_raw_counts(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    res.append(
+                        "{0}\t{1}".format(ngram, counts_for_hist.word_to_count[w])
+                    )
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_modified_counts(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    modified_count = len(counts_for_hist.word_to_context[w])
+                    raw_count = counts_for_hist.word_to_count[w]
+
+                    if modified_count == 0:
+                        res.append("{0}\t{1}".format(ngram, raw_count))
+                    else:
+                        res.append("{0}\t{1}".format(ngram, modified_count))
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_f(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    f = counts_for_hist.word_to_f[w]
+                    if f == 0:  # f(<s>) is always 0
+                        f = 1e-99
+
+                    res.append("{0}\t{1}".format(ngram, math.log(f, 10)))
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_f_and_bow(self, info_string):
+        # these are useful for debug.
+        print(info_string)
+        res = []
+        for this_order_counts in self.counts:
+            for hist, counts_for_hist in this_order_counts.items():
+                for w in counts_for_hist.word_to_count.keys():
+                    ngram = " ".join(hist) + " " + w
+                    ngram = ngram.strip(strip_chars)
+
+                    f = counts_for_hist.word_to_f[w]
+                    if f == 0:  # f(<s>) is always 0
+                        f = 1e-99
+
+                    bow = counts_for_hist.word_to_bow[w]
+                    if bow is None:
+                        res.append("{1}\t{0}".format(ngram, math.log(f, 10)))
+                    else:
+                        res.append(
+                            "{1}\t{0}\t{2}".format(
+                                ngram, math.log(f, 10), math.log(bow, 10)
+                            )
+                        )
+        res.sort(reverse=True)
+        for r in res:
+            print(r)
+
+    def print_as_arpa(
+        self, fout=io.TextIOWrapper(sys.stdout.buffer, encoding="latin-1")
+    ):
+        # print as ARPA format.
+
+        print("\\data\\", file=fout)
+        for hist_len in range(self.ngram_order):
+            # print the number of n-grams.
+            print(
+                "ngram {0}={1}".format(
+                    hist_len + 1,
+                    sum(
+                        [
+                            len(counts_for_hist.word_to_f)
+                            for counts_for_hist in self.counts[hist_len].values()
+                        ]
+                    ),
+                ),
+                file=fout,
+            )
+
+        print("", file=fout)
+
+        for hist_len in range(self.ngram_order):
+            print("\\{0}-grams:".format(hist_len + 1), file=fout)
+
+            this_order_counts = self.counts[hist_len]
+            for hist, counts_for_hist in this_order_counts.items():
+                for word in counts_for_hist.word_to_count.keys():
+                    ngram = hist + (word,)
+                    prob = counts_for_hist.word_to_f[word]
+                    bow = counts_for_hist.word_to_bow[word]
+
+                    if prob == 0:  # f(<s>) is always 0
+                        prob = 1e-99
+
+                    line = "{0}\t{1}".format("%.7f" % math.log10(prob), " ".join(ngram))
+                    if bow is not None:
+                        line += "\t{0}".format("%.7f" % math.log10(bow))
+                    print(line, file=fout)
+            print("", file=fout)
+        print("\\end\\", file=fout)
+
+
+if __name__ == "__main__":
+
+    ngram_counts = NgramCounts(args.ngram_order)
+
+    if args.text is None:
+        ngram_counts.add_raw_counts_from_standard_input()
+    else:
+        assert os.path.isfile(args.text)
+        ngram_counts.add_raw_counts_from_file(args.text)
+
+    ngram_counts.cal_discounting_constants()
+    ngram_counts.cal_f()
+    ngram_counts.cal_bow()
+
+    if args.lm is None:
+        ngram_counts.print_as_arpa()
+    else:
+        with open(args.lm, "w", encoding=default_encoding) as f:
+            ngram_counts.print_as_arpa(fout=f)

egs/tedlium2/ASR/shared/parse_options.sh

@@ -0,0 +1,97 @@
+#!/usr/bin/env bash
+
+# Copyright 2012  Johns Hopkins University (Author: Daniel Povey);
+#                 Arnab Ghoshal, Karel Vesely
+
+# 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
+#
+# THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
+# WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
+# MERCHANTABLITY OR NON-INFRINGEMENT.
+# See the Apache 2 License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Parse command-line options.
+# To be sourced by another script (as in ". parse_options.sh").
+# Option format is: --option-name arg
+# and shell variable "option_name" gets set to value "arg."
+# The exception is --help, which takes no arguments, but prints the
+# $help_message variable (if defined).
+
+
+###
+### The --config file options have lower priority to command line
+### options, so we need to import them first...
+###
+
+# Now import all the configs specified by command-line, in left-to-right order
+for ((argpos=1; argpos<$#; argpos++)); do
+  if [ "${!argpos}" == "--config" ]; then
+    argpos_plus1=$((argpos+1))
+    config=${!argpos_plus1}
+    [ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
+    . $config  # source the config file.
+  fi
+done
+
+
+###
+### Now we process the command line options
+###
+while true; do
+  [ -z "${1:-}" ] && break;  # break if there are no arguments
+  case "$1" in
+    # If the enclosing script is called with --help option, print the help
+    # message and exit.  Scripts should put help messages in $help_message
+    --help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
+      else printf "$help_message\n" 1>&2 ; fi;
+      exit 0 ;;
+    --*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
+      exit 1 ;;
+    # If the first command-line argument begins with "--" (e.g. --foo-bar),
+    # then work out the variable name as $name, which will equal "foo_bar".
+    --*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
+      # Next we test whether the variable in question is undefned-- if so it's
+      # an invalid option and we die.  Note: $0 evaluates to the name of the
+      # enclosing script.
+      # The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
+      # is undefined.  We then have to wrap this test inside "eval" because
+      # foo_bar is itself inside a variable ($name).
+      eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
+
+      oldval="`eval echo \\$$name`";
+      # Work out whether we seem to be expecting a Boolean argument.
+      if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
+        was_bool=true;
+      else
+        was_bool=false;
+      fi
+
+      # Set the variable to the right value-- the escaped quotes make it work if
+      # the option had spaces, like --cmd "queue.pl -sync y"
+      eval $name=\"$2\";
+
+      # Check that Boolean-valued arguments are really Boolean.
+      if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
+        echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
+        exit 1;
+      fi
+      shift 2;
+      ;;
+  *) break;
+  esac
+done
+
+
+# Check for an empty argument to the --cmd option, which can easily occur as a
+# result of scripting errors.
+[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
+
+
+true; # so this script returns exit code 0.

egs/tedlium2/ASR/transducer_stateless/README.md

@@ -0,0 +1,20 @@
+## Introduction
+
+The decoder, i.e., the prediction network, is from
+https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+(Rnn-Transducer with Stateless Prediction Network)
+
+You can use the following command to start the training:
+
+```bash
+cd egs/tedlium3/ASR
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --max-duration 300
+```

egs/tedlium2/ASR/transducer_stateless/asr_datamodule.py

@@ -0,0 +1,368 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2021  Xiaomi Corporation (Author: Mingshuang Luo)
+#
+# 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.
+
+
+import argparse
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+
+class TedLiumAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. TEDLium3 dev
+    and test).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=30,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset.",
+        )
+
+    def train_dataloaders(
+        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=10,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                    max_frames_mask_fraction=0.15,
+                    p=0.9,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to get Musan cuts")
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        logging.info("About to create train dataloader")
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts_test: CutSet) -> DataLoader:
+
+        logging.debug("About to create test dataset")
+        if self.args.on_the_fly_feats:
+            test = K2SpeechRecognitionDataset(
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            test = K2SpeechRecognitionDataset(
+                return_cuts=self.args.return_cuts,
+            )
+
+        test_sampler = DynamicBucketingSampler(
+            cuts_test,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=test_sampler,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "tedlium_cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "tedlium_cuts_test.jsonl.gz")

egs/tedlium2/ASR/transducer_stateless/beam_search.py

@@ -0,0 +1,539 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Mingshuang Luo)
+#
+# 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 dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import torch
+from model import Transducer
+
+
+def greedy_search(
+    model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int
+) -> List[int]:
+    """
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      max_sym_per_frame:
+        Maximum number of symbols per frame. If it is set to 0, the WER
+        would be 100%.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+
+    blank_id = model.decoder.blank_id
+    unk_id = model.decoder.unk_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor(
+        [blank_id] * context_size, device=device, dtype=torch.int64
+    ).reshape(1, context_size)
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+    hyp = [blank_id] * context_size
+
+    # Maximum symbols per utterance.
+    max_sym_per_utt = 1000
+
+    # symbols per frame
+    sym_per_frame = 0
+
+    # symbols per utterance decoded so far
+    sym_per_utt = 0
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        if sym_per_frame >= max_sym_per_frame:
+            sym_per_frame = 0
+            t += 1
+            continue
+
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        logits = model.joiner(
+            current_encoder_out, decoder_out, encoder_out_len, decoder_out_len
+        )
+        # logits is (1, 1, 1, vocab_size)
+
+        y = logits.argmax().item()
+        if y != blank_id and y != unk_id:
+            hyp.append(y)
+            decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape(
+                1, context_size
+            )
+
+            decoder_out = model.decoder(decoder_input, need_pad=False)
+
+            sym_per_utt += 1
+            sym_per_frame += 1
+        else:
+            sym_per_frame = 0
+            t += 1
+    hyp = hyp[context_size:]  # remove blanks
+
+    return hyp
+
+
+@dataclass
+class Hypothesis:
+    # The predicted tokens so far.
+    # Newly predicted tokens are appended to `ys`.
+    ys: List[int]
+
+    # The log prob of ys.
+    # It contains only one entry.
+    log_prob: torch.Tensor
+
+    @property
+    def key(self) -> str:
+        """Return a string representation of self.ys"""
+        return "_".join(map(str, self.ys))
+
+
+class HypothesisList(object):
+    def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None:
+        """
+        Args:
+          data:
+            A dict of Hypotheses. Its key is its `value.key`.
+        """
+        if data is None:
+            self._data = {}
+        else:
+            self._data = data
+
+    @property
+    def data(self) -> Dict[str, Hypothesis]:
+        return self._data
+
+    def add(self, hyp: Hypothesis) -> None:
+        """Add a Hypothesis to `self`.
+
+        If `hyp` already exists in `self`, its probability is updated using
+        `log-sum-exp` with the existed one.
+
+        Args:
+          hyp:
+            The hypothesis to be added.
+        """
+        key = hyp.key
+        if key in self:
+            old_hyp = self._data[key]  # shallow copy
+            torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob)
+        else:
+            self._data[key] = hyp
+
+    def get_most_probable(self, length_norm: bool = False) -> Hypothesis:
+        """Get the most probable hypothesis, i.e., the one with
+        the largest `log_prob`.
+
+        Args:
+          length_norm:
+            If True, the `log_prob` of a hypothesis is normalized by the
+            number of tokens in it.
+        Returns:
+          Return the hypothesis that has the largest `log_prob`.
+        """
+        if length_norm:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys))
+        else:
+            return max(self._data.values(), key=lambda hyp: hyp.log_prob)
+
+    def remove(self, hyp: Hypothesis) -> None:
+        """Remove a given hypothesis.
+
+        Caution:
+          `self` is modified **in-place**.
+
+        Args:
+          hyp:
+            The hypothesis to be removed from `self`.
+            Note: It must be contained in `self`. Otherwise,
+            an exception is raised.
+        """
+        key = hyp.key
+        assert key in self, f"{key} does not exist"
+        del self._data[key]
+
+    def filter(self, threshold: torch.Tensor) -> "HypothesisList":
+        """Remove all Hypotheses whose log_prob is less than threshold.
+
+        Caution:
+          `self` is not modified. Instead, a new HypothesisList is returned.
+
+        Returns:
+          Return a new HypothesisList containing all hypotheses from `self`
+          with `log_prob` being greater than the given `threshold`.
+        """
+        ans = HypothesisList()
+        for _, hyp in self._data.items():
+            if hyp.log_prob > threshold:
+                ans.add(hyp)  # shallow copy
+        return ans
+
+    def topk(self, k: int) -> "HypothesisList":
+        """Return the top-k hypothesis."""
+        hyps = list(self._data.items())
+
+        hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k]
+
+        ans = HypothesisList(dict(hyps))
+        return ans
+
+    def __contains__(self, key: str):
+        return key in self._data
+
+    def __iter__(self):
+        return iter(self._data.values())
+
+    def __len__(self) -> int:
+        return len(self._data)
+
+    def __str__(self) -> str:
+        s = []
+        for key in self:
+            s.append(key)
+        return ", ".join(s)
+
+
+def run_decoder(
+    ys: List[int],
+    model: Transducer,
+    decoder_cache: Dict[str, torch.Tensor],
+) -> torch.Tensor:
+    """Run the neural decoder model for a given hypothesis.
+
+    Args:
+      ys:
+        The current hypothesis.
+      model:
+        The transducer model.
+      decoder_cache:
+        Cache to save computations.
+    Returns:
+      Return a 1-D tensor of shape (decoder_out_dim,) containing
+      output of `model.decoder`.
+    """
+    context_size = model.decoder.context_size
+    key = "_".join(map(str, ys[-context_size:]))
+    if key in decoder_cache:
+        return decoder_cache[key]
+
+    device = model.device
+
+    decoder_input = torch.tensor([ys[-context_size:]], device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+    decoder_cache[key] = decoder_out
+
+    return decoder_out
+
+
+def run_joiner(
+    key: str,
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    decoder_out: torch.Tensor,
+    encoder_out_len: torch.Tensor,
+    decoder_out_len: torch.Tensor,
+    joint_cache: Dict[str, torch.Tensor],
+):
+    """Run the joint network given outputs from the encoder and decoder.
+
+    Args:
+      key:
+        A key into the `joint_cache`.
+      model:
+        The transducer model.
+      encoder_out:
+        A tensor of shape (1, 1, encoder_out_dim).
+      decoder_out:
+        A tensor of shape (1, 1, decoder_out_dim).
+      encoder_out_len:
+        A tensor with value [1].
+      decoder_out_len:
+        A tensor with value [1].
+      joint_cache:
+        A dict to save computations.
+    Returns:
+      Return a tensor from the output of log-softmax.
+      Its shape is (vocab_size,).
+    """
+    if key in joint_cache:
+        return joint_cache[key]
+
+    logits = model.joiner(
+        encoder_out,
+        decoder_out,
+        encoder_out_len,
+        decoder_out_len,
+    )
+
+    # TODO(fangjun): Scale the blank posterior
+    log_prob = logits.log_softmax(dim=-1)
+    # log_prob is (1, 1, 1, vocab_size)
+
+    log_prob = log_prob.squeeze()
+    # Now log_prob is (vocab_size,)
+
+    joint_cache[key] = log_prob
+
+    return log_prob
+
+
+def 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.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = model.decoder.unk_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor([blank_id] * context_size, device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+
+    B = HypothesisList()
+    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])
+
+    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 = list(B)
+        B = HypothesisList()
+
+        ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A])
+        # ys_log_probs is of shape (num_hyps, 1)
+
+        decoder_input = torch.tensor(
+            [hyp.ys[-context_size:] for hyp in A],
+            device=device,
+        )
+        # decoder_input is of shape (num_hyps, context_size)
+
+        decoder_out = model.decoder(decoder_input, need_pad=False)
+        # decoder_output is of shape (num_hyps, 1, decoder_output_dim)
+
+        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[:]
+            new_token = topk_token_indexes[i]
+            if new_token != blank_id and new_token != unk_id:
+                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)
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+
+    return ys
+
+
+def beam_search(
+    model: Transducer,
+    encoder_out: torch.Tensor,
+    beam: int = 4,
+) -> List[int]:
+    """
+    It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf
+
+    espnet/nets/beam_search_transducer.py#L247 is used as a reference.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder. Support only N==1 for now.
+      beam:
+        Beam size.
+    Returns:
+      Return the decoded result.
+    """
+    assert encoder_out.ndim == 3
+
+    # support only batch_size == 1 for now
+    assert encoder_out.size(0) == 1, encoder_out.size(0)
+    blank_id = model.decoder.blank_id
+    unk_id = model.decoder.unk_id
+    context_size = model.decoder.context_size
+
+    device = model.device
+
+    decoder_input = torch.tensor([blank_id] * context_size, device=device).reshape(
+        1, context_size
+    )
+
+    decoder_out = model.decoder(decoder_input, need_pad=False)
+
+    T = encoder_out.size(1)
+    t = 0
+
+    B = HypothesisList()
+    B.add(
+        Hypothesis(
+            ys=[blank_id] * context_size,
+            log_prob=torch.zeros(1, dtype=torch.float32, device=device),
+        )
+    )
+
+    max_sym_per_utt = 20000
+
+    sym_per_utt = 0
+
+    encoder_out_len = torch.tensor([1])
+    decoder_out_len = torch.tensor([1])
+
+    decoder_cache: Dict[str, torch.Tensor] = {}
+
+    while t < T and sym_per_utt < max_sym_per_utt:
+        # fmt: off
+        current_encoder_out = encoder_out[:, t:t+1, :]
+        # fmt: on
+        A = B
+        B = HypothesisList()
+
+        joint_cache: Dict[str, torch.Tensor] = {}
+
+        while True:
+            y_star = A.get_most_probable()
+            A.remove(y_star)
+
+            decoder_out = run_decoder(
+                ys=y_star.ys, model=model, decoder_cache=decoder_cache
+            )
+
+            key = "_".join(map(str, y_star.ys[-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,
+            )
+
+            # First, process the blank symbol
+            skip_log_prob = log_prob[blank_id]
+            new_y_star_log_prob = y_star.log_prob + skip_log_prob
+
+            # ys[:] returns a copy of ys
+            B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob))
+
+            # Second, process other non-blank labels
+            values, indices = log_prob.topk(beam + 1)
+            for idx in range(values.size(0)):
+                i = indices[idx].item()
+                if i == blank_id or i == unk_id:
+                    continue
+
+                new_ys = y_star.ys + [i]
+
+                new_log_prob = y_star.log_prob + values[idx]
+                A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob))
+
+            # Check whether B contains more than "beam" elements more probable
+            # than the most probable in A
+            A_most_probable = A.get_most_probable()
+
+            kept_B = B.filter(A_most_probable.log_prob)
+
+            if len(kept_B) >= beam:
+                B = kept_B.topk(beam)
+                break
+
+        t += 1
+
+    best_hyp = B.get_most_probable(length_norm=True)
+    ys = best_hyp.ys[context_size:]  # [context_size:] to remove blanks
+    return ys

egs/tedlium2/ASR/transducer_stateless/decode.py

@@ -0,0 +1,492 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+#                                            Mingshuang Luo)
+#
+# 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
+./transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 11 \
+        --exp-dir ./transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) beam search
+./transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 11 \
+        --exp-dir ./transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method beam_search \
+        --beam-size 4
+
+(3) modified beam search
+./transducer_stateless/decode.py \
+        --epoch 29 \
+        --avg 11 \
+        --exp-dir ./transducer_stateless/exp \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from beam_search import beam_search, greedy_search, modified_beam_search
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=13,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        beam_search or modified_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=3,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    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,
+        unk_id=params.unk_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,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
+    hyps = []
+    batch_size = encoder_out.size(0)
+
+    for i in range(batch_size):
+        # fmt: off
+        encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+        # fmt: on
+        if params.decoding_method == "greedy_search":
+            hyp = greedy_search(
+                model=model,
+                encoder_out=encoder_out_i,
+                max_sym_per_frame=params.max_sym_per_frame,
+            )
+        elif params.decoding_method == "beam_search":
+            hyp = beam_search(
+                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
+            )
+        else:
+            raise ValueError(f"Unsupported decoding method: {params.decoding_method}")
+        hyps.append(sp.decode(hyp).split())
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    else:
+        return {f"beam_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      sp:
+        The BPE model.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 100
+    else:
+        log_interval = 2
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((cut_id, ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        results = sorted(results)
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+    if "beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <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)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    # we need cut ids to display recognition results.
+    args.return_cuts = True
+    tedlium = TedLiumAsrDataModule(args)
+    dev_cuts = tedlium.dev_cuts()
+    test_cuts = tedlium.test_cuts()
+
+    dev_dl = tedlium.valid_dataloaders(dev_cuts)
+    test_dl = tedlium.test_dataloaders(test_cuts)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/transducer_stateless/decoder.py

@@ -0,0 +1,100 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+#                                                  Mingshuang Luo)
+#
+# 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.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        unk_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          unk_id:
+            The ID of the unk symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+        self.unk_id = unk_id
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        if context_size > 1:
+            self.conv = nn.Conv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            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).
+        """
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(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 embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        return embedding_out

egs/tedlium2/ASR/transducer_stateless/encoder_interface.py

@@ -0,0 +1,43 @@
+# 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.
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+
+class EncoderInterface(nn.Module):
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (batch_size, input_seq_len, num_features)
+            containing the input features.
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames
+            in `x` before padding.
+        Returns:
+          Return a tuple containing two tensors:
+            - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)
+              containing unnormalized probabilities, i.e., the output of a
+              linear layer.
+            - encoder_out_lens, a tensor of shape (batch_size,) containing
+              the number of frames in `encoder_out` before padding.
+        """
+        raise NotImplementedError("Please implement it in a subclass")

egs/tedlium2/ASR/transducer_stateless/export.py

@@ -0,0 +1,252 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
+# 					                         Mingshuang Luo)
+#
+# 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.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./transducer_stateless/export.py \
+  --exp-dir ./transducer_stateless/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 29 \
+  --avg 11
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `transducer_stateless/decode.py`, you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/tedlium3/ASR
+    ./transducer_stateless/decode.py \
+        --exp-dir ./transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=20,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=10,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    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) -> 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,
+        unk_id=params.unk_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 main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <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)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/tedlium2/ASR/transducer_stateless/joiner.py

@@ -0,0 +1,79 @@
+# 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.
+
+from typing import List
+
+import torch
+import torch.nn as nn
+
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, output_dim: int):
+        super().__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.output_linear = nn.Linear(input_dim, output_dim)
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+        encoder_out_len: torch.Tensor,
+        decoder_out_len: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, self.input_dim).
+          decoder_out:
+            Output from the decoder. Its shape is (N, U, self.input_dim).
+          encoder_out_len:
+            A 1-D tensor of shape (N,) containing valid number of frames
+            before padding in `encoder_out`.
+          decoder_out_len:
+            A 1-D tensor of shape (N,) containing valid number of frames
+            before padding in `decoder_out`.
+        Returns:
+          Return a tensor of shape (sum_all_TU, self.output_dim).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 3
+        assert encoder_out.size(0) == decoder_out.size(0)
+        assert encoder_out.size(2) == self.input_dim
+        assert decoder_out.size(2) == self.input_dim
+
+        N = encoder_out.size(0)
+
+        encoder_out_len: List[int] = encoder_out_len.tolist()
+        decoder_out_len: List[int] = decoder_out_len.tolist()
+
+        encoder_out_list = [encoder_out[i, : encoder_out_len[i], :] for i in range(N)]
+
+        decoder_out_list = [decoder_out[i, : decoder_out_len[i], :] for i in range(N)]
+
+        x = [
+            e.unsqueeze(1) + d.unsqueeze(0)
+            for e, d in zip(encoder_out_list, decoder_out_list)
+        ]
+
+        x = [p.reshape(-1, self.input_dim) for p in x]
+        x = torch.cat(x)
+
+        activations = torch.tanh(x)
+
+        logits = self.output_linear(activations)
+
+        return logits

egs/tedlium2/ASR/transducer_stateless/model.py

@@ -0,0 +1,143 @@
+# 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.
+
+import random
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        modified_transducer_prob: float = 0.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          modified_transducer_prob:
+            The probability to use modified transducer loss.
+        Returns:
+          Return the transducer loss.
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+        sos_y_padded = sos_y_padded.to(torch.int64)
+
+        decoder_out = self.decoder(sos_y_padded)
+
+        # +1 here since a blank is prepended to each utterance.
+        logits = self.joiner(
+            encoder_out=encoder_out,
+            decoder_out=decoder_out,
+            encoder_out_len=x_lens,
+            decoder_out_len=y_lens + 1,
+        )
+
+        # rnnt_loss requires 0 padded targets
+        # Note: y does not start with SOS
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        # We don't put this `import` at the beginning of the file
+        # as it is required only in the training, not during the
+        # reference stage
+        import optimized_transducer
+
+        assert 0 <= modified_transducer_prob <= 1
+
+        if modified_transducer_prob == 0:
+            one_sym_per_frame = False
+        elif random.random() < modified_transducer_prob:
+            # random.random() returns a float in the range [0, 1)
+            one_sym_per_frame = True
+        else:
+            one_sym_per_frame = False
+
+        loss = optimized_transducer.transducer_loss(
+            logits=logits,
+            targets=y_padded,
+            logit_lengths=x_lens,
+            target_lengths=y_lens,
+            blank=blank_id,
+            reduction="sum",
+            one_sym_per_frame=one_sym_per_frame,
+            from_log_softmax=False,
+        )
+
+        return loss

egs/tedlium2/ASR/transducer_stateless/pretrained.py

@@ -0,0 +1,337 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		 2022  Xiaomi Crop.        (authors: Mingshuang Luo)
+#
+# 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
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/exp/pretrained.pt \
+        --bpe-model ./data/lang_bpe_500/bpe.model \
+        --method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(2) beam search
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/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
+./transducer_stateless/pretrained.py \
+        --checkpoint ./transducer_stateless/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
+
+You can also use `./transducer_stateless/exp/epoch-xx.pt`.
+
+Note: ./transducer_stateless/exp/pretrained.pt is generated by
+./transducer_stateless/export.py
+"""
+
+
+import argparse
+import logging
+import math
+from typing import List
+
+import kaldifeat
+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 conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from model import Transducer
+from torch.nn.utils.rnn import pad_sequence
+
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict
+
+
+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.
+        Used only when method is ctc-decoding.
+        """,
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_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(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="Used only when --method is beam_search and modified_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=3,
+        help="""Maximum number of symbols per frame. Used only when
+        --method is greedy_search.
+        """,
+    )
+
+    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,
+        unk_id=params.unk_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]:
+    """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}. 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> 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(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)
+
+    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)
+
+    with torch.no_grad():
+        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)
+    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
+            )
+        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}")
+
+        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/tedlium2/ASR/transducer_stateless/subsampling.py

@@ -0,0 +1,153 @@
+# 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.
+
+
+import torch
+import torch.nn as nn
+
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2),
+            nn.ReLU(),
+        )
+        self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True)
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x

egs/tedlium2/ASR/transducer_stateless/test_decoder.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						   Mingshuang Luo)
+#
+# 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.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/tedlium3/ASR
+    python ./transducer_stateless/test_decoder.py
+"""
+
+import torch
+from decoder import Decoder
+
+
+def test_decoder():
+    vocab_size = 3
+    blank_id = 0
+    unk_id = 2
+    embedding_dim = 128
+    context_size = 4
+
+    decoder = Decoder(
+        vocab_size=vocab_size,
+        embedding_dim=embedding_dim,
+        blank_id=blank_id,
+        unk_id=unk_id,
+        context_size=context_size,
+    )
+    N = 100
+    U = 20
+    x = torch.randint(low=0, high=vocab_size, size=(N, U))
+    y = decoder(x)
+    assert y.shape == (N, U, embedding_dim)
+
+    # for inference
+    x = torch.randint(low=0, high=vocab_size, size=(N, context_size))
+    y = decoder(x, need_pad=False)
+    assert y.shape == (N, 1, embedding_dim)
+
+
+def main():
+    test_decoder()
+
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/transducer_stateless/train.py

@@ -0,0 +1,737 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# 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:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./transducer_stateless/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir transducer_stateless/exp \
+  --max-duration 300
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import TedLiumAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from local.convert_transcript_words_to_bpe_ids import convert_texts_into_ids
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    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(
+        "--modified-transducer-prob",
+        type=float,
+        default=0.25,
+        help="""The probability to use modified transducer loss.
+        In modified transduer, it limits the maximum number of symbols
+        per frame to 1. See also the option --max-sym-per-frame in
+        transducer_stateless/decode.py
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # 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,
+            # parameters for Noam
+            "warm_step": 80000,  # For the 100h subset, use 8k
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # 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) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.encoder_out_dim,
+        blank_id=params.blank_id,
+        unk_id=params.unk_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 load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    unk_id = params.unk_id
+    y = convert_texts_into_ids(texts, unk_id, sp=sp)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            modified_transducer_prob=params.modified_transducer_prob,
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+        if batch_idx % params.log_interval == 0:
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <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)
+
+    logging.info("About to create 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}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    tedlium = TedLiumAsrDataModule(args)
+
+    train_cuts = tedlium.train_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 17 seconds
+        return 1.0 <= c.duration <= 17.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    train_dl = tedlium.train_dataloaders(train_cuts)
+    valid_cuts = tedlium.dev_cuts()
+    valid_dl = tedlium.valid_dataloaders(valid_cuts)
+
+    scan_pessimistic_batches_for_oom(
+        model=model,
+        train_dl=train_dl,
+        optimizer=optimizer,
+        sp=sp,
+        params=params,
+    )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/learning_rate", cur_lr, params.batch_idx_train)
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            optimizer.zero_grad()
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+            )
+            loss.backward()
+            clip_grad_norm_(model.parameters(), 5.0, 2.0)
+            optimizer.step()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    TedLiumAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/tedlium2/ASR/transducer_stateless/transformer.py

@@ -0,0 +1,416 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# 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.
+
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+from icefall.utils import make_pad_mask
+
+
+class Transformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          dropout:
+            Dropout in encoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.output_dim = output_dim
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        lengths = ((x_lens - 1) // 2 - 1) // 2
+        assert x.size(0) == lengths.max().item()
+
+        mask = make_pad_mask(lengths)
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)