diff --git a/.github/workflows/style_check.yml b/.github/workflows/style_check.yml index 3c971e231d..908c3cc430 100644 --- a/.github/workflows/style_check.yml +++ b/.github/workflows/style_check.yml @@ -69,7 +69,7 @@ jobs: working-directory: ${{github.workspace}} run: | black --check --diff . - + - name: Run isort shell: bash working-directory: ${{github.workspace}} diff --git a/egs/multi_ja_en/ASR/README.md b/egs/multi_ja_en/ASR/README.md new file mode 100644 index 0000000000..09964a4ab1 --- /dev/null +++ b/egs/multi_ja_en/ASR/README.md @@ -0,0 +1,17 @@ +# Introduction + +A bilingual Japanese-English ASR model that utilizes ReazonSpeech, developed by the developers of ReazonSpeech. + +**ReazonSpeech** is an open-source dataset that contains a diverse set of natural Japanese speech, collected from terrestrial television streams. It contains more than 35,000 hours of audio. + + +# Included Training Sets + +1. LibriSpeech (English) +2. ReazonSpeech (Japanese) + +|Datset| Number of hours| URL| +|---|---:|---| +|**TOTAL**|35,960|---| +|LibriSpeech|960|https://www.openslr.org/12/| +|ReazonSpeech (all) |35,000|https://huggingface.co/datasets/reazon-research/reazonspeech| diff --git a/egs/multi_ja_en/ASR/RESULTS.md b/egs/multi_ja_en/ASR/RESULTS.md new file mode 100644 index 0000000000..c1bbb9ac5e --- /dev/null +++ b/egs/multi_ja_en/ASR/RESULTS.md @@ -0,0 +1,53 @@ +## Results + +### Zipformer + +#### Non-streaming + +The training command is: + +```shell +./zipformer/train.py \ + --bilingual 1 \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --max-duration 600 +``` + +The decoding command is: + +```shell +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search +``` + +To export the model with onnx: + +```shell +./zipformer/export-onnx.py --tokens data/lang_bbpe_2000/tokens.txt --use-averaged-model 0 --epoch 35 --avg 1 --exp-dir zipformer/exp --num-encoder-layers "2,2,3,4,3,2" --downsampling-factor "1,2,4,8,4,2" --feedforward-dim "512,768,1024,1536,1024,768" --num-heads "4,4,4,8,4,4" --encoder-dim "192,256,384,512,384,256" --query-head-dim 32 --value-head-dim 12 --pos-head-dim 4 --pos-dim 48 --encoder-unmasked-dim "192,192,256,256,256,192" --cnn-module-kernel "31,31,15,15,15,31" --decoder-dim 512 --joiner-dim 512 --causal False --chunk-size "16,32,64,-1" --left-context-frames "64,128,256,-1" --fp16 True +``` +Word Error Rates (WERs) listed below: + +| Datasets | ReazonSpeech | ReazonSpeech | LibriSpeech | LibriSpeech | +|----------------------|--------------|---------------|--------------------|-------------------| +| Zipformer WER (%) | dev | test | test-clean | test-other | +| greedy_search | 5.9 | 4.07 | 3.46 | 8.35 | +| modified_beam_search | 4.87 | 3.61 | 3.28 | 8.07 | +| fast_beam_search | 41.04 | 36.59 | 16.14 | 22.0 | + + +Character Error Rates (CERs) for Japanese listed below: +| Decoding Method | In-Distribution CER | JSUT | CommonVoice | TEDx | +| :------------------: | :-----------------: | :--: | :---------: | :---: | +| greedy search | 12.56 | 6.93 | 9.75 | 9.67 | +| modified beam search | 11.59 | 6.97 | 9.55 | 9.51 | + +Pre-trained model can be found here: https://huggingface.co/reazon-research/reazonspeech-k2-v2-ja-en/tree/main + diff --git a/egs/multi_ja_en/ASR/local/compute_fbank_reazonspeech.py b/egs/multi_ja_en/ASR/local/compute_fbank_reazonspeech.py new file mode 100644 index 0000000000..af78414063 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/compute_fbank_reazonspeech.py @@ -0,0 +1,146 @@ +#!/usr/bin/env python3 +# Copyright 2023 The University of Electro-Communications (Author: Teo Wen Shen) # noqa +# +# 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 +import os +from pathlib import Path +from typing import List, Tuple + +import torch + +# fmt: off +from lhotse import ( # See the following for why LilcomChunkyWriter is preferred; https://github.com/k2-fsa/icefall/pull/404; https://github.com/lhotse-speech/lhotse/pull/527 + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + RecordingSet, + SupervisionSet, +) + +# fmt: on + +# 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) + +RNG_SEED = 42 +concat_params = {"gap": 1.0, "maxlen": 10.0} + + +def make_cutset_blueprints( + manifest_dir: Path, +) -> List[Tuple[str, CutSet]]: + cut_sets = [] + + # Create test dataset + logging.info("Creating test cuts.") + cut_sets.append( + ( + "test", + CutSet.from_manifests( + recordings=RecordingSet.from_file( + manifest_dir / "reazonspeech_recordings_test.jsonl.gz" + ), + supervisions=SupervisionSet.from_file( + manifest_dir / "reazonspeech_supervisions_test.jsonl.gz" + ), + ), + ) + ) + + # Create dev dataset + logging.info("Creating dev cuts.") + cut_sets.append( + ( + "dev", + CutSet.from_manifests( + recordings=RecordingSet.from_file( + manifest_dir / "reazonspeech_recordings_dev.jsonl.gz" + ), + supervisions=SupervisionSet.from_file( + manifest_dir / "reazonspeech_supervisions_dev.jsonl.gz" + ), + ), + ) + ) + + # Create train dataset + logging.info("Creating train cuts.") + cut_sets.append( + ( + "train", + CutSet.from_manifests( + recordings=RecordingSet.from_file( + manifest_dir / "reazonspeech_recordings_train.jsonl.gz" + ), + supervisions=SupervisionSet.from_file( + manifest_dir / "reazonspeech_supervisions_train.jsonl.gz" + ), + ), + ) + ) + return cut_sets + + +def get_args(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument("-m", "--manifest-dir", type=Path) + return parser.parse_args() + + +def main(): + args = get_args() + + extractor = Fbank(FbankConfig(num_mel_bins=80)) + num_jobs = min(16, os.cpu_count()) + + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + if (args.manifest_dir / ".reazonspeech-fbank.done").exists(): + logging.info( + "Previous fbank computed for ReazonSpeech found. " + f"Delete {args.manifest_dir / '.reazonspeech-fbank.done'} to allow recomputing fbank." + ) + return + else: + cut_sets = make_cutset_blueprints(args.manifest_dir) + for part, cut_set in cut_sets: + logging.info(f"Processing {part}") + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + num_jobs=num_jobs, + storage_path=(args.manifest_dir / f"feats_{part}").as_posix(), + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(args.manifest_dir / f"reazonspeech_cuts_{part}.jsonl.gz") + + logging.info("All fbank computed for ReazonSpeech.") + (args.manifest_dir / ".reazonspeech-fbank.done").touch() + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/local/display_manifest_statistics.py b/egs/multi_ja_en/ASR/local/display_manifest_statistics.py new file mode 100644 index 0000000000..ace1dd73f5 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/display_manifest_statistics.py @@ -0,0 +1,58 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# 2022 The University of Electro-Communications (author: Teo Wen Shen) # noqa +# +# 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 +from pathlib import Path + +from lhotse import CutSet, load_manifest + +ARGPARSE_DESCRIPTION = """ +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 +pruned_transducer_stateless5/train.py for usage. +""" + + +def get_parser(): + parser = argparse.ArgumentParser( + description=ARGPARSE_DESCRIPTION, + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + + parser.add_argument("--manifest-dir", type=Path, help="Path to cutset manifests") + + return parser.parse_args() + + +def main(): + args = get_parser() + + for part in ["train", "dev"]: + path = args.manifest_dir / f"reazonspeech_cuts_{part}.jsonl.gz" + cuts: CutSet = load_manifest(path) + + print("\n---------------------------------\n") + print(path.name + ":") + cuts.describe() + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/local/prepare_char.py b/egs/multi_ja_en/ASR/local/prepare_char.py new file mode 120000 index 0000000000..42743b5449 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/prepare_char.py @@ -0,0 +1 @@ +../../../aishell/ASR/local/prepare_char.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/local/prepare_for_bpe_model.py b/egs/multi_ja_en/ASR/local/prepare_for_bpe_model.py new file mode 100755 index 0000000000..27832ad1bc --- /dev/null +++ b/egs/multi_ja_en/ASR/local/prepare_for_bpe_model.py @@ -0,0 +1,66 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Zengrui Jin) +# +# 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 tokenizes the training transcript by CJK characters +# and saves the result to transcript_chars.txt, which is used +# to train the BPE model later. + +import argparse +import re +from pathlib import Path + +from tqdm.auto import tqdm + +from icefall.utils import tokenize_by_ja_char + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Output directory. + The generated transcript_chars.txt is saved to this directory. + """, + ) + + parser.add_argument( + "--text", + type=str, + help="Training transcript.", + ) + + return parser.parse_args() + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + text = Path(args.text) + + assert lang_dir.exists() and text.exists(), f"{lang_dir} or {text} does not exist!" + + transcript_path = lang_dir / "transcript_chars.txt" + + with open(text, "r", encoding="utf-8") as fin: + with open(transcript_path, "w+", encoding="utf-8") as fout: + for line in tqdm(fin): + fout.write(tokenize_by_ja_char(line) + "\n") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/local/prepare_lang.py b/egs/multi_ja_en/ASR/local/prepare_lang.py new file mode 120000 index 0000000000..747f2ab398 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/prepare_lang.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/prepare_lang.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/local/prepare_lang_bbpe.py b/egs/multi_ja_en/ASR/local/prepare_lang_bbpe.py new file mode 100755 index 0000000000..6134710add --- /dev/null +++ b/egs/multi_ja_en/ASR/local/prepare_lang_bbpe.py @@ -0,0 +1,268 @@ +#!/usr/bin/env python3 +# 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. + + +""" + +This script takes as input `lang_dir`, which should contain:: + + - lang_dir/bbpe.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 +import re +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.byte_utils import byte_encode +from icefall.utils import str2bool, tokenize_by_ja_char + + +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 is defined in local/train_bpe_model.py + assert token2id[""] == 0 + assert word2id[""] == 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], oov: 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. + oov: + The out of vocabulary word in lexicon. + 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. + encode_words = [byte_encode(tokenize_by_ja_char(w)) for w in words] + words_pieces_ids: List[List[int]] = sp.encode(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)) + + lexicon.append((oov, ["▁", sp.id_to_piece(sp.unk_id())])) + + token2id: Dict[str, int] = {sp.id_to_piece(i): i for i in range(sp.vocab_size())} + + 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( + "--oov", + type=str, + default="", + help="The out of vocabulary word in lexicon.", + ) + + 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 / "bbpe.model" + + word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt") + + words = word_sym_table.symbols + + excluded = ["", "!SIL", "", args.oov, "#0", "", ""] + + for w in excluded: + if w in words: + words.remove(w) + + lexicon, token_sym_table = generate_lexicon(model_file, words, args.oov) + + 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("") + word_sym_table.add("") + + 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() diff --git a/egs/multi_ja_en/ASR/local/prepare_lang_char.py b/egs/multi_ja_en/ASR/local/prepare_lang_char.py new file mode 100644 index 0000000000..19c5f4a318 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/prepare_lang_char.py @@ -0,0 +1,75 @@ +#!/usr/bin/env python3 +# Copyright 2022 The University of Electro-Communications (Author: Teo Wen Shen) # noqa +# +# 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 pathlib import Path + +from lhotse import CutSet + + +def get_args(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + + parser.add_argument( + "train_cut", metavar="train-cut", type=Path, help="Path to the train cut" + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default=Path("data/lang_char"), + help=( + "Name of lang dir. " + "If not set, this will default to lang_char_{trans-mode}" + ), + ) + + return parser.parse_args() + + +def main(): + args = get_args() + logging.basicConfig( + format=("%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"), + level=logging.INFO, + ) + + sysdef_string = set(["", "", "", " "]) + + token_set = set() + logging.info(f"Creating vocabulary from {args.train_cut}.") + train_cut: CutSet = CutSet.from_file(args.train_cut) + for cut in train_cut: + for sup in cut.supervisions: + token_set.update(sup.text) + + token_set = [""] + sorted(token_set - sysdef_string) + ["", ""] + args.lang_dir.mkdir(parents=True, exist_ok=True) + (args.lang_dir / "tokens.txt").write_text( + "\n".join(f"{t}\t{i}" for i, t in enumerate(token_set)) + ) + + (args.lang_dir / "lang_type").write_text("char") + logging.info("Done.") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/local/prepare_words.py b/egs/multi_ja_en/ASR/local/prepare_words.py new file mode 120000 index 0000000000..ef2b4eaf35 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/prepare_words.py @@ -0,0 +1 @@ +../../../aishell2/ASR/local/prepare_words.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/local/text2segments.py b/egs/multi_ja_en/ASR/local/text2segments.py new file mode 100644 index 0000000000..e0f3a15c44 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/text2segments.py @@ -0,0 +1,95 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- + +# Copyright 2021 Xiaomi Corp. (authors: Mingshuang Luo) +# 2022 Xiaomi Corp. (authors: Weiji Zhuang) +# +# 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 "text", which refers to the transcript file: + - text +and generates the output file with word segmentation implemented using MeCab: + - text_words_segmentation +""" + +import argparse +from multiprocessing import Pool + +import MeCab +from tqdm import tqdm + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Japanese Word Segmentation for text", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--num-process", + "-n", + default=20, + type=int, + help="the number of processes", + ) + parser.add_argument( + "--input-file", + "-i", + default="data/lang_char/text", + type=str, + help="the input text file", + ) + parser.add_argument( + "--output-file", + "-o", + default="data/lang_char/text_words_segmentation", + type=str, + help="the text implemented with word segmentation using MeCab", + ) + + return parser + + +def cut(lines): + if lines is not None: + mecab = MeCab.Tagger("-Owakati") # Use '-Owakati' option for word segmentation + segmented_line = mecab.parse(lines).strip() + return segmented_line.split() # Return as a list of words + else: + return None + + +def main(): + parser = get_parser() + args = parser.parse_args() + + num_process = args.num_process + input_file = args.input_file + output_file = args.output_file + + with open(input_file, "r", encoding="utf-8") as fr: + lines = fr.readlines() + + with Pool(processes=num_process) as p: + new_lines = list(tqdm(p.imap(cut, lines), total=len(lines))) + + with open(output_file, "w", encoding="utf-8") as fw: + for line in new_lines: + fw.write(" ".join(line) + "\n") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/local/text2token.py b/egs/multi_ja_en/ASR/local/text2token.py new file mode 100755 index 0000000000..ce64847c93 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/text2token.py @@ -0,0 +1,177 @@ +#!/usr/bin/env python3 +# Copyright 2017 Johns Hopkins University (authors: Shinji Watanabe) +# 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. + + +import argparse +import codecs +import re +import sys +from typing import List + +from romkan import to_roma # Replace with python-romkan v0.2.1 + +is_python2 = sys.version_info[0] == 2 + + +def exist_or_not(i, match_pos): + start_pos = None + end_pos = None + for pos in match_pos: + if pos[0] <= i < pos[1]: + start_pos = pos[0] + end_pos = pos[1] + break + + return start_pos, end_pos + + +def get_parser(): + parser = argparse.ArgumentParser( + description="convert raw text to tokenized text", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--nchar", + "-n", + default=1, + type=int, + help="number of characters to split, i.e., \ + aabb -> a a b b with -n 1 and aa bb with -n 2", + ) + parser.add_argument( + "--skip-ncols", "-s", default=0, type=int, help="skip first n columns" + ) + parser.add_argument("--space", default="", type=str, help="space symbol") + parser.add_argument( + "--non-lang-syms", + "-l", + default=None, + type=str, + help="list of non-linguistic symbols, e.g., etc.", + ) + parser.add_argument("text", type=str, default=False, nargs="?", help="input text") + parser.add_argument( + "--trans_type", + "-t", + type=str, + default="char", + choices=["char", "romaji"], + help="Transcript type. char/romaji", + ) + return parser + + +def token2id( + texts, token_table, token_type: str = "romaji", oov: str = "" +) -> List[List[int]]: + """Convert token to id. + Args: + texts: + The input texts, it refers to the Japanese text here. + token_table: + The token table is built based on "data/lang_xxx/token.txt" + token_type: + The type of token, such as "romaji". + oov: + Out of vocabulary token. When a word(token) in the transcript + does not exist in the token list, it is replaced with `oov`. + + Returns: + The list of ids for the input texts. + """ + if texts is None: + raise ValueError("texts can't be None!") + else: + oov_id = token_table[oov] + ids: List[List[int]] = [] + for text in texts: + chars_list = list(str(text)) + if token_type == "romaji": + text = [to_roma(c) for c in chars_list] + sub_ids = [ + token_table[txt] if txt in token_table else oov_id for txt in text + ] + ids.append(sub_ids) + return ids + + +def main(): + parser = get_parser() + args = parser.parse_args() + + rs = [] + if args.non_lang_syms is not None: + with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f: + nls = [x.rstrip() for x in f.readlines()] + rs = [re.compile(re.escape(x)) for x in nls] + + if args.text: + f = codecs.open(args.text, encoding="utf-8") + else: + f = codecs.getreader("utf-8")(sys.stdin if is_python2 else sys.stdin.buffer) + + sys.stdout = codecs.getwriter("utf-8")( + sys.stdout if is_python2 else sys.stdout.buffer + ) + line = f.readline() + n = args.nchar + while line: + x = line.split() + print(" ".join(x[: args.skip_ncols]), end=" ") + a = " ".join(x[args.skip_ncols :]) # noqa E203 + + # get all matched positions + match_pos = [] + for r in rs: + i = 0 + while i >= 0: + m = r.search(a, i) + if m: + match_pos.append([m.start(), m.end()]) + i = m.end() + else: + break + if len(match_pos) > 0: + chars = [] + i = 0 + while i < len(a): + start_pos, end_pos = exist_or_not(i, match_pos) + if start_pos is not None: + chars.append(a[start_pos:end_pos]) + i = end_pos + else: + chars.append(a[i]) + i += 1 + a = chars + + if args.trans_type == "romaji": + a = [to_roma(c) for c in list(str(a))] + + a = [a[j : j + n] for j in range(0, len(a), n)] # noqa E203 + + a_flat = [] + for z in a: + a_flat.append("".join(z)) + + a_chars = "".join(a_flat) + print(a_chars) + line = f.readline() + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/local/train_bbpe_model.py b/egs/multi_ja_en/ASR/local/train_bbpe_model.py new file mode 100755 index 0000000000..d104f2717d --- /dev/null +++ b/egs/multi_ja_en/ASR/local/train_bbpe_model.py @@ -0,0 +1,114 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# 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 re +import shutil +import tempfile +from pathlib import Path + +import sentencepiece as spm + +from icefall import byte_encode +from icefall.utils import tokenize_by_ja_char + + +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 _convert_to_bchar(in_path: str, out_path: str): + with open(out_path, "w") as f: + for line in open(in_path, "r").readlines(): + f.write(byte_encode(tokenize_by_ja_char(line)) + "\n") + + +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}" + model_file = Path(model_prefix + ".model") + if model_file.is_file(): + print(f"{model_file} exists - skipping") + return + + character_coverage = 1.0 + input_sentence_size = 100000000 + + user_defined_symbols = ["", ""] + 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. + + temp = tempfile.NamedTemporaryFile() + train_text = temp.name + + _convert_to_bchar(args.transcript, train_text) + + 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, + ) + + shutil.copyfile(model_file, f"{lang_dir}/bbpe.model") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/local/utils/asr_datamodule.py b/egs/multi_ja_en/ASR/local/utils/asr_datamodule.py new file mode 100644 index 0000000000..be18e65c18 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/utils/asr_datamodule.py @@ -0,0 +1,355 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 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 inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, List, Optional + +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SimpleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import OnTheFlyFeatures +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class ReazonSpeechAsrDataModule: + """ + DataModule for k2 ASR experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + 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/dev/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( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=False, + 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=False, + 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. + """ + + transforms = [] + 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}") + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + 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, + ) + + 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=self.args.drop_last, + ) + else: + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + 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: CutSet) -> DataLoader: + logging.info("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + 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 / "reazonspeech_cuts_train.jsonl.gz" + ) + + @lru_cache() + def valid_cuts(self) -> CutSet: + logging.info("About to get dev cuts") + return load_manifest_lazy( + self.args.manifest_dir / "reazonspeech_cuts_dev.jsonl.gz" + ) + + @lru_cache() + def test_cuts(self) -> List[CutSet]: + logging.info("About to get test cuts") + return load_manifest_lazy( + self.args.manifest_dir / "reazonspeech_cuts_test.jsonl.gz" + ) diff --git a/egs/multi_ja_en/ASR/local/utils/tokenizer.py b/egs/multi_ja_en/ASR/local/utils/tokenizer.py new file mode 100644 index 0000000000..ba71cff893 --- /dev/null +++ b/egs/multi_ja_en/ASR/local/utils/tokenizer.py @@ -0,0 +1,252 @@ +import argparse +from pathlib import Path +from typing import Callable, List, Union + +import sentencepiece as spm +from k2 import SymbolTable + + +class Tokenizer: + text2word: Callable[[str], List[str]] + + @staticmethod + def add_arguments(parser: argparse.ArgumentParser): + group = parser.add_argument_group(title="Lang related options") + group.add_argument("--lang", type=Path, help="Path to lang directory.") + + group.add_argument( + "--lang-type", + type=str, + default=None, + help=( + "Either 'bpe' or 'char'. If not provided, it expects lang_dir/lang_type to exists. " + "Note: 'bpe' directly loads sentencepiece.SentencePieceProcessor" + ), + ) + + @staticmethod + def Load(lang_dir: Path, lang_type="", oov=""): + + if not lang_type: + assert (lang_dir / "lang_type").exists(), "lang_type not specified." + lang_type = (lang_dir / "lang_type").read_text().strip() + + tokenizer = None + + if lang_type == "bpe": + assert ( + lang_dir / "bpe.model" + ).exists(), f"No BPE .model could be found in {lang_dir}." + tokenizer = spm.SentencePieceProcessor() + tokenizer.Load(str(lang_dir / "bpe.model")) + elif lang_type == "char": + tokenizer = CharTokenizer(lang_dir, oov=oov) + else: + raise NotImplementedError(f"{lang_type} not supported at the moment.") + + return tokenizer + + load = Load + + def PieceToId(self, piece: str) -> int: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + piece_to_id = PieceToId + + def IdToPiece(self, id: int) -> str: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + id_to_piece = IdToPiece + + def GetPieceSize(self) -> int: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + get_piece_size = GetPieceSize + + def __len__(self) -> int: + return self.get_piece_size() + + def EncodeAsIdsBatch(self, input: List[str]) -> List[List[int]]: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + def EncodeAsPiecesBatch(self, input: List[str]) -> List[List[str]]: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + def EncodeAsIds(self, input: str) -> List[int]: + return self.EncodeAsIdsBatch([input])[0] + + def EncodeAsPieces(self, input: str) -> List[str]: + return self.EncodeAsPiecesBatch([input])[0] + + def Encode( + self, input: Union[str, List[str]], out_type=int + ) -> Union[List, List[List]]: + if not input: + return [] + + if isinstance(input, list): + if out_type is int: + return self.EncodeAsIdsBatch(input) + if out_type is str: + return self.EncodeAsPiecesBatch(input) + + if out_type is int: + return self.EncodeAsIds(input) + if out_type is str: + return self.EncodeAsPieces(input) + + encode = Encode + + def DecodeIdsBatch(self, input: List[List[int]]) -> List[str]: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + def DecodePiecesBatch(self, input: List[List[str]]) -> List[str]: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + def DecodeIds(self, input: List[int]) -> str: + return self.DecodeIdsBatch([input])[0] + + def DecodePieces(self, input: List[str]) -> str: + return self.DecodePiecesBatch([input])[0] + + def Decode( + self, + input: Union[int, List[int], List[str], List[List[int]], List[List[str]]], + ) -> Union[List[str], str]: + + if not input: + return "" + + if isinstance(input, int): + return self.id_to_piece(input) + elif isinstance(input, str): + raise TypeError( + "Unlike spm.SentencePieceProcessor, cannot decode from type str." + ) + + if isinstance(input[0], list): + if not input[0] or isinstance(input[0][0], int): + return self.DecodeIdsBatch(input) + + if isinstance(input[0][0], str): + return self.DecodePiecesBatch(input) + + if isinstance(input[0], int): + return self.DecodeIds(input) + if isinstance(input[0], str): + return self.DecodePieces(input) + + raise RuntimeError("Unknown input type") + + decode = Decode + + def SplitBatch(self, input: List[str]) -> List[List[str]]: + raise NotImplementedError( + "You need to implement this function in the child class." + ) + + def Split(self, input: Union[List[str], str]) -> Union[List[List[str]], List[str]]: + if isinstance(input, list): + return self.SplitBatch(input) + elif isinstance(input, str): + return self.SplitBatch([input])[0] + raise RuntimeError("Unknown input type") + + split = Split + + +class CharTokenizer(Tokenizer): + def __init__(self, lang_dir: Path, oov="", sep=""): + assert ( + lang_dir / "tokens.txt" + ).exists(), f"tokens.txt could not be found in {lang_dir}." + token_table = SymbolTable.from_file(lang_dir / "tokens.txt") + assert ( + "#0" not in token_table + ), "This tokenizer does not support disambig symbols." + self._id2sym = token_table._id2sym + self._sym2id = token_table._sym2id + self.oov = oov + self.oov_id = self._sym2id[oov] + self.sep = sep + if self.sep: + self.text2word = lambda x: x.split(self.sep) + else: + self.text2word = lambda x: list(x.replace(" ", "")) + + def piece_to_id(self, piece: str) -> int: + try: + return self._sym2id[piece] + except KeyError: + return self.oov_id + + def id_to_piece(self, id: int) -> str: + return self._id2sym[id] + + def get_piece_size(self) -> int: + return len(self._sym2id) + + def EncodeAsIdsBatch(self, input: List[str]) -> List[List[int]]: + return [[self.piece_to_id(i) for i in self.text2word(text)] for text in input] + + def EncodeAsPiecesBatch(self, input: List[str]) -> List[List[str]]: + return [ + [i if i in self._sym2id else self.oov for i in self.text2word(text)] + for text in input + ] + + def DecodeIdsBatch(self, input: List[List[int]]) -> List[str]: + return [self.sep.join(self.id_to_piece(i) for i in text) for text in input] + + def DecodePiecesBatch(self, input: List[List[str]]) -> List[str]: + return [self.sep.join(text) for text in input] + + def SplitBatch(self, input: List[str]) -> List[List[str]]: + return [self.text2word(text) for text in input] + + +def test_CharTokenizer(): + test_single_string = "こんにちは" + test_multiple_string = [ + "今日はいい天気ですよね", + "諏訪湖は綺麗でしょう", + "这在词表外", + "分かち 書き に し た 文章 です", + "", + ] + test_empty_string = "" + sp = Tokenizer.load(Path("lang_char"), "char", oov="") + splitter = sp.split + print(sp.encode(test_single_string, out_type=str)) + print(sp.encode(test_single_string, out_type=int)) + print(sp.encode(test_multiple_string, out_type=str)) + print(sp.encode(test_multiple_string, out_type=int)) + print(sp.encode(test_empty_string, out_type=str)) + print(sp.encode(test_empty_string, out_type=int)) + print(sp.decode(sp.encode(test_single_string, out_type=str))) + print(sp.decode(sp.encode(test_single_string, out_type=int))) + print(sp.decode(sp.encode(test_multiple_string, out_type=str))) + print(sp.decode(sp.encode(test_multiple_string, out_type=int))) + print(sp.decode(sp.encode(test_empty_string, out_type=str))) + print(sp.decode(sp.encode(test_empty_string, out_type=int))) + print(splitter(test_single_string)) + print(splitter(test_multiple_string)) + print(splitter(test_empty_string)) + + +if __name__ == "__main__": + test_CharTokenizer() diff --git a/egs/multi_ja_en/ASR/local/validate_bpe_lexicon.py b/egs/multi_ja_en/ASR/local/validate_bpe_lexicon.py new file mode 120000 index 0000000000..721bb48e7c --- /dev/null +++ b/egs/multi_ja_en/ASR/local/validate_bpe_lexicon.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/validate_bpe_lexicon.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/local/validate_manifest.py b/egs/multi_ja_en/ASR/local/validate_manifest.py new file mode 100644 index 0000000000..7f67c64b6c --- /dev/null +++ b/egs/multi_ja_en/ASR/local/validate_manifest.py @@ -0,0 +1,96 @@ +#!/usr/bin/env python3 +# Copyright 2022 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 checks the following assumptions of the generated manifest: + +- Single supervision per cut +- Supervision time bounds are within cut time bounds + +We will add more checks later if needed. + +Usage example: + + python3 ./local/validate_manifest.py \ + ./data/fbank/librispeech_cuts_train-clean-100.jsonl.gz + +""" + +import argparse +import logging +from pathlib import Path + +from lhotse import CutSet, load_manifest +from lhotse.cut import Cut + + +def get_args(): + parser = argparse.ArgumentParser() + + parser.add_argument( + "--manifest", + type=Path, + help="Path to the manifest file", + ) + + return parser.parse_args() + + +def validate_one_supervision_per_cut(c: Cut): + if len(c.supervisions) != 1: + raise ValueError(f"{c.id} has {len(c.supervisions)} supervisions") + + +def validate_supervision_and_cut_time_bounds(c: Cut): + s = c.supervisions[0] + + # Removed because when the cuts were trimmed from supervisions, + # the start time of the supervision can be lesser than cut start time. + # https://github.com/lhotse-speech/lhotse/issues/813 + # if s.start < c.start: + # raise ValueError( + # f"{c.id}: Supervision start time {s.start} is less " + # f"than cut start time {c.start}" + # ) + + if s.end > c.end: + raise ValueError( + f"{c.id}: Supervision end time {s.end} is larger " + f"than cut end time {c.end}" + ) + + +def main(): + args = get_args() + + manifest = Path(args.manifest) + logging.info(f"Validating {manifest}") + + assert manifest.is_file(), f"{manifest} does not exist" + cut_set = load_manifest(manifest) + assert isinstance(cut_set, CutSet) + + for c in cut_set: + validate_one_supervision_per_cut(c) + validate_supervision_and_cut_time_bounds(c) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + main() diff --git a/egs/multi_ja_en/ASR/prepare.sh b/egs/multi_ja_en/ASR/prepare.sh new file mode 100755 index 0000000000..7a6a634183 --- /dev/null +++ b/egs/multi_ja_en/ASR/prepare.sh @@ -0,0 +1,185 @@ +#!/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 + +stage=-1 +stop_stage=100 + +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +vocab_sizes=( + 2000 +) + +# 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" + +log "Dataset: musan" +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Soft link fbank of musan" + mkdir -p data/fbank + if [ -e ../../librispeech/ASR/data/fbank/.musan.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_feats) . + ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_cuts.jsonl.gz) . + cd ../.. + else + log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 4 --stop-stage 4" + exit 1 + fi +fi + +log "Dataset: LibriSpeech" +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 1: Soft link fbank of LibriSpeech" + mkdir -p data/fbank + if [ -e ../../librispeech/ASR/data/fbank/.librispeech.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/librispeech_cuts*) . + ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/librispeech_feats*) . + cd ../.. + else + log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 1 --stop-stage 1 and ../../librispeech/ASR/prepare.sh --stage 3 --stop-stage 3" + exit 1 + fi +fi + +log "Dataset: ReazonSpeech" +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 2: Soft link fbank of ReazonSpeech" + mkdir -p data/fbank + if [ -e ../../reazonspeech/ASR/data/manifests/.reazonspeech.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../reazonspeech/ASR/data/manifests/reazonspeech_cuts*) . + cd .. + mkdir -p manifests + cd manifests + ln -svf $(realpath ../../../../reazonspeech/ASR/data/manifests/feats_*) . + cd ../.. + else + log "Abort! Please run ../../reazonspeech/ASR/prepare.sh --stage 0 --stop-stage 2" + exit 1 + fi +fi + +# New Stage 3: Prepare char based lang for ReazonSpeech +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + lang_char_dir=data/lang_char + log "Stage 3: Prepare char based lang for ReazonSpeech" + mkdir -p $lang_char_dir + + # Prepare text + if [ ! -f $lang_char_dir/text ]; then + gunzip -c ../../reazonspeech/ASR/data/manifests/reazonspeech_supervisions_train.jsonl.gz \ + | jq '.text' | sed 's/"//g' \ + | ./local/text2token.py -t "char" > $lang_char_dir/text + fi + + # jp word segmentation for text + if [ ! -f $lang_char_dir/text_words_segmentation ]; then + python3 ./local/text2segments.py \ + --input-file $lang_char_dir/text \ + --output-file $lang_char_dir/text_words_segmentation + fi + + cat $lang_char_dir/text_words_segmentation | sed 's/ /\n/g' \ + | sort -u | sed '/^$/d' | uniq > $lang_char_dir/words_no_ids.txt + + if [ ! -f $lang_char_dir/words.txt ]; then + python3 ./local/prepare_words.py \ + --input-file $lang_char_dir/words_no_ids.txt \ + --output-file $lang_char_dir/words.txt + fi + + if [ ! -f $lang_char_dir/L_disambig.pt ]; then + python3 ./local/prepare_char.py --lang-dir data/lang_char + fi +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Prepare Byte BPE based lang" + mkdir -p data/fbank + if [ ! -d ../../reazonspeech/ASR/data/lang_char ] && [ ! -d ./data/lang_char ]; then + log "Abort! Please run ../../reazonspeech/ASR/prepare.sh --stage 3 --stop-stage 3" + exit 1 + fi + + if [ ! -d ../../librispeech/ASR/data/lang_bpe_500 ] && [ ! -d ./data/lang_bpe_500 ]; then + log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 5 --stop-stage 5" + exit 1 + fi + + cd data/ + # if [ ! -d ./lang_char ]; then + # ln -svf $(realpath ../../../reazonspeech/ASR/data/lang_char) . + # fi + if [ ! -d ./lang_bpe_500 ]; then + ln -svf $(realpath ../../../librispeech/ASR/data/lang_bpe_500) . + fi + cd ../ + + for vocab_size in ${vocab_sizes[@]}; do + lang_dir=data/lang_bbpe_${vocab_size} + mkdir -p $lang_dir + + cat data/lang_char/text data/lang_bpe_500/transcript_words.txt \ + > $lang_dir/text + + if [ ! -f $lang_dir/transcript_chars.txt ]; then + ./local/prepare_for_bpe_model.py \ + --lang-dir ./$lang_dir \ + --text $lang_dir/text + fi + + if [ ! -f $lang_dir/text_words_segmentation ]; then + python3 ./local/text2segments.py \ + --input-file ./data/lang_char/text \ + --output-file $lang_dir/text_words_segmentation + + cat ./data/lang_bpe_500/transcript_words.txt \ + >> $lang_dir/text_words_segmentation + fi + + cat $lang_dir/text_words_segmentation | sed 's/ /\n/g' \ + | sort -u | sed '/^$/d' | uniq > $lang_dir/words_no_ids.txt + + if [ ! -f $lang_dir/words.txt ]; then + python3 ./local/prepare_words.py \ + --input-file $lang_dir/words_no_ids.txt \ + --output-file $lang_dir/words.txt + fi + + if [ ! -f $lang_dir/bbpe.model ]; then + ./local/train_bbpe_model.py \ + --lang-dir $lang_dir \ + --vocab-size $vocab_size \ + --transcript $lang_dir/text + fi + + if [ ! -f $lang_dir/L_disambig.pt ]; then + ./local/prepare_lang_bbpe.py --lang-dir $lang_dir + + log "Validating $lang_dir/lexicon.txt" + ln -svf $(realpath ../../multi_zh_en/ASR/local/validate_bpe_lexicon.py) local/ + ./local/validate_bpe_lexicon.py \ + --lexicon $lang_dir/lexicon.txt \ + --bpe-model $lang_dir/bbpe.model + fi + done +fi + +log "prepare.sh: PREPARATION DONE" diff --git a/egs/multi_ja_en/ASR/shared b/egs/multi_ja_en/ASR/shared new file mode 120000 index 0000000000..4c5e91438c --- /dev/null +++ b/egs/multi_ja_en/ASR/shared @@ -0,0 +1 @@ +../../../icefall/shared/ \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/asr_datamodule.py b/egs/multi_ja_en/ASR/zipformer/asr_datamodule.py new file mode 120000 index 0000000000..a48591198f --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../local/utils/asr_datamodule.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/beam_search.py b/egs/multi_ja_en/ASR/zipformer/beam_search.py new file mode 120000 index 0000000000..8e2c0a65c5 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/beam_search.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/ctc_decode.py b/egs/multi_ja_en/ASR/zipformer/ctc_decode.py new file mode 120000 index 0000000000..faa8bd562e --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/ctc_decode.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/ctc_decode.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/decode.py b/egs/multi_ja_en/ASR/zipformer/decode.py new file mode 100755 index 0000000000..26ce3e018c --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/decode.py @@ -0,0 +1,792 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# 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 +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + +import argparse +import logging +import math +import re +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 ReazonSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from train import add_model_arguments, get_model, get_params + +from icefall import byte_encode, smart_byte_decode +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + tokenize_by_ja_char, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +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 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bbpe_2000/bbpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bbpe_2000", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + 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, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + 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""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + 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`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).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) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, 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(smart_byte_decode(hyp).split()) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + 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, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in hyp_tokens: + hyps.append([word_table[i] for i in hyp]) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + 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, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(smart_byte_decode(hyp).split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + 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, + num_paths=params.num_paths, + ref_texts=sp.encode(byte_encode(tokenize_by_ja_char(supervisions["text"]))), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(smart_byte_decode(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(smart_byte_decode(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(smart_byte_decode(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(smart_byte_decode(sp.decode(hyp)).split()) + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: 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, + word_table: Optional[k2.SymbolTable] = None, + 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. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + 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 = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + texts = [tokenize_by_ja_char(str(text)).split() for text in texts] + # print(texts) + # exit() + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + 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): + this_batch.append((cut_id, ref_text, 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() + ReazonSpeechAsrDataModule.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", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + 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}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + 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) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + 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)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = 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 + data_module = ReazonSpeechAsrDataModule(args) + multi_dataset = MultiDataset(args) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}" + ) + return T > 0 + + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dl = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dl): + logging.info(f"Start decoding test set: {test_set}") + + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_ja_en/ASR/zipformer/decode_stream.py b/egs/multi_ja_en/ASR/zipformer/decode_stream.py new file mode 120000 index 0000000000..b8d8ddfc4c --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/decode_stream.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decode_stream.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/decoder.py b/egs/multi_ja_en/ASR/zipformer/decoder.py new file mode 120000 index 0000000000..5a8018680d --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/do_not_use_it_directly.py b/egs/multi_ja_en/ASR/zipformer/do_not_use_it_directly.py new file mode 100755 index 0000000000..072679cfc9 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/do_not_use_it_directly.py @@ -0,0 +1,1261 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# 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_stateless7_streaming/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_streaming/exp \ + --lang data/lang_char \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_streaming/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_streaming/exp \ + --lang data/lang_char \ + --max-duration 550 +""" + + +import argparse +import copy +import logging +import math +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import ReazonSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from tokenizer import Tokenizer +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer_for_ncnn_export_only import Zipformer + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] +LOG_EPS = math.log(1e-10) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +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=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=Path, + default="pruned_transducer_stateless7_streaming/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + 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", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--pad-feature", + type=int, + default=0, + help=""" + Number of frames to pad at the end. + """, + ) + + add_model_arguments(parser) + + 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. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + 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": 1000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + is_pnnx=True, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=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, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` 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. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + 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] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = 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. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + 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: Union[nn.Module, DDP], + sp: Tokenizer, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer 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 Zipformer 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. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).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) + + if params.pad_feature: + feature_lens += params.pad_feature + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.pad_feature), + value=LOG_EPS, + ) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + 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, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + 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: Union[nn.Module, DDP], + sp: Tokenizer, + 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: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: Tokenizer, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> 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. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + 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. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except Exception as e: # noqa + logging.error(e, exc_info=True) + display_and_save_batch(batch, params=params, sp=sp) + raise e + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + 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 params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + 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() + log_mode = logging.info + log_mode(f"Epoch {params.cur_epoch}, validation: {valid_info}") + log_mode( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + 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, master_port=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 = Tokenizer.load(args.lang, args.lang_type) + + # is defined in local/prepare_lang_char.py + params.blank_id = sp.piece_to_id("") + 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}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 0.3 or c.duration > 30.0: + logging.debug( + f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = sp.encode(c.supervisions[0].text, out_type=str) + + if T < len(tokens): + logging.info( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + reazonspeech_corpus = ReazonSpeechAsrDataModule(args) + train_cuts = reazonspeech_corpus.train_cuts() + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = reazonspeech_corpus.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = reazonspeech_corpus.valid_cuts() + valid_dl = reazonspeech_corpus.valid_dataloaders(valid_cuts) + + if params.start_batch <= 0 and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: Tokenizer, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: Tokenizer, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception 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]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + raise RuntimeError("Please don't use this file directly!") + parser = get_parser() + ReazonSpeechAsrDataModule.add_arguments(parser) + Tokenizer.add_arguments(parser) + args = parser.parse_args() + + 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() diff --git a/egs/multi_ja_en/ASR/zipformer/encoder_interface.py b/egs/multi_ja_en/ASR/zipformer/encoder_interface.py new file mode 120000 index 0000000000..c2eaca6712 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/export-onnx.py b/egs/multi_ja_en/ASR/zipformer/export-onnx.py new file mode 120000 index 0000000000..70a15683c2 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/export-onnx.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/export.py b/egs/multi_ja_en/ASR/zipformer/export.py new file mode 120000 index 0000000000..dfc1bec080 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/export.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/generate_averaged_model.py b/egs/multi_ja_en/ASR/zipformer/generate_averaged_model.py new file mode 120000 index 0000000000..5a015ee6c1 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/generate_averaged_model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/generate_averaged_model.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/joiner.py b/egs/multi_ja_en/ASR/zipformer/joiner.py new file mode 120000 index 0000000000..5b8a36332e --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/model.py b/egs/multi_ja_en/ASR/zipformer/model.py new file mode 120000 index 0000000000..cd7e07d72b --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/multi_dataset.py b/egs/multi_ja_en/ASR/zipformer/multi_dataset.py new file mode 100644 index 0000000000..b0cdc1f6a2 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/multi_dataset.py @@ -0,0 +1,143 @@ +import argparse +import logging +from functools import lru_cache +from pathlib import Path +from typing import Dict + +from lhotse import CutSet, load_manifest_lazy + + +class MultiDataset: + def __init__(self, args: argparse.Namespace): + """ + Args: + manifest_dir: + It is expected to contain the following files: + - reazonspeech_cuts_train.jsonl.gz + - librispeech_cuts_train-clean-100.jsonl.gz + - librispeech_cuts_train-clean-360.jsonl.gz + - librispeech_cuts_train-other-500.jsonl.gz + """ + self.fbank_dir = Path(args.manifest_dir) + + def train_cuts(self) -> CutSet: + logging.info("About to get multidataset train cuts") + + logging.info("Loading Reazonspeech in lazy mode") + reazonspeech_cuts = load_manifest_lazy( + self.fbank_dir / "reazonspeech_cuts_train.jsonl.gz" + ) + + logging.info("Loading LibriSpeech in lazy mode") + train_clean_100_cuts = self.train_clean_100_cuts() + train_clean_360_cuts = self.train_clean_360_cuts() + train_other_500_cuts = self.train_other_500_cuts() + + return CutSet.mux( + reazonspeech_cuts, + train_clean_100_cuts, + train_clean_360_cuts, + train_other_500_cuts, + weights=[ + len(reazonspeech_cuts), + len(train_clean_100_cuts), + len(train_clean_360_cuts), + len(train_other_500_cuts), + ], + ) + + def dev_cuts(self) -> CutSet: + logging.info("About to get multidataset dev cuts") + + logging.info("Loading Reazonspeech DEV set in lazy mode") + reazonspeech_dev_cuts = load_manifest_lazy( + self.fbank_dir / "reazonspeech_cuts_dev.jsonl.gz" + ) + + logging.info("Loading LibriSpeech DEV set in lazy mode") + dev_clean_cuts = self.dev_clean_cuts() + dev_other_cuts = self.dev_other_cuts() + + return CutSet.mux( + reazonspeech_dev_cuts, + dev_clean_cuts, + dev_other_cuts, + weights=[ + len(reazonspeech_dev_cuts), + len(dev_clean_cuts), + len(dev_other_cuts), + ], + ) + + def test_cuts(self) -> Dict[str, CutSet]: + logging.info("About to get multidataset test cuts") + + logging.info("Loading Reazonspeech set in lazy mode") + reazonspeech_test_cuts = load_manifest_lazy( + self.fbank_dir / "reazonspeech_cuts_test.jsonl.gz" + ) + reazonspeech_dev_cuts = load_manifest_lazy( + self.fbank_dir / "reazonspeech_cuts_dev.jsonl.gz" + ) + + logging.info("Loading LibriSpeech set in lazy mode") + test_clean_cuts = self.test_clean_cuts() + test_other_cuts = self.test_other_cuts() + + test_cuts = { + "reazonspeech_test": reazonspeech_test_cuts, + "reazonspeech_dev": reazonspeech_dev_cuts, + "librispeech_test_clean": test_clean_cuts, + "librispeech_test_other": test_other_cuts, + } + + return test_cuts + + @lru_cache() + def train_clean_100_cuts(self) -> CutSet: + logging.info("About to get train-clean-100 cuts") + return load_manifest_lazy( + self.fbank_dir / "librispeech_cuts_train-clean-100.jsonl.gz" + ) + + @lru_cache() + def train_clean_360_cuts(self) -> CutSet: + logging.info("About to get train-clean-360 cuts") + return load_manifest_lazy( + self.fbank_dir / "librispeech_cuts_train-clean-360.jsonl.gz" + ) + + @lru_cache() + def train_other_500_cuts(self) -> CutSet: + logging.info("About to get train-other-500 cuts") + return load_manifest_lazy( + self.fbank_dir / "librispeech_cuts_train-other-500.jsonl.gz" + ) + + @lru_cache() + def dev_clean_cuts(self) -> CutSet: + logging.info("About to get dev-clean cuts") + return load_manifest_lazy( + self.fbank_dir / "librispeech_cuts_dev-clean.jsonl.gz" + ) + + @lru_cache() + def dev_other_cuts(self) -> CutSet: + logging.info("About to get dev-other cuts") + return load_manifest_lazy( + self.fbank_dir / "librispeech_cuts_dev-other.jsonl.gz" + ) + + @lru_cache() + def test_clean_cuts(self) -> CutSet: + logging.info("About to get test-clean cuts") + return load_manifest_lazy( + self.fbank_dir / "librispeech_cuts_test-clean.jsonl.gz" + ) + + @lru_cache() + def test_other_cuts(self) -> CutSet: + logging.info("About to get test-other cuts") + return load_manifest_lazy( + self.fbank_dir / "librispeech_cuts_test-other.jsonl.gz" + ) diff --git a/egs/multi_ja_en/ASR/zipformer/my_profile.py b/egs/multi_ja_en/ASR/zipformer/my_profile.py new file mode 120000 index 0000000000..3a90b26289 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/my_profile.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/my_profile.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/onnx_decode.py b/egs/multi_ja_en/ASR/zipformer/onnx_decode.py new file mode 120000 index 0000000000..0573b88c5b --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/onnx_decode.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_decode.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/onnx_pretrained.py b/egs/multi_ja_en/ASR/zipformer/onnx_pretrained.py new file mode 120000 index 0000000000..8f32f4ee7a --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/onnx_pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/optim.py b/egs/multi_ja_en/ASR/zipformer/optim.py new file mode 120000 index 0000000000..5eaa3cffd4 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/pretrained.py b/egs/multi_ja_en/ASR/zipformer/pretrained.py new file mode 120000 index 0000000000..0bd71dde4d --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/pretrained.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/scaling.py b/egs/multi_ja_en/ASR/zipformer/scaling.py new file mode 120000 index 0000000000..6f398f431d --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/scaling_converter.py b/egs/multi_ja_en/ASR/zipformer/scaling_converter.py new file mode 120000 index 0000000000..b0ecee05e1 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/streaming_beam_search.py b/egs/multi_ja_en/ASR/zipformer/streaming_beam_search.py new file mode 120000 index 0000000000..b1ed545579 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/streaming_beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/streaming_decode.py b/egs/multi_ja_en/ASR/zipformer/streaming_decode.py new file mode 100755 index 0000000000..935f86de18 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/streaming_decode.py @@ -0,0 +1,935 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang, +# Fangjun Kuang, +# Zengwei Yao) +# 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: + +Monolingual: +./zipformer/streaming_decode.py \ + --epoch 28 \ + --avg 15 \ + --causal 1 \ + --chunk-size 32 \ + --left-context-frames 256 \ + --exp-dir ./zipformer/exp-large \ + --lang data/lang_char \ + --num-encoder-layers 2,2,4,5,4,2 \ + --feedforward-dim 512,768,1536,2048,1536,768 \ + --encoder-dim 192,256,512,768,512,256 \ + --encoder-unmasked-dim 192,192,256,320,256,192 + +Bilingual: +./zipformer/streaming_decode.py \ + --bilingual 1 \ + --epoch 28 \ + --avg 15 \ + --causal 1 \ + --chunk-size 32 \ + --left-context-frames 256 \ + --exp-dir ./zipformer/exp-large \ + --lang data/lang_char \ + --num-encoder-layers 2,2,4,5,4,2 \ + --feedforward-dim 512,768,1536,2048,1536,768 \ + --encoder-dim 192,256,512,768,512,256 \ + --encoder-unmasked-dim 192,192,256,320,256,192 \ + +""" + +import argparse +import logging +import math +import os +import pdb +import subprocess as sp +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import numpy as np +import sentencepiece as spm +import torch +from asr_datamodule import ReazonSpeechAsrDataModule +from decode_stream import DecodeStream +from kaldifeat import Fbank, FbankOptions +from lhotse import CutSet +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from streaming_beam_search import ( + fast_beam_search_one_best, + greedy_search, + modified_beam_search, +) +from tokenizer import Tokenizer +from torch import Tensor, nn +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--bilingual", + type=str2bool, + default=False, + help="Whether the model is bilingual or not. 1 = bilingual.", + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/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( + "--lang-dir", + type=Path, + default="data/lang_char", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Supported decoding methods are: + greedy_search + modified_beam_search + fast_beam_search + """, + ) + + parser.add_argument( + "--num_active_paths", + type=int, + default=4, + help="""An interger indicating how many candidates we will keep for each + frame. Used only when --decoding-method is 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=32, + 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( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel.", + ) + + add_model_arguments(parser) + + return parser + + +def get_init_states( + model: nn.Module, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), +) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = model.encoder.get_init_states(batch_size, device) + + embed_states = model.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device) + states.append(processed_lens) + + return states + + +def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]: + """Stack list of zipformer states that correspond to separate utterances + into a single emformer state, so that it can be used as an input for + zipformer when those utterances are formed into a batch. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. For element-n, + state_list[n] is a list of cached tensors of all encoder layers. For layer-i, + state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, + cached_val2, cached_conv1, cached_conv2). + state_list[n][-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + state_list[n][-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Note: + It is the inverse of :func:`unstack_states`. + """ + batch_size = len(state_list) + assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0]) + tot_num_layers = (len(state_list[0]) - 2) // 6 + + batch_states = [] + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key = torch.cat( + [state_list[i][layer_offset] for i in range(batch_size)], dim=1 + ) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn = torch.cat( + [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1 = torch.cat( + [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2 = torch.cat( + [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1 = torch.cat( + [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2 = torch.cat( + [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0 + ) + batch_states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + cached_embed_left_pad = torch.cat( + [state_list[i][-2] for i in range(batch_size)], dim=0 + ) + batch_states.append(cached_embed_left_pad) + + processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0) + batch_states.append(processed_lens) + + return batch_states + + +def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]: + """Unstack the zipformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Note: + It is the inverse of :func:`stack_states`. + + Args: + batch_states: A list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + state_list[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Returns: + state_list: A list of list. Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. + """ + assert (len(batch_states) - 2) % 6 == 0, len(batch_states) + tot_num_layers = (len(batch_states) - 2) // 6 + + processed_lens = batch_states[-1] + batch_size = processed_lens.shape[0] + + state_list = [[] for _ in range(batch_size)] + + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk( + chunks=batch_size, dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1_list = batch_states[layer_offset + 2].chunk( + chunks=batch_size, dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2_list = batch_states[layer_offset + 3].chunk( + chunks=batch_size, dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1_list = batch_states[layer_offset + 4].chunk( + chunks=batch_size, dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2_list = batch_states[layer_offset + 5].chunk( + chunks=batch_size, dim=0 + ) + for i in range(batch_size): + state_list[i] += [ + cached_key_list[i], + cached_nonlin_attn_list[i], + cached_val1_list[i], + cached_val2_list[i], + cached_conv1_list[i], + cached_conv2_list[i], + ] + + cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(cached_embed_left_pad_list[i]) + + processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(processed_lens_list[i]) + + return state_list + + +def streaming_forward( + features: Tensor, + feature_lens: Tensor, + model: nn.Module, + states: List[Tensor], + chunk_size: int, + left_context_len: int, +) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Returns encoder outputs, output lengths, and updated states. + """ + cached_embed_left_pad = states[-2] + (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward( + x=features, + x_lens=feature_lens, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == chunk_size, (x.size(1), chunk_size) + + src_key_padding_mask = make_pad_mask(x_lens) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + encoder_states = states[:-2] + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = model.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + return encoder_out, encoder_out_lens, new_states + + +def decode_one_chunk( + params: AttributeDict, + model: nn.Module, + decode_streams: List[DecodeStream], +) -> List[int]: + """Decode one chunk frames of features for each decode_streams and + return the indexes of finished streams in a List. + + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + decode_streams: + A List of DecodeStream, each belonging to a utterance. + Returns: + Return a List containing which DecodeStreams are finished. + """ + chunk_size = int(params.chunk_size) + left_context_len = int(params.left_context_frames) + + features = [] + feature_lens = [] + states = [] + processed_lens = [] # Used in fast-beam-search + + for stream in decode_streams: + feat, feat_len = stream.get_feature_frames(chunk_size * 2) + features.append(feat) + feature_lens.append(feat_len) + states.append(stream.states) + processed_lens.append(stream.done_frames) + + feature_lens = torch.tensor(feature_lens, device=model.device) + features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS) + + # Make sure the length after encoder_embed is at least 1. + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + tail_length = chunk_size * 2 + 7 + 2 * 3 + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPS, + ) + + states = stack_states(states) + + encoder_out, encoder_out_lens, new_states = streaming_forward( + features=features, + feature_lens=feature_lens, + model=model, + states=states, + chunk_size=chunk_size, + left_context_len=left_context_len, + ) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + if params.decoding_method == "greedy_search": + greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams) + elif params.decoding_method == "fast_beam_search": + processed_lens = torch.tensor(processed_lens, device=model.device) + processed_lens = processed_lens + encoder_out_lens + fast_beam_search_one_best( + model=model, + encoder_out=encoder_out, + processed_lens=processed_lens, + streams=decode_streams, + beam=params.beam, + max_states=params.max_states, + max_contexts=params.max_contexts, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=decode_streams, + encoder_out=encoder_out, + num_active_paths=params.num_active_paths, + ) + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + + states = unstack_states(new_states) + + finished_streams = [] + for i in range(len(decode_streams)): + decode_streams[i].states = states[i] + decode_streams[i].done_frames += encoder_out_lens[i] + if decode_streams[i].done: + finished_streams.append(i) + # finished_streams.append(i) + + return finished_streams + + +def decode_dataset( + cuts: CutSet, + params: AttributeDict, + model: nn.Module, + sp: Tokenizer, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + cuts: + Lhotse Cutset 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. + """ + device = model.device + + opts = FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + + log_interval = 100 + + decode_results = [] + # Contain decode streams currently running. + decode_streams = [] + for num, cut in enumerate(cuts): + # each utterance has a DecodeStream. + initial_states = get_init_states(model=model, batch_size=1, device=device) + decode_stream = DecodeStream( + params=params, + cut_id=cut.id, + initial_states=initial_states, + decoding_graph=decoding_graph, + device=device, + ) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + + # The trained model is using normalized samples + # - this is to avoid sending [-32k,+32k] signal in... + # - some lhotse AudioTransform classes can make the signal + # be out of range [-1, 1], hence the tolerance 10 + assert ( + np.abs(audio).max() <= 10 + ), "Should be normalized to [-1, 1], 10 for tolerance..." + + samples = torch.from_numpy(audio).squeeze(0) + + fbank = Fbank(opts) + feature = fbank(samples.to(device)) + decode_stream.set_features(feature, tail_pad_len=30) + decode_stream.ground_truth = cut.supervisions[0].text + + decode_streams.append(decode_stream) + + while len(decode_streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + decode_streams[i].id, + decode_streams[i].ground_truth.split(), + sp.decode(decode_streams[i].decoding_result()).split(), + ) + ) + del decode_streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + # decode final chunks of last sequences + while len(decode_streams): + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + + if not finished_streams: + print("No finished streams, breaking the loop") + break + + for i in sorted(finished_streams, reverse=True): + try: + decode_results.append( + ( + decode_streams[i].id, + decode_streams[i].ground_truth.split(), + sp.decode(decode_streams[i].decoding_result()).split(), + ) + ) + del decode_streams[i] + except IndexError as e: + print(f"IndexError: {e}") + print(f"decode_streams length: {len(decode_streams)}") + print(f"finished_streams: {finished_streams}") + print(f"i: {i}") + continue + + if params.decoding_method == "greedy_search": + key = "greedy_search" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ) + elif params.decoding_method == "modified_beam_search": + key = f"num_active_paths_{params.num_active_paths}" + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + torch.cuda.synchronize() + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[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() + ReazonSpeechAsrDataModule.add_arguments(parser) + Tokenizer.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + params.res_dir = params.exp_dir / "streaming" / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + assert params.causal, params.causal + assert "," not in params.chunk_size, "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + # for fast_beam_search + if params.decoding_method == "fast_beam_search": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + 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}") + + if not params.bilingual: + sp = Tokenizer.load(params.lang, params.lang_type) + else: + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + 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)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + model.device = device + + decoding_graph = None + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, 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 + reazonspeech_corpus = ReazonSpeechAsrDataModule(args) + + if params.bilingual: + multi_dataset = MultiDataset(args) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}" + ) + return T > 0 + + test_sets_cuts = multi_dataset.test_cuts() + test_sets = test_sets_cuts.keys() + test_cuts = [test_sets_cuts[k] for k in test_sets] + + valid_cuts = reazonspeech_corpus.valid_cuts() + test_cuts = reazonspeech_corpus.test_cuts() + + test_sets = ["valid", "test"] + test_cuts = [valid_cuts, test_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): + logging.info(f"Decoding {test_set}") + if params.bilingual: + test_cut = test_cut.filter(remove_short_utt) + results_dict = decode_dataset( + cuts=test_cut, + 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() diff --git a/egs/multi_ja_en/ASR/zipformer/subsampling.py b/egs/multi_ja_en/ASR/zipformer/subsampling.py new file mode 120000 index 0000000000..01ae9002c6 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/test_scaling.py b/egs/multi_ja_en/ASR/zipformer/test_scaling.py new file mode 120000 index 0000000000..7157984369 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/test_scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/test_scaling.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/test_subsampling.py b/egs/multi_ja_en/ASR/zipformer/test_subsampling.py new file mode 120000 index 0000000000..bf0ee3d115 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/test_subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/test_subsampling.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/tokenizer.py b/egs/multi_ja_en/ASR/zipformer/tokenizer.py new file mode 120000 index 0000000000..958c99e855 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/tokenizer.py @@ -0,0 +1 @@ +../local/utils/tokenizer.py \ No newline at end of file diff --git a/egs/multi_ja_en/ASR/zipformer/train.py b/egs/multi_ja_en/ASR/zipformer/train.py new file mode 100755 index 0000000000..bfb037f503 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/train.py @@ -0,0 +1,1462 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey) +# +# 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" + +# For non-streaming model training: +./zipformer/train.py \ + --bilingual 1 \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --max-duration 600 + +# For streaming model training: +./zipformer/train.py \ + --bilingual 1 \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --causal 1 \ + --max-duration 600 + +It supports training with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + +import argparse +import copy +import logging +import os +import re +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import ReazonSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from multi_dataset import MultiDataset +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from tokenizer import Tokenizer +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import byte_encode, diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, + tokenize_by_ja_char, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--bilingual", + type=str2bool, + default=False, + help="Whether the model is bilingual or not. 1 = bilingual.", + ) + + 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=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + # changed - not used in monolingual streaming + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bbpe_2000/bbpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.015, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + 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( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + 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. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + 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 zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` 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. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + 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] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = 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. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + 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) + + +# fix implementation for sentencepiece_processor: spm.SentencePieceProcessor, stuff +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + tokenizer: Tokenizer, + sentencepiece_processor: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute 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 Zipformer 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. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).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) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + if not params.bilingual: + y = tokenizer.encode(texts, out_type=int) + else: + y = sentencepiece_processor.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + losses = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + simple_loss, pruned_loss, ctc_loss = losses[:3] + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + tokenizer: Tokenizer, + sentencepiece_processor: 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, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + 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: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + tokenizer: Tokenizer, + sentencepiece_processor: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> 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. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + 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. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch( + batch, + params=params, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + ) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise_grad_scale_is_too_small_error(cur_grad_scale) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + 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 params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + 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}") + + # Use lang_dir for further operations + # tokenizer = Tokenizer.load(args.lang, args.lang_type) + + # sentencepiece_processor = spm.SentencePieceProcessor() + # sentencepiece_processor.load(params.bpe_model) + tokenizer = None + sentencepiece_processor = None + + # is defined in local/prepare_lang_char.py + + if not params.bilingual: + tokenizer = Tokenizer.load(args.lang, args.lang_type) + params.blank_id = tokenizer.piece_to_id("") + params.vocab_size = tokenizer.get_piece_size() + else: + sentencepiece_processor = spm.SentencePieceProcessor() + sentencepiece_processor.load(params.bpe_model) + params.blank_id = sentencepiece_processor.piece_to_id("") + params.vocab_size = sentencepiece_processor.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + reazonspeech_corpus = ReazonSpeechAsrDataModule(args) + if params.bilingual: + multi_dataset = MultiDataset(args) + train_cuts = multi_dataset.train_cuts() + else: + train_cuts = reazonspeech_corpus.train_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + # if c.duration < 1.0 or c.duration > 30.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + # return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_samples - 7) // 2 + 1) // 2 + if not params.bilingual: + tokens = tokenizer.encode(c.supervisions[0].text, out_type=str) + else: + tokens = sentencepiece_processor.encode( + c.supervisions[0].text, out_type=str + ) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_samples}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + def tokenize_and_encode_text(c: Cut): + # Text normalize for each sample + text = c.supervisions[0].text + text = byte_encode(tokenize_by_ja_char(text)) + c.supervisions[0].text = text + return c + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.bilingual: + train_cuts = train_cuts.map(tokenize_and_encode_text) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = reazonspeech_corpus.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + if params.bilingual: + valid_cuts = reazonspeech_corpus.valid_cuts() + else: + valid_cuts = multi_dataset.dev_cuts() + valid_dl = reazonspeech_corpus.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + tokenizer: Tokenizer, + sentencepiece_processor: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + tokenizer: + The BPE Tokenizer model. + sentencepiece_processor: + The BPE SentencePieceProcessor model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + if params.bilingual: + y = sentencepiece_processor.encode(supervisions["text"], out_type=int) + else: + y = tokenizer.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + tokenizer: Tokenizer, + sentencepiece_processor: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception 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]}) ..." + ) + display_and_save_batch( + batch, + params=params, + tokenizer=tokenizer, + sentencepiece_processor=sentencepiece_processor, + ) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + ReazonSpeechAsrDataModule.add_arguments(parser) + Tokenizer.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() diff --git a/egs/multi_ja_en/ASR/zipformer/zipformer.py b/egs/multi_ja_en/ASR/zipformer/zipformer.py new file mode 120000 index 0000000000..23011dda71 --- /dev/null +++ b/egs/multi_ja_en/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/reazonspeech/ASR/RESULTS.md b/egs/reazonspeech/ASR/RESULTS.md index c0b4fe54a7..92610d75bb 100644 --- a/egs/reazonspeech/ASR/RESULTS.md +++ b/egs/reazonspeech/ASR/RESULTS.md @@ -47,3 +47,41 @@ The decoding command is: --blank-penalty 0 ``` +#### Streaming + +We have not completed evaluation of our models yet and will add evaluation results here once it's completed. + +The training command is: +```shell +./zipformer/train.py \ + --world-size 8 \ + --num-epochs 40 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp-large \ + --causal 1 \ + --num-encoder-layers 2,2,4,5,4,2 \ + --feedforward-dim 512,768,1536,2048,1536,768 \ + --encoder-dim 192,256,512,768,512,256 \ + --encoder-unmasked-dim 192,192,256,320,256,192 \ + --lang data/lang_char \ + --max-duration 1600 +``` + +The decoding command is: + +```shell +./zipformer/streaming_decode.py \ + --epoch 28 \ + --avg 15 \ + --causal 1 \ + --chunk-size 32 \ + --left-context-frames 256 \ + --exp-dir ./zipformer/exp-large \ + --lang data/lang_char \ + --num-encoder-layers 2,2,4,5,4,2 \ + --feedforward-dim 512,768,1536,2048,1536,768 \ + --encoder-dim 192,256,512,768,512,256 \ + --encoder-unmasked-dim 192,192,256,320,256,192 +``` + diff --git a/egs/reazonspeech/ASR/local/utils/tokenizer.py b/egs/reazonspeech/ASR/local/utils/tokenizer.py index c9be72be10..ba71cff893 100644 --- a/egs/reazonspeech/ASR/local/utils/tokenizer.py +++ b/egs/reazonspeech/ASR/local/utils/tokenizer.py @@ -12,7 +12,6 @@ class Tokenizer: @staticmethod def add_arguments(parser: argparse.ArgumentParser): group = parser.add_argument_group(title="Lang related options") - group.add_argument("--lang", type=Path, help="Path to lang directory.") group.add_argument( diff --git a/egs/reazonspeech/ASR/zipformer/streaming_decode.py b/egs/reazonspeech/ASR/zipformer/streaming_decode.py index 4c18c75634..7e3199e099 100755 --- a/egs/reazonspeech/ASR/zipformer/streaming_decode.py +++ b/egs/reazonspeech/ASR/zipformer/streaming_decode.py @@ -1,6 +1,7 @@ #!/usr/bin/env python3 -# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang) -# +# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang, +# Fangjun Kuang, +# Zengwei Yao) # See ../../../../LICENSE for clarification regarding multiple authors # # Licensed under the Apache License, Version 2.0 (the "License"); @@ -17,28 +18,23 @@ """ Usage: -./pruned_transducer_stateless7_streaming/streaming_decode.py \ - --epoch 28 \ - --avg 15 \ - --decode-chunk-len 32 \ - --exp-dir ./pruned_transducer_stateless7_streaming/exp \ - --decoding_method greedy_search \ - --lang data/lang_char \ - --num-decode-streams 2000 +./zipformer/streaming_decode.py--epoch 28 --avg 15 --causal 1 --chunk-size 32 --left-context-frames 256 --exp-dir ./zipformer/exp-large --lang data/lang_char --num-encoder-layers 2,2,4,5,4,2 --feedforward-dim 512,768,1536,2048,1536,768 --encoder-dim 192,256,512,768,512,256 --encoder-unmasked-dim 192,192,256,320,256,192 + """ import argparse import logging import math +import os +import pdb +import subprocess as sp from pathlib import Path from typing import Dict, List, Optional, Tuple import k2 import numpy as np import torch -import torch.nn as nn from asr_datamodule import ReazonSpeechAsrDataModule -from decode import save_results from decode_stream import DecodeStream from kaldifeat import Fbank, FbankOptions from lhotse import CutSet @@ -48,9 +44,9 @@ modified_beam_search, ) from tokenizer import Tokenizer +from torch import Tensor, nn from torch.nn.utils.rnn import pad_sequence -from train import add_model_arguments, get_params, get_transducer_model -from zipformer import stack_states, unstack_states +from train import add_model_arguments, get_model, get_params from icefall.checkpoint import ( average_checkpoints, @@ -58,7 +54,14 @@ find_checkpoints, load_checkpoint, ) -from icefall.utils import AttributeDict, setup_logger, str2bool +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) LOG_EPS = math.log(1e-10) @@ -73,7 +76,7 @@ def get_parser(): type=int, default=28, help="""It specifies the checkpoint to use for decoding. - Note: Epoch counts from 0. + Note: Epoch counts from 1. You can specify --avg to use more checkpoints for model averaging.""", ) @@ -87,12 +90,6 @@ def get_parser(): """, ) - parser.add_argument( - "--gpu", - type=int, - default=0, - ) - parser.add_argument( "--avg", type=int, @@ -116,7 +113,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="pruned_transducer_stateless2/exp", + default="zipformer/exp", help="The experiment dir", ) @@ -127,6 +124,13 @@ def get_parser(): help="Path to the BPE model", ) + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_char", + help="The lang dir containing word table and LG graph", + ) + parser.add_argument( "--decoding-method", type=str, @@ -138,14 +142,6 @@ def get_parser(): """, ) - parser.add_argument( - "--decoding-graph", - type=str, - default="", - help="""Used only when --decoding-method is - fast_beam_search""", - ) - parser.add_argument( "--num_active_paths", type=int, @@ -157,7 +153,7 @@ def get_parser(): parser.add_argument( "--beam", type=float, - default=4.0, + 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. @@ -194,18 +190,235 @@ def get_parser(): help="The number of streams that can be decoded parallel.", ) - parser.add_argument( - "--res-dir", - type=Path, - default=None, - help="The path to save results.", - ) - add_model_arguments(parser) return parser +def get_init_states( + model: nn.Module, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), +) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = model.encoder.get_init_states(batch_size, device) + + embed_states = model.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device) + states.append(processed_lens) + + return states + + +def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]: + """Stack list of zipformer states that correspond to separate utterances + into a single emformer state, so that it can be used as an input for + zipformer when those utterances are formed into a batch. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. For element-n, + state_list[n] is a list of cached tensors of all encoder layers. For layer-i, + state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, + cached_val2, cached_conv1, cached_conv2). + state_list[n][-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + state_list[n][-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Note: + It is the inverse of :func:`unstack_states`. + """ + batch_size = len(state_list) + assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0]) + tot_num_layers = (len(state_list[0]) - 2) // 6 + + batch_states = [] + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key = torch.cat( + [state_list[i][layer_offset] for i in range(batch_size)], dim=1 + ) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn = torch.cat( + [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1 = torch.cat( + [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2 = torch.cat( + [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1 = torch.cat( + [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2 = torch.cat( + [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0 + ) + batch_states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + cached_embed_left_pad = torch.cat( + [state_list[i][-2] for i in range(batch_size)], dim=0 + ) + batch_states.append(cached_embed_left_pad) + + processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0) + batch_states.append(processed_lens) + + return batch_states + + +def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]: + """Unstack the zipformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Note: + It is the inverse of :func:`stack_states`. + + Args: + batch_states: A list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + state_list[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Returns: + state_list: A list of list. Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. + """ + assert (len(batch_states) - 2) % 6 == 0, len(batch_states) + tot_num_layers = (len(batch_states) - 2) // 6 + + processed_lens = batch_states[-1] + batch_size = processed_lens.shape[0] + + state_list = [[] for _ in range(batch_size)] + + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk( + chunks=batch_size, dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1_list = batch_states[layer_offset + 2].chunk( + chunks=batch_size, dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2_list = batch_states[layer_offset + 3].chunk( + chunks=batch_size, dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1_list = batch_states[layer_offset + 4].chunk( + chunks=batch_size, dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2_list = batch_states[layer_offset + 5].chunk( + chunks=batch_size, dim=0 + ) + for i in range(batch_size): + state_list[i] += [ + cached_key_list[i], + cached_nonlin_attn_list[i], + cached_val1_list[i], + cached_val2_list[i], + cached_conv1_list[i], + cached_conv2_list[i], + ] + + cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(cached_embed_left_pad_list[i]) + + processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(processed_lens_list[i]) + + return state_list + + +def streaming_forward( + features: Tensor, + feature_lens: Tensor, + model: nn.Module, + states: List[Tensor], + chunk_size: int, + left_context_len: int, +) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Returns encoder outputs, output lengths, and updated states. + """ + cached_embed_left_pad = states[-2] + (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward( + x=features, + x_lens=feature_lens, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == chunk_size, (x.size(1), chunk_size) + + src_key_padding_mask = make_pad_mask(x_lens) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + encoder_states = states[:-2] + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = model.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + return encoder_out, encoder_out_lens, new_states + + def decode_one_chunk( params: AttributeDict, model: nn.Module, @@ -224,27 +437,32 @@ def decode_one_chunk( Returns: Return a List containing which DecodeStreams are finished. """ - device = model.device + # pdb.set_trace() + # print(model) + # print(model.device) + # device = model.device + chunk_size = int(params.chunk_size) + left_context_len = int(params.left_context_frames) features = [] feature_lens = [] states = [] - processed_lens = [] + processed_lens = [] # Used in fast-beam-search for stream in decode_streams: - feat, feat_len = stream.get_feature_frames(params.decode_chunk_len) + feat, feat_len = stream.get_feature_frames(chunk_size * 2) features.append(feat) feature_lens.append(feat_len) states.append(stream.states) processed_lens.append(stream.done_frames) - feature_lens = torch.tensor(feature_lens, device=device) + feature_lens = torch.tensor(feature_lens, device=model.device) features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS) - # We subsample features with ((x_len - 7) // 2 + 1) // 2 and the max downsampling - # factor in encoders is 8. - # After feature embedding (x_len - 7) // 2, we have (23 - 7) // 2 = 8. - tail_length = 23 + # Make sure the length after encoder_embed is at least 1. + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + tail_length = chunk_size * 2 + 7 + 2 * 3 if features.size(1) < tail_length: pad_length = tail_length - features.size(1) feature_lens += pad_length @@ -256,12 +474,14 @@ def decode_one_chunk( ) states = stack_states(states) - processed_lens = torch.tensor(processed_lens, device=device) - encoder_out, encoder_out_lens, new_states = model.encoder.streaming_forward( - x=features, - x_lens=feature_lens, + encoder_out, encoder_out_lens, new_states = streaming_forward( + features=features, + feature_lens=feature_lens, + model=model, states=states, + chunk_size=chunk_size, + left_context_len=left_context_len, ) encoder_out = model.joiner.encoder_proj(encoder_out) @@ -269,6 +489,7 @@ def decode_one_chunk( if params.decoding_method == "greedy_search": greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams) elif params.decoding_method == "fast_beam_search": + processed_lens = torch.tensor(processed_lens, device=model.device) processed_lens = processed_lens + encoder_out_lens fast_beam_search_one_best( model=model, @@ -295,8 +516,9 @@ def decode_one_chunk( for i in range(len(decode_streams)): decode_streams[i].states = states[i] decode_streams[i].done_frames += encoder_out_lens[i] - if decode_streams[i].done: - finished_streams.append(i) + # if decode_streams[i].done: + # finished_streams.append(i) + finished_streams.append(i) return finished_streams @@ -305,7 +527,7 @@ def decode_dataset( cuts: CutSet, params: AttributeDict, model: nn.Module, - sp: Tokenizer, + tokenizer: Tokenizer, decoding_graph: Optional[k2.Fsa] = None, ) -> Dict[str, List[Tuple[List[str], List[str]]]]: """Decode dataset. @@ -317,7 +539,7 @@ def decode_dataset( It is returned by :func:`get_params`. model: The neural model. - sp: + tokenizer: The BPE model. decoding_graph: The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used @@ -338,14 +560,14 @@ def decode_dataset( opts.frame_opts.samp_freq = 16000 opts.mel_opts.num_bins = 80 - log_interval = 50 + log_interval = 100 decode_results = [] # Contain decode streams currently running. decode_streams = [] for num, cut in enumerate(cuts): # each utterance has a DecodeStream. - initial_states = model.encoder.get_init_state(device=device) + initial_states = get_init_states(model=model, batch_size=1, device=device) decode_stream = DecodeStream( params=params, cut_id=cut.id, @@ -361,15 +583,19 @@ def decode_dataset( assert audio.dtype == np.float32, audio.dtype # The trained model is using normalized samples - assert audio.max() <= 1, "Should be normalized to [-1, 1])" + # - this is to avoid sending [-32k,+32k] signal in... + # - some lhotse AudioTransform classes can make the signal + # be out of range [-1, 1], hence the tolerance 10 + assert ( + np.abs(audio).max() <= 10 + ), "Should be normalized to [-1, 1], 10 for tolerance..." samples = torch.from_numpy(audio).squeeze(0) fbank = Fbank(opts) feature = fbank(samples.to(device)) - decode_stream.set_features(feature, tail_pad_len=params.decode_chunk_len) - decode_stream.ground_truth = cut.supervisions[0].custom[params.transcript_mode] - + decode_stream.set_features(feature, tail_pad_len=30) + decode_stream.ground_truth = cut.supervisions[0].text decode_streams.append(decode_stream) while len(decode_streams) >= params.num_decode_streams: @@ -380,8 +606,8 @@ def decode_dataset( decode_results.append( ( decode_streams[i].id, - sp.text2word(decode_streams[i].ground_truth), - sp.text2word(sp.decode(decode_streams[i].decoding_result())), + decode_streams[i].ground_truth.split(), + tokenizer.decode(decode_streams[i].decoding_result()).split(), ) ) del decode_streams[i] @@ -391,18 +617,37 @@ def decode_dataset( # decode final chunks of last sequences while len(decode_streams): + # print("INSIDE LEN DECODE STREAMS") + # pdb.set_trace() + # print(model.device) + # test_device = model.device + # print("done") finished_streams = decode_one_chunk( params=params, model=model, decode_streams=decode_streams ) + # print('INSIDE FOR LOOP ') + # print(finished_streams) + + if not finished_streams: + print("No finished streams, breaking the loop") + break + for i in sorted(finished_streams, reverse=True): - decode_results.append( - ( - decode_streams[i].id, - sp.text2word(decode_streams[i].ground_truth), - sp.text2word(sp.decode(decode_streams[i].decoding_result())), + try: + decode_results.append( + ( + decode_streams[i].id, + decode_streams[i].ground_truth.split(), + tokenizer.decode(decode_streams[i].decoding_result()).split(), + ) ) - ) - del decode_streams[i] + del decode_streams[i] + except IndexError as e: + print(f"IndexError: {e}") + print(f"decode_streams length: {len(decode_streams)}") + print(f"finished_streams: {finished_streams}") + print(f"i: {i}") + continue if params.decoding_method == "greedy_search": key = "greedy_search" @@ -416,9 +661,54 @@ def decode_dataset( key = f"num_active_paths_{params.num_active_paths}" else: raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + torch.cuda.synchronize() return {key: decode_results} +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[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() @@ -430,16 +720,20 @@ def main(): params = get_params() params.update(vars(args)) - if not params.res_dir: - params.res_dir = params.exp_dir / "streaming" / params.decoding_method + params.res_dir = params.exp_dir / "streaming" / params.decoding_method if params.iter > 0: params.suffix = f"iter-{params.iter}-avg-{params.avg}" else: params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" - # for streaming - params.suffix += f"-streaming-chunk-size-{params.decode_chunk_len}" + assert params.causal, params.causal + assert "," not in params.chunk_size, "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" # for fast_beam_search if params.decoding_method == "fast_beam_search": @@ -455,21 +749,21 @@ def main(): device = torch.device("cpu") if torch.cuda.is_available(): - device = torch.device("cuda", params.gpu) + device = torch.device("cuda", 0) logging.info(f"Device: {device}") - sp = Tokenizer.load(params.lang, params.lang_type) + sp_token = Tokenizer.load(params.lang, params.lang_type) - # and is defined in local/prepare_lang_char.py - params.blank_id = sp.piece_to_id("") - params.unk_id = sp.piece_to_id("") - params.vocab_size = sp.get_piece_size() + # and is defined in local/train_bpe_model.py + params.blank_id = sp_token.piece_to_id("") + params.unk_id = sp_token.piece_to_id("") + params.vocab_size = sp_token.get_piece_size() logging.info(params) logging.info("About to create model") - model = get_transducer_model(params) + model = get_model(params) if not params.use_averaged_model: if params.iter > 0: @@ -553,42 +847,51 @@ def main(): model.device = device decoding_graph = None - if params.decoding_graph: - decoding_graph = k2.Fsa.from_dict( - torch.load(params.decoding_graph, map_location=device) - ) - elif params.decoding_method == "fast_beam_search": + if params.decoding_method == "fast_beam_search": decoding_graph = k2.trivial_graph(params.vocab_size - 1, 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 reazonspeech_corpus = ReazonSpeechAsrDataModule(args) - for subdir in ["valid"]: + valid_cuts = reazonspeech_corpus.valid_cuts() + test_cuts = reazonspeech_corpus.test_cuts() + + test_sets = ["valid", "test"] + test_cuts = [valid_cuts, test_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): results_dict = decode_dataset( - cuts=getattr(reazonspeech_corpus, f"{subdir}_cuts")(), + cuts=test_cut, params=params, model=model, - sp=sp, + tokenizer=sp_token, decoding_graph=decoding_graph, ) - tot_err = save_results( - params=params, test_set_name=subdir, results_dict=results_dict + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, ) - with ( - params.res_dir - / ( - f"{subdir}-{params.decode_chunk_len}" - f"_{params.avg}_{params.epoch}.cer" - ) - ).open("w") as fout: - if len(tot_err) == 1: - fout.write(f"{tot_err[0][1]}") - else: - fout.write("\n".join(f"{k}\t{v}") for k, v in tot_err) + # valid_cuts = reazonspeech_corpus.valid_cuts() + + # for valid_cut in valid_cuts: + # results_dict = decode_dataset( + # cuts=valid_cut, + # params=params, + # model=model, + # sp=sp, + # decoding_graph=decoding_graph, + # ) + # save_results( + # params=params, + # test_set_name="valid", + # results_dict=results_dict, + # ) logging.info("Done!") diff --git a/icefall/__init__.py b/icefall/__init__.py index b1e4313e9b..3077b8162c 100644 --- a/icefall/__init__.py +++ b/icefall/__init__.py @@ -68,6 +68,7 @@ str2bool, subsequent_chunk_mask, tokenize_by_CJK_char, + tokenize_by_ja_char, write_error_stats, ) diff --git a/icefall/utils.py b/icefall/utils.py index 41eebadd46..aab479e568 100644 --- a/icefall/utils.py +++ b/icefall/utils.py @@ -1758,6 +1758,30 @@ def tokenize_by_CJK_char(line: str) -> str: return " ".join([w.strip() for w in chars if w.strip()]) +def tokenize_by_ja_char(line: str) -> str: + """ + Tokenize a line of text with Japanese characters. + + Note: All non-Japanese characters will be upper case. + + Example: + input = "こんにちは世界は hello world の日本語" + output = "こ ん に ち は 世 界 は HELLO WORLD の 日 本 語" + + Args: + line: + The input text. + + Return: + A new string tokenized by Japanese characters. + """ + pattern = re.compile(r"([\u3040-\u309F\u30A0-\u30FF\u4E00-\u9FFF])") + chars = pattern.split(line.strip()) + return " ".join( + [w.strip().upper() if not pattern.match(w) else w for w in chars if w.strip()] + ) + + def display_and_save_batch( batch: dict, params: AttributeDict,