diff --git a/README.md b/README.md index 107bbaee0..be922c191 100644 --- a/README.md +++ b/README.md @@ -32,6 +32,7 @@ We provide the following recipes: - [WenetSpeech][wenetspeech] - [Alimeeting][alimeeting] - [Aishell4][aishell4] + - [TAL_CSASR][tal_csasr] ### yesno @@ -286,6 +287,21 @@ The best CER(%) results: We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1z3lkURVv9M7uTiIgf3Np9IntMHEknaks?usp=sharing) +### TAL_CSASR + +We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][TAL_CSASR_pruned_transducer_stateless5]. + +#### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss + +The best results for Chinese CER(%) and English WER(%) respectivly (zh: Chinese, en: English): +|decoding-method | dev | dev_zh | dev_en | test | test_zh | test_en | +|--|--|--|--|--|--|--| +|greedy_search| 7.30 | 6.48 | 19.19 |7.39| 6.66 | 19.13| +|modified_beam_search| 7.15 | 6.35 | 18.95 | 7.22| 6.50 | 18.70 | +|fast_beam_search| 7.18 | 6.39| 18.90 | 7.27| 6.55 | 18.77| + +We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1DmIx-NloI1CMU5GdZrlse7TRu4y3Dpf8?usp=sharing) + ## Deployment with C++ Once you have trained a model in icefall, you may want to deploy it with C++, @@ -315,6 +331,7 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad [WenetSpeech_pruned_transducer_stateless2]: egs/wenetspeech/ASR/pruned_transducer_stateless2 [Alimeeting_pruned_transducer_stateless2]: egs/alimeeting/ASR/pruned_transducer_stateless2 [Aishell4_pruned_transducer_stateless5]: egs/aishell4/ASR/pruned_transducer_stateless5 +[TAL_CSASR_pruned_transducer_stateless5]: egs/tal_csasr/ASR/pruned_transducer_stateless5 [yesno]: egs/yesno/ASR [librispeech]: egs/librispeech/ASR [aishell]: egs/aishell/ASR @@ -325,4 +342,5 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad [wenetspeech]: egs/wenetspeech/ASR [alimeeting]: egs/alimeeting/ASR [aishell4]: egs/aishell4/ASR +[tal_csasr]: egs/tal_csasr/ASR [k2]: https://github.com/k2-fsa/k2 diff --git a/egs/tal_csasr/ASR/README.md b/egs/tal_csasr/ASR/README.md new file mode 100644 index 000000000..a705a2f44 --- /dev/null +++ b/egs/tal_csasr/ASR/README.md @@ -0,0 +1,19 @@ + +# Introduction + +This recipe includes some different ASR models trained with TAL_CSASR. + +[./RESULTS.md](./RESULTS.md) contains the latest results. + +# Transducers + +There are various folders containing the name `transducer` in this folder. +The following table lists the differences among them. + +| | Encoder | Decoder | Comment | +|---------------------------------------|---------------------|--------------------|-----------------------------| +| `pruned_transducer_stateless5` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner| + +The decoder in `transducer_stateless` is modified from the paper +[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/). +We place an additional Conv1d layer right after the input embedding layer. diff --git a/egs/tal_csasr/ASR/RESULTS.md b/egs/tal_csasr/ASR/RESULTS.md new file mode 100644 index 000000000..ddff0ab61 --- /dev/null +++ b/egs/tal_csasr/ASR/RESULTS.md @@ -0,0 +1,87 @@ +## Results + +### TAL_CSASR Mix Chars and BPEs training results (Pruned Transducer Stateless5) + +#### 2022-06-22 + +Using the codes from this PR https://github.com/k2-fsa/icefall/pull/428. + +The WERs are + +|decoding-method | epoch(iter) | avg | dev | test | +|--|--|--|--|--| +|greedy_search | 30 | 24 | 7.49 | 7.58| +|modified_beam_search | 30 | 24 | 7.33 | 7.38| +|fast_beam_search | 30 | 24 | 7.32 | 7.42| +|greedy_search(use-averaged-model=True) | 30 | 24 | 7.30 | 7.39| +|modified_beam_search(use-averaged-model=True) | 30 | 24 | 7.15 | 7.22| +|fast_beam_search(use-averaged-model=True) | 30 | 24 | 7.18 | 7.27| +|greedy_search | 348000 | 30 | 7.46 | 7.54| +|modified_beam_search | 348000 | 30 | 7.24 | 7.36| +|fast_beam_search | 348000 | 30 | 7.25 | 7.39 | + +The results (CER(%) and WER(%)) for Chinese CER and English WER respectivly (zh: Chinese, en: English): +|decoding-method | epoch(iter) | avg | dev | dev_zh | dev_en | test | test_zh | test_en | +|--|--|--|--|--|--|--|--|--| +|greedy_search(use-averaged-model=True) | 30 | 24 | 7.30 | 6.48 | 19.19 |7.39| 6.66 | 19.13| +|modified_beam_search(use-averaged-model=True) | 30 | 24 | 7.15 | 6.35 | 18.95 | 7.22| 6.50 | 18.70 | +|fast_beam_search(use-averaged-model=True) | 30 | 24 | 7.18 | 6.39| 18.90 | 7.27| 6.55 | 18.77| + +The training command for reproducing is given below: + +``` +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5" + +./pruned_transducer_stateless5/train.py \ + --world-size 6 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless5/exp \ + --lang-dir data/lang_char \ + --max-duration 90 +``` + +The tensorboard training log can be found at +https://tensorboard.dev/experiment/KaACzXOVR0OM6cy0qbN5hw/#scalars + +The decoding command is: +``` +epoch=30 +avg=24 +use_average_model=True + +## greedy search +./pruned_transducer_stateless5/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir pruned_transducer_stateless5/exp \ + --lang-dir ./data/lang_char \ + --max-duration 800 \ + --use-averaged-model $use_average_model + +## modified beam search +./pruned_transducer_stateless5/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir pruned_transducer_stateless5/exp \ + --lang-dir ./data/lang_char \ + --max-duration 800 \ + --decoding-method modified_beam_search \ + --beam-size 4 \ + --use-averaged-model $use_average_model + +## fast beam search +./pruned_transducer_stateless5/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir ./pruned_transducer_stateless5/exp \ + --lang-dir ./data/lang_char \ + --max-duration 1500 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 \ + --use-averaged-model $use_average_model +``` + +A pre-trained model and decoding logs can be found at diff --git a/egs/tal_csasr/ASR/local/__init__.py b/egs/tal_csasr/ASR/local/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/tal_csasr/ASR/local/compute_fbank_musan.py b/egs/tal_csasr/ASR/local/compute_fbank_musan.py new file mode 120000 index 000000000..5833f2484 --- /dev/null +++ b/egs/tal_csasr/ASR/local/compute_fbank_musan.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/compute_fbank_musan.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/local/compute_fbank_tal_csasr.py b/egs/tal_csasr/ASR/local/compute_fbank_tal_csasr.py new file mode 100755 index 000000000..367e098f7 --- /dev/null +++ b/egs/tal_csasr/ASR/local/compute_fbank_tal_csasr.py @@ -0,0 +1,115 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the tal_csasr dataset. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def compute_fbank_tal_csasr(num_mel_bins: int = 80): + src_dir = Path("data/manifests/tal_csasr") + output_dir = Path("data/fbank") + num_jobs = min(15, os.cpu_count()) + + dataset_parts = ( + "train_set", + "dev_set", + "test_set", + ) + prefix = "tal_csasr" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + cuts_filename = f"{prefix}_cuts_{partition}.{suffix}" + if (output_dir / cuts_filename).is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + if "train" in partition: + cut_set = ( + cut_set + + cut_set.perturb_speed(0.9) + + cut_set.perturb_speed(1.1) + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / cuts_filename) + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + + return parser.parse_args() + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_tal_csasr(num_mel_bins=args.num_mel_bins) diff --git a/egs/tal_csasr/ASR/local/display_manifest_statistics.py b/egs/tal_csasr/ASR/local/display_manifest_statistics.py new file mode 100644 index 000000000..7521bb55b --- /dev/null +++ b/egs/tal_csasr/ASR/local/display_manifest_statistics.py @@ -0,0 +1,96 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang +# Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This file displays duration statistics of utterances in a manifest. +You can use the displayed value to choose minimum/maximum duration +to remove short and long utterances during the training. +See the function `remove_short_and_long_utt()` +in ../../../librispeech/ASR/transducer/train.py +for usage. +""" + + +from lhotse import load_manifest + + +def main(): + paths = [ + "./data/fbank/tal_csasr_cuts_train_set.jsonl.gz", + "./data/fbank/tal_csasr_cuts_dev_set.jsonl.gz", + "./data/fbank/tal_csasr_cuts_test_set.jsonl.gz", + ] + + for path in paths: + print(f"Displaying the statistics for {path}") + cuts = load_manifest(path) + cuts.describe() + + +if __name__ == "__main__": + main() + +""" +Displaying the statistics for ./data/fbank/tal_csasr_cuts_train_set.jsonl.gz +Cuts count: 1050000 +Total duration (hours): 1679.0 +Speech duration (hours): 1679.0 (100.0%) +*** +Duration statistics (seconds): +mean 5.8 +std 4.1 +min 0.3 +25% 2.8 +50% 4.4 +75% 7.3 +99% 18.0 +99.5% 18.8 +99.9% 20.8 +max 36.5 +Displaying the statistics for ./data/fbank/tal_csasr_cuts_dev_set.jsonl.gz +Cuts count: 5000 +Total duration (hours): 8.0 +Speech duration (hours): 8.0 (100.0%) +*** +Duration statistics (seconds): +mean 5.8 +std 4.0 +min 0.5 +25% 2.8 +50% 4.5 +75% 7.4 +99% 17.0 +99.5% 17.7 +99.9% 19.5 +max 21.5 +Displaying the statistics for ./data/fbank/tal_csasr_cuts_test_set.jsonl.gz +Cuts count: 15000 +Total duration (hours): 23.6 +Speech duration (hours): 23.6 (100.0%) +*** +Duration statistics (seconds): +mean 5.7 +std 4.0 +min 0.5 +25% 2.8 +50% 4.4 +75% 7.2 +99% 17.2 +99.5% 17.9 +99.9% 19.6 +max 32.3 +""" diff --git a/egs/tal_csasr/ASR/local/prepare_char.py b/egs/tal_csasr/ASR/local/prepare_char.py new file mode 100755 index 000000000..2c5b8b8b3 --- /dev/null +++ b/egs/tal_csasr/ASR/local/prepare_char.py @@ -0,0 +1,265 @@ +#!/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/text_with_bpe, + - 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 re +from pathlib import Path +from typing import Dict, List + +import k2 +import sentencepiece as spm +import torch +from prepare_lang import ( + Lexicon, + add_disambig_symbols, + add_self_loops, + write_lexicon, + write_mapping, +) + + +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] if i in token2id else token2id[""] 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 contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool: + """Check if all the given tokens are in token symbol table. + + Args: + token_sym_table: + Token symbol table that contains all the valid tokens. + tokens: + A list of tokens. + Returns: + Return True if there is any token not in the token_sym_table, + otherwise False. + """ + for tok in tokens: + if tok not in token_sym_table: + return True + return False + + +def generate_lexicon( + token_sym_table: Dict[str, int], + words: List[str], + bpe_model=None, +) -> Lexicon: + """Generate a lexicon from a word list and token_sym_table. + + Args: + token_sym_table: + Token symbol table that mapping token to token ids. + words: + A list of strings representing words. + Returns: + Return a dict whose keys are words and values are the corresponding + tokens. + """ + sp = "" + if bpe_model is not None: + sp = spm.SentencePieceProcessor() + sp.load(str(bpe_model)) + + lexicon = [] + zhPattern = re.compile(r"([\u4e00-\u9fa5])") + for word in words: + match = zhPattern.search(word) + tokens = [] + if match: + tokens = list(word.strip(" \t")) + else: + tokens = sp.encode_as_pieces(word.strip(" \t")) + + if contain_oov(token_sym_table, tokens): + continue + lexicon.append((word, tokens)) + + # The OOV word is + lexicon.append(("", [""])) + return lexicon + + +def generate_tokens(text_file: str) -> Dict[str, int]: + """Generate tokens from the given text file. + + Args: + text_file: + A file that contains text lines to generate tokens. + Returns: + Return a dict whose keys are tokens and values are token ids ranged + from 0 to len(keys) - 1. + """ + tokens: Dict[str, int] = dict() + tokens[""] = 0 + tokens[""] = 1 + tokens[""] = 2 + whitespace = re.compile(r"([\t\r\n]+)") + with open(text_file, "r", encoding="utf-8") as f: + for line in f: + line = re.sub(whitespace, "", line) + chars = line.split(" ") + for char in chars: + if char not in tokens: + tokens[char] = len(tokens) + + return tokens + + +def main(): + lang_dir = Path("data/lang_char") + text_file = lang_dir / "text_with_bpe" + bpe_model = lang_dir / "bpe.model" + + word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt") + + words = word_sym_table.symbols + + excluded = ["", "!SIL", "", "", "#0", "", ""] + for w in excluded: + if w in words: + words.remove(w) + + token_sym_table = generate_tokens(text_file) + + lexicon = generate_lexicon(token_sym_table, words, bpe_model=bpe_model) + + 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 __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/local/prepare_lang.py b/egs/tal_csasr/ASR/local/prepare_lang.py new file mode 100755 index 000000000..e5ae89ec4 --- /dev/null +++ b/egs/tal_csasr/ASR/local/prepare_lang.py @@ -0,0 +1,390 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This script takes as input a lexicon file "data/lang_phone/lexicon.txt" +consisting of words and tokens (i.e., phones) and does the following: + +1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt + +2. Generate tokens.txt, the token table mapping a token to a unique integer. + +3. Generate words.txt, the word table mapping a word to a unique integer. + +4. Generate L.pt, in k2 format. It can be loaded by + + d = torch.load("L.pt") + lexicon = k2.Fsa.from_dict(d) + +5. Generate L_disambig.pt, in k2 format. +""" +import argparse +import math +from collections import defaultdict +from pathlib import Path +from typing import Any, Dict, List, Tuple + +import k2 +import torch + +from icefall.lexicon import read_lexicon, write_lexicon + +Lexicon = List[Tuple[str, List[str]]] + + +def write_mapping(filename: str, sym2id: Dict[str, int]) -> None: + """Write a symbol to ID mapping to a file. + + Note: + No need to implement `read_mapping` as it can be done + through :func:`k2.SymbolTable.from_file`. + + Args: + filename: + Filename to save the mapping. + sym2id: + A dict mapping symbols to IDs. + Returns: + Return None. + """ + with open(filename, "w", encoding="utf-8") as f: + for sym, i in sym2id.items(): + f.write(f"{sym} {i}\n") + + +def get_tokens(lexicon: Lexicon) -> List[str]: + """Get tokens from a lexicon. + + Args: + lexicon: + It is the return value of :func:`read_lexicon`. + Returns: + Return a list of unique tokens. + """ + ans = set() + for _, tokens in lexicon: + ans.update(tokens) + sorted_ans = sorted(list(ans)) + return sorted_ans + + +def get_words(lexicon: Lexicon) -> List[str]: + """Get words from a lexicon. + + Args: + lexicon: + It is the return value of :func:`read_lexicon`. + Returns: + Return a list of unique words. + """ + ans = set() + for word, _ in lexicon: + ans.add(word) + sorted_ans = sorted(list(ans)) + return sorted_ans + + +def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]: + """It adds pseudo-token disambiguation symbols #1, #2 and so on + at the ends of tokens to ensure that all pronunciations are different, + and that none is a prefix of another. + + See also add_lex_disambig.pl from kaldi. + + Args: + lexicon: + It is returned by :func:`read_lexicon`. + Returns: + Return a tuple with two elements: + + - The output lexicon with disambiguation symbols + - The ID of the max disambiguation symbol that appears + in the lexicon + """ + + # (1) Work out the count of each token-sequence in the + # lexicon. + count = defaultdict(int) + for _, tokens in lexicon: + count[" ".join(tokens)] += 1 + + # (2) For each left sub-sequence of each token-sequence, note down + # that it exists (for identifying prefixes of longer strings). + issubseq = defaultdict(int) + for _, tokens in lexicon: + tokens = tokens.copy() + tokens.pop() + while tokens: + issubseq[" ".join(tokens)] = 1 + tokens.pop() + + # (3) For each entry in the lexicon: + # if the token sequence is unique and is not a + # prefix of another word, no disambig symbol. + # Else output #1, or #2, #3, ... if the same token-seq + # has already been assigned a disambig symbol. + ans = [] + + # We start with #1 since #0 has its own purpose + first_allowed_disambig = 1 + max_disambig = first_allowed_disambig - 1 + last_used_disambig_symbol_of = defaultdict(int) + + for word, tokens in lexicon: + tokenseq = " ".join(tokens) + assert tokenseq != "" + if issubseq[tokenseq] == 0 and count[tokenseq] == 1: + ans.append((word, tokens)) + continue + + cur_disambig = last_used_disambig_symbol_of[tokenseq] + if cur_disambig == 0: + cur_disambig = first_allowed_disambig + else: + cur_disambig += 1 + + if cur_disambig > max_disambig: + max_disambig = cur_disambig + last_used_disambig_symbol_of[tokenseq] = cur_disambig + tokenseq += f" #{cur_disambig}" + ans.append((word, tokenseq.split())) + return ans, max_disambig + + +def generate_id_map(symbols: List[str]) -> Dict[str, int]: + """Generate ID maps, i.e., map a symbol to a unique ID. + + Args: + symbols: + A list of unique symbols. + Returns: + A dict containing the mapping between symbols and IDs. + """ + return {sym: i for i, sym in enumerate(symbols)} + + +def add_self_loops( + arcs: List[List[Any]], disambig_token: int, disambig_word: int +) -> List[List[Any]]: + """Adds self-loops to states of an FST to propagate disambiguation symbols + through it. They are added on each state with non-epsilon output symbols + on at least one arc out of the state. + + See also fstaddselfloops.pl from Kaldi. One difference is that + Kaldi uses OpenFst style FSTs and it has multiple final states. + This function uses k2 style FSTs and it does not need to add self-loops + to the final state. + + The input label of a self-loop is `disambig_token`, while the output + label is `disambig_word`. + + Args: + arcs: + A list-of-list. The sublist contains + `[src_state, dest_state, label, aux_label, score]` + disambig_token: + It is the token ID of the symbol `#0`. + disambig_word: + It is the word ID of the symbol `#0`. + + Return: + Return new `arcs` containing self-loops. + """ + states_needs_self_loops = set() + for arc in arcs: + src, dst, ilabel, olabel, score = arc + if olabel != 0: + states_needs_self_loops.add(src) + + ans = [] + for s in states_needs_self_loops: + ans.append([s, s, disambig_token, disambig_word, 0]) + + return arcs + ans + + +def lexicon_to_fst( + lexicon: Lexicon, + token2id: Dict[str, int], + word2id: Dict[str, int], + sil_token: str = "SIL", + sil_prob: float = 0.5, + need_self_loops: bool = False, +) -> k2.Fsa: + """Convert a lexicon to an FST (in k2 format) with optional silence at + the beginning and end of each word. + + Args: + lexicon: + The input lexicon. See also :func:`read_lexicon` + token2id: + A dict mapping tokens to IDs. + word2id: + A dict mapping words to IDs. + sil_token: + The silence token. + sil_prob: + The probability for adding a silence at the beginning and end + of the word. + need_self_loops: + If True, add self-loop to states with non-epsilon output symbols + on at least one arc out of the state. The input label for this + self loop is `token2id["#0"]` and the output label is `word2id["#0"]`. + Returns: + Return an instance of `k2.Fsa` representing the given lexicon. + """ + assert sil_prob > 0.0 and sil_prob < 1.0 + # CAUTION: we use score, i.e, negative cost. + sil_score = math.log(sil_prob) + no_sil_score = math.log(1.0 - sil_prob) + + start_state = 0 + loop_state = 1 # words enter and leave from here + sil_state = 2 # words terminate here when followed by silence; this state + # has a silence transition to loop_state. + next_state = 3 # the next un-allocated state, will be incremented as we go. + arcs = [] + + assert token2id[""] == 0 + assert word2id[""] == 0 + + eps = 0 + + sil_token = token2id[sil_token] + + arcs.append([start_state, loop_state, eps, eps, no_sil_score]) + arcs.append([start_state, sil_state, eps, eps, sil_score]) + arcs.append([sil_state, loop_state, sil_token, eps, 0]) + + for word, tokens in lexicon: + assert len(tokens) > 0, f"{word} has no pronunciations" + cur_state = loop_state + + word = word2id[word] + tokens = [token2id[i] for i in tokens] + + for i in range(len(tokens) - 1): + w = word if i == 0 else eps + arcs.append([cur_state, next_state, tokens[i], w, 0]) + + cur_state = next_state + next_state += 1 + + # now for the last token of this word + # It has two out-going arcs, one to the loop state, + # the other one to the sil_state. + i = len(tokens) - 1 + w = word if i == 0 else eps + arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score]) + arcs.append([cur_state, sil_state, tokens[i], w, sil_score]) + + if need_self_loops: + disambig_token = token2id["#0"] + disambig_word = word2id["#0"] + arcs = add_self_loops( + arcs, + disambig_token=disambig_token, + disambig_word=disambig_word, + ) + + final_state = next_state + arcs.append([loop_state, final_state, -1, -1, 0]) + arcs.append([final_state]) + + arcs = sorted(arcs, key=lambda arc: arc[0]) + arcs = [[str(i) for i in arc] for arc in arcs] + arcs = [" ".join(arc) for arc in arcs] + arcs = "\n".join(arcs) + + fsa = k2.Fsa.from_str(arcs, acceptor=False) + return fsa + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", type=str, help="The lang dir, data/lang_phone" + ) + return parser.parse_args() + + +def main(): + out_dir = Path(get_args().lang_dir) + lexicon_filename = out_dir / "lexicon.txt" + sil_token = "SIL" + sil_prob = 0.5 + + lexicon = read_lexicon(lexicon_filename) + tokens = get_tokens(lexicon) + words = get_words(lexicon) + + lexicon_disambig, max_disambig = add_disambig_symbols(lexicon) + + for i in range(max_disambig + 1): + disambig = f"#{i}" + assert disambig not in tokens + tokens.append(f"#{i}") + + assert "" not in tokens + tokens = [""] + tokens + + assert "" not in words + assert "#0" not in words + assert "" not in words + assert "" not in words + + words = [""] + words + ["#0", "", ""] + + token2id = generate_id_map(tokens) + word2id = generate_id_map(words) + + write_mapping(out_dir / "tokens.txt", token2id) + write_mapping(out_dir / "words.txt", word2id) + write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig) + + L = lexicon_to_fst( + lexicon, + token2id=token2id, + word2id=word2id, + sil_token=sil_token, + sil_prob=sil_prob, + ) + + L_disambig = lexicon_to_fst( + lexicon_disambig, + token2id=token2id, + word2id=word2id, + sil_token=sil_token, + sil_prob=sil_prob, + need_self_loops=True, + ) + torch.save(L.as_dict(), out_dir / "L.pt") + torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt") + + if False: + # Just for debugging, will remove it + L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt") + L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt") + L_disambig.labels_sym = L.labels_sym + L_disambig.aux_labels_sym = L.aux_labels_sym + L.draw(out_dir / "L.png", title="L") + L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig") + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/local/prepare_words.py b/egs/tal_csasr/ASR/local/prepare_words.py new file mode 100755 index 000000000..41ab3b2cb --- /dev/null +++ b/egs/tal_csasr/ASR/local/prepare_words.py @@ -0,0 +1,84 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- + +# Copyright 2021 Xiaomi Corp. (authors: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This script takes as input words.txt without ids: + - words_no_ids.txt +and generates the new words.txt with related ids. + - words.txt +""" + + +import argparse +import logging + +from tqdm import tqdm + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Prepare words.txt", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--input", + default="data/lang_char/words_no_ids.txt", + type=str, + help="the words file without ids for WenetSpeech", + ) + parser.add_argument( + "--output", + default="data/lang_char/words.txt", + type=str, + help="the words file with ids for WenetSpeech", + ) + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + input_file = args.input + output_file = args.output + + f = open(input_file, "r", encoding="utf-8") + lines = f.readlines() + new_lines = [] + add_words = [" 0", "!SIL 1", " 2", " 3"] + new_lines.extend(add_words) + + logging.info("Starting reading the input file") + for i in tqdm(range(len(lines))): + x = lines[i] + idx = 4 + i + new_line = str(x.strip("\n")) + " " + str(idx) + new_lines.append(new_line) + + logging.info("Starting writing the words.txt") + f_out = open(output_file, "w", encoding="utf-8") + for line in new_lines: + f_out.write(line) + f_out.write("\n") + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/local/test_prepare_lang.py b/egs/tal_csasr/ASR/local/test_prepare_lang.py new file mode 100755 index 000000000..d4cf62bba --- /dev/null +++ b/egs/tal_csasr/ASR/local/test_prepare_lang.py @@ -0,0 +1,106 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +# Copyright (c) 2021 Xiaomi Corporation (authors: Fangjun Kuang) + +import os +import tempfile + +import k2 +from prepare_lang import ( + add_disambig_symbols, + generate_id_map, + get_phones, + get_words, + lexicon_to_fst, + read_lexicon, + write_lexicon, + write_mapping, +) + + +def generate_lexicon_file() -> str: + fd, filename = tempfile.mkstemp() + os.close(fd) + s = """ + !SIL SIL + SPN + SPN + f f + a a + foo f o o + bar b a r + bark b a r k + food f o o d + food2 f o o d + fo f o + """.strip() + with open(filename, "w") as f: + f.write(s) + return filename + + +def test_read_lexicon(filename: str): + lexicon = read_lexicon(filename) + phones = get_phones(lexicon) + words = get_words(lexicon) + print(lexicon) + print(phones) + print(words) + lexicon_disambig, max_disambig = add_disambig_symbols(lexicon) + print(lexicon_disambig) + print("max disambig:", f"#{max_disambig}") + + phones = ["", "SIL", "SPN"] + phones + for i in range(max_disambig + 1): + phones.append(f"#{i}") + words = [""] + words + + phone2id = generate_id_map(phones) + word2id = generate_id_map(words) + + print(phone2id) + print(word2id) + + write_mapping("phones.txt", phone2id) + write_mapping("words.txt", word2id) + + write_lexicon("a.txt", lexicon) + write_lexicon("a_disambig.txt", lexicon_disambig) + + fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id) + fsa.labels_sym = k2.SymbolTable.from_file("phones.txt") + fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt") + fsa.draw("L.pdf", title="L") + + fsa_disambig = lexicon_to_fst( + lexicon_disambig, phone2id=phone2id, word2id=word2id + ) + fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt") + fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt") + fsa_disambig.draw("L_disambig.pdf", title="L_disambig") + + +def main(): + filename = generate_lexicon_file() + test_read_lexicon(filename) + os.remove(filename) + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/local/text2segments.py b/egs/tal_csasr/ASR/local/text2segments.py new file mode 100644 index 000000000..3df727c67 --- /dev/null +++ b/egs/tal_csasr/ASR/local/text2segments.py @@ -0,0 +1,83 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- + +# Copyright 2021 Xiaomi Corp. (authors: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This script takes as input "text", which refers to the transcript file for +WenetSpeech: + - text +and generates the output file text_word_segmentation which is implemented +with word segmenting: + - text_words_segmentation +""" + + +import argparse + +import jieba +from tqdm import tqdm + +jieba.enable_paddle() + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Chinese Word Segmentation for text", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--input-file", + default="data/lang_char/text", + type=str, + help="the input text file for WenetSpeech", + ) + parser.add_argument( + "--output-file", + default="data/lang_char/text_words_segmentation", + type=str, + help="the text implemented with words segmenting for WenetSpeech", + ) + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + input_file = args.input_file + output_file = args.output_file + + f = open(input_file, "r", encoding="utf-8") + lines = f.readlines() + new_lines = [] + for i in tqdm(range(len(lines))): + x = lines[i].rstrip() + seg_list = jieba.cut(x, use_paddle=True) + new_line = " ".join(seg_list) + new_lines.append(new_line) + + f_new = open(output_file, "w", encoding="utf-8") + for line in new_lines: + f_new.write(line) + f_new.write("\n") + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/local/text2token.py b/egs/tal_csasr/ASR/local/text2token.py new file mode 100755 index 000000000..71be2a613 --- /dev/null +++ b/egs/tal_csasr/ASR/local/text2token.py @@ -0,0 +1,195 @@ +#!/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 pypinyin import lazy_pinyin, pinyin + +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 symobles, 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", "pinyin", "lazy_pinyin"], + help="""Transcript type. char/pinyin/lazy_pinyin""", + ) + return parser + + +def token2id( + texts, token_table, token_type: str = "lazy_pinyin", oov: str = "" +) -> List[List[int]]: + """Convert token to id. + Args: + texts: + The input texts, it refers to the chinese text here. + token_table: + The token table is built based on "data/lang_xxx/token.txt" + token_type: + The type of token, such as "pinyin" and "lazy_pinyin". + 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 == "lazy_pinyin": + text = lazy_pinyin(chars_list) + sub_ids = [ + token_table[txt] if txt in token_table else oov_id + for txt in text + ] + ids.append(sub_ids) + else: # token_type = "pinyin" + text = pinyin(chars_list) + sub_ids = [ + token_table[txt[0]] if txt[0] 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 == "pinyin": + a = pinyin(list(str(a))) + a = [one[0] for one in a] + + if args.trans_type == "lazy_pinyin": + a = lazy_pinyin(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/tal_csasr/ASR/local/text_normalize.py b/egs/tal_csasr/ASR/local/text_normalize.py new file mode 100755 index 000000000..e97b3a5a3 --- /dev/null +++ b/egs/tal_csasr/ASR/local/text_normalize.py @@ -0,0 +1,147 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +# Copyright 2022 Xiaomi Corp. (authors: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This script takes as input "text_full", which includes all transcript files +for a dataset: + - text_full +and generates the output file text_normalize which is implemented +to normalize text: + - text +""" + + +import argparse + +from tqdm import tqdm + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Normalizing for text", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--input", + default="data/lang_char/text_full", + type=str, + help="the input text file", + ) + parser.add_argument( + "--output", + default="data/lang_char/text", + type=str, + help="the text implemented with normalizer", + ) + + return parser + + +def text_normalize(str_line: str): + line = str_line.strip().rstrip("\n") + line = line.replace("", "") + line = line.replace("<%>", "") + line = line.replace("<->", "") + line = line.replace("<$>", "") + line = line.replace("<#>", "") + line = line.replace("<_>", "") + line = line.replace("", "") + line = line.replace("`", "") + line = line.replace("'", "") + line = line.replace("&", "") + line = line.replace(",", "") + line = line.replace("Ａ", "A") + line = line.replace("Ｃ", "C") + line = line.replace("Ｄ", "D") + line = line.replace("Ｅ", "E") + line = line.replace("Ｇ", "G") + line = line.replace("Ｈ", "H") + line = line.replace("Ｉ", "I") + line = line.replace("Ｎ", "N") + line = line.replace("Ｕ", "U") + line = line.replace("Ｗ", "W") + line = line.replace("Ｙ", "Y") + line = line.replace("ａ", "A") + line = line.replace("ｂ", "B") + line = line.replace("ｃ", "C") + line = line.replace("ｋ", "K") + line = line.replace("ｔ", "T") + line = line.replace("，", "") + line = line.replace("丶", "") + line = line.replace("。", "") + line = line.replace("、", "") + line = line.replace("？", "") + line = line.replace("·", "") + line = line.replace("*", "") + line = line.replace("！", "") + line = line.replace("$", "") + line = line.replace("+", "") + line = line.replace("-", "") + line = line.replace("\\", "") + line = line.replace("?", "") + line = line.replace("￥", "") + line = line.replace("%", "") + line = line.replace(".", "") + line = line.replace("<", "") + line = line.replace("＆", "") + line = line.replace("~", "") + line = line.replace("=", "") + line = line.replace(":", "") + line = line.replace("!", "") + line = line.replace("/", "") + line = line.replace("‘", "") + line = line.replace("’", "") + line = line.replace("“", "") + line = line.replace("”", "") + line = line.replace("[", "") + line = line.replace("]", "") + line = line.replace("@", "") + line = line.replace("#", "") + line = line.replace("：", "") + line = line.replace("；", "") + line = line.replace("…", "") + line = line.replace("《", "") + line = line.replace("》", "") + line = line.upper() + + return line + + +def main(): + parser = get_parser() + args = parser.parse_args() + + input_file = args.input + output_file = args.output + + f = open(input_file, "r", encoding="utf-8") + lines = f.readlines() + new_lines = [] + for i in tqdm(range(len(lines))): + new_line = text_normalize(lines[i]) + new_lines.append(new_line) + + f_new = open(output_file, "w", encoding="utf-8") + for line in new_lines: + f_new.write(line) + f_new.write("\n") + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/local/tokenize_with_bpe_model.py b/egs/tal_csasr/ASR/local/tokenize_with_bpe_model.py new file mode 100644 index 000000000..d7fd838f2 --- /dev/null +++ b/egs/tal_csasr/ASR/local/tokenize_with_bpe_model.py @@ -0,0 +1,95 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- +# +# Copyright 2021 Mobvoi Inc. (authors: Binbin Zhang) +# Copyright 2022 Xiaomi Corp. (authors: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This script takes as input text (it includes Chinese and English): + - text +and generates the text_with_bpe. + - text_with_bpe +""" + + +import argparse +import logging + +import sentencepiece as spm +from tqdm import tqdm + +from icefall.utils import tokenize_by_bpe_model + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Prepare text_with_bpe", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--input", + default="data/lang_char/text", + type=str, + help="the text includes Chinese and English words", + ) + parser.add_argument( + "--output", + default="data/lang_char/text_with_bpe", + type=str, + help="the text_with_bpe tokenized by bpe model", + ) + parser.add_argument( + "--bpe-model", + default="data/lang_char/bpe.model", + type=str, + help="the bpe model for processing the English parts", + ) + + return parser + + +def main(): + parser = get_parser() + args = parser.parse_args() + + input_file = args.input + output_file = args.output + bpe_model = args.bpe_model + + sp = spm.SentencePieceProcessor() + sp.load(bpe_model) + + f = open(input_file, "r", encoding="utf-8") + lines = f.readlines() + + logging.info("Starting reading the text") + new_lines = [] + for i in tqdm(range(len(lines))): + x = lines[i] + txt_tokens = tokenize_by_bpe_model(sp, x) + new_line = txt_tokens.replace("/", " ") + new_lines.append(new_line) + + logging.info("Starting writing the text_with_bpe") + f_out = open(output_file, "w", encoding="utf-8") + for line in new_lines: + f_out.write(line) + f_out.write("\n") + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/prepare.sh b/egs/tal_csasr/ASR/prepare.sh new file mode 100755 index 000000000..340521ad8 --- /dev/null +++ b/egs/tal_csasr/ASR/prepare.sh @@ -0,0 +1,172 @@ +#!/usr/bin/env bash + +set -eou pipefail + +stage=-1 +stop_stage=100 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/tal_csasr +# You can find three directories:train_set, dev_set, and test_set. +# You can get it from https://ai.100tal.com/dataset +# +# - $dl_dir/musan +# This directory contains the following directories downloaded from +# http://www.openslr.org/17/ +# +# - music +# - noise +# - speech + +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "Stage 0: Download data" + # Before you run this script, you must get the TAL_CSASR dataset + # from https://ai.100tal.com/dataset + mv $dl_dir/TALCS_corpus $dl_dir/tal_csasr + + # If you have pre-downloaded it to /path/to/TALCS_corpus, + # you can create a symlink + # + # ln -sfv /path/to/TALCS_corpus $dl_dir/tal_csasr + + # If you have pre-downloaded it to /path/to/musan, + # you can create a symlink + # + # ln -sfv /path/to/musan $dl_dir/musan + # + if [ ! -d $dl_dir/musan ]; then + lhotse download musan $dl_dir + fi +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare tal_csasr manifest" + # We assume that you have downloaded the TALCS_corpus + # to $dl_dir/tal_csasr + if [ ! -f data/manifests/tal_csasr/.manifests.done ]; then + mkdir -p data/manifests/tal_csasr + lhotse prepare tal-csasr $dl_dir/tal_csasr data/manifests/tal_csasr + touch data/manifests/tal_csasr/.manifests.done + fi +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Prepare musan manifest" + # We assume that you have downloaded the musan corpus + # to data/musan + if [ ! -f data/manifests/.musan_manifests.done ]; then + log "It may take 6 minutes" + mkdir -p data/manifests + lhotse prepare musan $dl_dir/musan data/manifests + touch data/manifests/.musan_manifests.done + fi +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Compute fbank for musan" + if [ ! -f data/fbank/.msuan.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_musan.py + touch data/fbank/.msuan.done + fi +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Compute fbank for tal_csasr" + if [ ! -f data/fbank/.tal_csasr.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_tal_csasr.py + touch data/fbank/.tal_csasr.done + fi +fi + +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Prepare char based lang" + lang_char_dir=data/lang_char + mkdir -p $lang_char_dir + + # Download BPE models trained with LibriSpeech + # Here we use the BPE model with 5000 units trained with Librispeech. + # You can also use other BPE models if available. + if [ ! -f $lang_char_dir/bpe.model ]; then + wget -O $lang_char_dir/bpe.model \ + https://huggingface.co/luomingshuang/bpe_models_trained_with_Librispeech/resolve/main/lang_bpe_5000/bpe.model + fi + + # Prepare text. + # Note: in Linux, you can install jq with the following command: + # 1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64 + # 2. chmod +x ./jq + # 3. cp jq /usr/bin + if [ ! -f $lang_char_dir/text_full ]; then + gunzip -c data/manifests/tal_csasr/tal_csasr_supervisions_train_set.jsonl.gz \ + | jq ".text" | sed 's/"//g' \ + | ./local/text2token.py -t "char" > $lang_char_dir/text_train + + gunzip -c data/manifests/tal_csasr/tal_csasr_supervisions_dev_set.jsonl.gz \ + | jq ".text" | sed 's/"//g' \ + | ./local/text2token.py -t "char" > $lang_char_dir/text_dev + + gunzip -c data/manifests/tal_csasr/tal_csasr_supervisions_test_set.jsonl.gz \ + | jq ".text" | sed 's/"//g' \ + | ./local/text2token.py -t "char" > $lang_char_dir/text_test + + for r in text_train text_dev text_test ; do + cat $lang_char_dir/$r >> $lang_char_dir/text_full + done + fi + + # Prepare text normalize + if [ ! -f $lang_char_dir/text ]; then + python ./local/text_normalize.py \ + --input $lang_char_dir/text_full \ + --output $lang_char_dir/text + fi + + # Prepare words segments + if [ ! -f $lang_char_dir/text_words_segmentation ]; then + python ./local/text2segments.py \ + --input $lang_char_dir/text \ + --output $lang_char_dir/text_words_segmentation + + cat $lang_char_dir/text_words_segmentation | sed "s/ /\n/g" \ + | sort -u | sed "/^$/d" \ + | uniq > $lang_char_dir/words_no_ids.txt + fi + + # Prepare words.txt + if [ ! -f $lang_char_dir/words.txt ]; then + ./local/prepare_words.py \ + --input $lang_char_dir/words_no_ids.txt \ + --output $lang_char_dir/words.txt + fi + + # Tokenize text with BPE model + python ./local/tokenize_with_bpe_model.py \ + --input $lang_char_dir/text \ + --output $lang_char_dir/text_with_bpe \ + --bpe-model $lang_char_dir/bpe.model + + if [ ! -f $lang_char_dir/L_disambig.pt ]; then + python local/prepare_char.py + fi +fi diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/__init__.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py new file mode 100644 index 000000000..49bfb148b --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py @@ -0,0 +1,434 @@ +# Copyright 2021 Piotr Żelasko +# +# 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 + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SingleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import ( # noqa F401 for AudioSamples + AudioSamples, + OnTheFlyFeatures, +) +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class TAL_CSASRAsrDataModule: + """ + 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/valid/test cuts.", + ) + + group.add_argument( + "--max-duration", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + + group.add_argument( + "--num-buckets", + type=int, + default=300, + 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=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + group.add_argument( + "--input-strategy", + type=str, + default="PrecomputedFeatures", + help="AudioSamples or PrecomputedFeatures", + ) + + 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. + """ + logging.info("About to get Musan cuts") + cuts_musan = load_manifest( + self.args.manifest_dir / "musan_cuts.jsonl.gz" + ) + + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + transforms.append( + CutMix( + cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True + ) + ) + else: + logging.info("Disable MUSAN") + + if self.args.concatenate_cuts: + logging.info( + f"Using cut concatenation with duration factor " + f"{self.args.duration_factor} and gap {self.args.gap}." + ) + # Cut concatenation should be the first transform in the list, + # so that if we e.g. mix noise in, it will fill the gaps between + # different utterances. + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + 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( + input_strategy=eval(self.args.input_strategy)(), + 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, + num_cuts_for_bins_estimate=20000, + buffer_size=60000, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SingleCutSampler.") + train_sampler = SingleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_dl.sampler.load_state_dict(sampler_state_dict) + + 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, + rank=0, + world_size=1, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + rank=0, + world_size=1, + shuffle=False, + ) + logging.info("About to create test dataloader") + 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 / "tal_csasr_cuts_train_set.jsonl.gz" + ) + + @lru_cache() + def valid_cuts(self) -> CutSet: + logging.info("About to get dev cuts") + return load_manifest_lazy( + self.args.manifest_dir / "tal_csasr_cuts_dev_set.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 / "tal_csasr_cuts_test_set.jsonl.gz" + ) diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/beam_search.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/beam_search.py new file mode 120000 index 000000000..ed78bd4bb --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/beam_search.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/conformer.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/conformer.py new file mode 120000 index 000000000..b2af4e1df --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/conformer.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py new file mode 100755 index 000000000..305729a99 --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py @@ -0,0 +1,755 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 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 +./pruned_transducer_stateless5/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless5/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./pruned_transducer_stateless5/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless5/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless5/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless5/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./pruned_transducer_stateless5/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless5/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +import argparse +import logging +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 TAL_CSASRAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.cut import Cut +from local.text_normalize import text_normalize +from train import add_model_arguments, get_params, get_transducer_model + +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, + write_error_stats, +) + + +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=False, + 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="pruned_transducer_stateless5/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An 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=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, + sp: spm.SentencePieceProcessor = 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. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = 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) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + zh_hyps = [] + en_hyps = [] + pattern = re.compile(r"([\u4e00-\u9fff])") + en_letter = "[\u0041-\u005a|\u0061-\u007a]+" # English letters + zh_char = "[\u4e00-\u9fa5]+" # Chinese chars + 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 i in range(encoder_out.size(0)): + hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + chars = pattern.split(hyp.upper()) + chars_new = [] + zh_text = [] + en_text = [] + for char in chars: + if char != "": + tokens = char.strip().split(" ") + chars_new.extend(tokens) + for token in tokens: + zh_text.extend(re.findall(zh_char, token)) + en_text.extend(re.findall(en_letter, token)) + hyps.append(chars_new) + zh_hyps.append(zh_text) + en_hyps.append(en_text) + 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 i in range(encoder_out.size(0)): + hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + chars = pattern.split(hyp.upper()) + chars_new = [] + zh_text = [] + en_text = [] + for char in chars: + if char != "": + tokens = char.strip().split(" ") + chars_new.extend(tokens) + for token in tokens: + zh_text.extend(re.findall(zh_char, token)) + en_text.extend(re.findall(en_letter, token)) + hyps.append(chars_new) + zh_hyps.append(zh_text) + en_hyps.append(en_text) + 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 i in range(encoder_out.size(0)): + hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + chars = pattern.split(hyp.upper()) + chars_new = [] + zh_text = [] + en_text = [] + for char in chars: + if char != "": + tokens = char.strip().split(" ") + chars_new.extend(tokens) + for token in tokens: + zh_text.extend(re.findall(zh_char, token)) + en_text.extend(re.findall(en_letter, token)) + hyps.append(chars_new) + zh_hyps.append(zh_text) + en_hyps.append(en_text) + 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}" + ) + for i in range(encoder_out.size(0)): + hyp = sp.decode( + [lexicon.token_table[idx] for idx in hyp_tokens[i]] + ) + chars = pattern.split(hyp.upper()) + chars_new = [] + zh_text = [] + en_text = [] + for char in chars: + if char != "": + tokens = char.strip().split(" ") + chars_new.extend(tokens) + for token in tokens: + zh_text.extend(re.findall(zh_char, token)) + en_text.extend(re.findall(en_letter, token)) + hyps.append(chars_new) + zh_hyps.append(zh_text) + en_hyps.append(en_text) + if params.decoding_method == "greedy_search": + return {"greedy_search": (hyps, zh_hyps, en_hyps)} + elif params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): (hyps, zh_hyps, en_hyps) + } + else: + return {f"beam_size_{params.beam_size}": (hyps, zh_hyps, en_hyps)} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + decoding_graph: Optional[k2.Fsa] = None, + sp: spm.SentencePieceProcessor = None, +) -> Dict[str, List[Tuple[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. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + zh_results = defaultdict(list) + en_results = defaultdict(list) + pattern = re.compile(r"([\u4e00-\u9fff])") + en_letter = "[\u0041-\u005a|\u0061-\u007a]+" # English letters + zh_char = "[\u4e00-\u9fa5]+" # Chinese chars + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + zh_texts = [] + en_texts = [] + for i in range(len(texts)): + text = texts[i] + chars = pattern.split(text.upper()) + chars_new = [] + zh_text = [] + en_text = [] + for char in chars: + if char != "": + tokens = char.strip().split(" ") + chars_new.extend(tokens) + for token in tokens: + zh_text.extend(re.findall(zh_char, token)) + en_text.extend(re.findall(en_letter, token)) + zh_texts.append(zh_text) + en_texts.append(en_text) + texts[i] = chars_new + hyps_dict = decode_one_batch( + params=params, + model=model, + lexicon=lexicon, + decoding_graph=decoding_graph, + batch=batch, + sp=sp, + ) + + for name, hyps_texts in hyps_dict.items(): + this_batch = [] + this_batch_zh = [] + this_batch_en = [] + # print(hyps_texts) + hyps, zh_hyps, en_hyps = hyps_texts + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + this_batch.append((ref_text, hyp_words)) + + for hyp_words, ref_text in zip(zh_hyps, zh_texts): + this_batch_zh.append((ref_text, hyp_words)) + + for hyp_words, ref_text in zip(en_hyps, en_texts): + this_batch_en.append((ref_text, hyp_words)) + + results[name].extend(this_batch) + zh_results[name + "_zh"].extend(this_batch_zh) + en_results[name + "_en"].extend(this_batch_en) + + 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, zh_results, en_results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + 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" + ) + 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() + TAL_CSASRAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{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}") + + bpe_model = params.lang_dir + "/bpe.model" + sp = spm.SentencePieceProcessor() + sp.load(bpe_model) + + lexicon = Lexicon(params.lang_dir) + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_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 params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + def text_normalize_for_cut(c: Cut): + # Text normalize for each sample + text = c.supervisions[0].text + text = text.strip("\n").strip("\t") + c.supervisions[0].text = text_normalize(text) + return c + + tal_csasr = TAL_CSASRAsrDataModule(args) + + dev_cuts = tal_csasr.valid_cuts() + dev_cuts = dev_cuts.map(text_normalize_for_cut) + dev_dl = tal_csasr.valid_dataloaders(dev_cuts) + + test_cuts = tal_csasr.test_cuts() + test_cuts = test_cuts.map(text_normalize_for_cut) + test_dl = tal_csasr.test_dataloaders(test_cuts) + + test_sets = ["dev", "test"] + test_dl = [dev_dl, test_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict, zh_results_dict, en_results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + lexicon=lexicon, + decoding_graph=decoding_graph, + sp=sp, + ) + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + save_results( + params=params, + test_set_name=test_set, + results_dict=zh_results_dict, + ) + save_results( + params=params, + test_set_name=test_set, + results_dict=en_results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/decoder.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decoder.py new file mode 120000 index 000000000..8a5e07bd5 --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decoder.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/encoder_interface.py new file mode 120000 index 000000000..2fc10439b --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/encoder_interface.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/export.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/export.py new file mode 100755 index 000000000..8f900208a --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/export.py @@ -0,0 +1,285 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# 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. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" +Usage: +./pruned_transducer_stateless5/export.py \ + --exp-dir ./pruned_transducer_stateless5/exp \ + --lang-dir ./data/lang_char \ + --epoch 30 \ + --avg 24 \ + --use-averaged-model True + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `pruned_transducer_stateless5/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/tal_csasr/ASR + ./pruned_transducer_stateless5/decode.py \ + --exp-dir ./pruned_transducer_stateless5/exp \ + --epoch 30 \ + --avg 24 \ + --max-duration 800 \ + --decoding-method greedy_search \ + --lang-dir ./data/lang_char +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for averaging. + 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=False, + 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="pruned_transducer_stateless5/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + bpe_model = params.lang_dir + "/bpe.model" + sp = spm.SentencePieceProcessor() + sp.load(bpe_model) + + lexicon = Lexicon(params.lang_dir) + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_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.eval() + + model.to("cpu") + model.eval() + + if params.jit: + # We won't use the forward() method of the model in C++, so just ignore + # it here. + # Otherwise, one of its arguments is a ragged tensor and is not + # torch scriptabe. + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torch.jit.script") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/joiner.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/joiner.py new file mode 120000 index 000000000..f31b5fd9b --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/joiner.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless2/joiner.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/model.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/model.py new file mode 120000 index 000000000..be059ba7c --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/model.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless2/model.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/optim.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/optim.py new file mode 120000 index 000000000..661206562 --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/optim.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless2/optim.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/pretrained.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/pretrained.py new file mode 100755 index 000000000..dbe213b24 --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/pretrained.py @@ -0,0 +1,375 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# 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. +""" +Usage: + +(1) greedy search +./pruned_transducer_stateless5/pretrained.py \ + --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \ + --lang-dir ./data/lang_char \ + --decoding-method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) modified beam search +./pruned_transducer_stateless5/pretrained.py \ + --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \ + --lang-dir ./data/lang_char \ + --decoding-method modified_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) fast beam search +./pruned_transducer_stateless5/pretrained.py \ + --checkpoint ./pruned_transducer_stateless5/exp/pretrained.pt \ + --lang-dir ./data/lang_char \ + --decoding-method fast_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +You can also use `./pruned_transducer_stateless5/exp/epoch-xx.pt`. + +Note: ./pruned_transducer_stateless5/exp/pretrained.pt is generated by +./pruned_transducer_stateless5/export.py +""" + + +import argparse +import logging +import math +import re +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.lexicon import Lexicon + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + add_model_arguments(parser) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + bpe_model = params.lang_dir + "/bpe.model" + sp = spm.SentencePieceProcessor() + sp.load(bpe_model) + + lexicon = Lexicon(params.lang_dir) + params.blank_di = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens = model.encoder( + x=features, x_lens=feature_lengths + ) + + num_waves = encoder_out.size(0) + hyps = [] + msg = f"Using {params.method}" + if params.method == "beam_search": + msg += f" with beam size {params.beam_size}" + logging.info(msg) + + pattern = re.compile(r"([\u4e00-\u9fff])") + if params.method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for i in range(encoder_out.size(0)): + hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + chars = pattern.split(hyp.upper()) + chars_new = [] + for char in chars: + if char != "": + chars_new.extend(char.strip().split(" ")) + hyps.append(chars_new) + elif params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for i in range(encoder_out.size(0)): + hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + chars = pattern.split(hyp.upper()) + chars_new = [] + for char in chars: + if char != "": + chars_new.extend(char.strip().split(" ")) + hyps.append(chars_new) + elif params.method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for i in range(encoder_out.size(0)): + hyp = sp.decode([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + chars = pattern.split(hyp.upper()) + chars_new = [] + for char in chars: + if char != "": + chars_new.extend(char.strip().split(" ")) + hyps.append(chars_new) + else: + for i in range(num_waves): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError(f"Unsupported method: {params.method}") + + hyp = sp.decode([lexicon.token_table[idx] for idx in hyp]) + chars = pattern.split(hyp.upper()) + chars_new = [] + for char in chars: + if char != "": + chars_new.extend(char.strip().split(" ")) + hyps.append(chars_new) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/scaling.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/scaling.py new file mode 120000 index 000000000..be7b111c6 --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/scaling.py @@ -0,0 +1 @@ +../../../../egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/test_model.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/test_model.py new file mode 100755 index 000000000..9aad32014 --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/test_model.py @@ -0,0 +1,65 @@ +#!/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. + + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./pruned_transducer_stateless4/test_model.py +""" + +from train import get_params, get_transducer_model + + +def test_model_1(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.num_encoder_layers = 24 + params.dim_feedforward = 1536 # 384 * 4 + params.encoder_dim = 384 + model = get_transducer_model(params) + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + +# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf +def test_model_M(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.num_encoder_layers = 18 + params.dim_feedforward = 1024 + params.encoder_dim = 256 + params.nhead = 4 + params.decoder_dim = 512 + params.joiner_dim = 512 + model = get_transducer_model(params) + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + +def main(): + # test_model_1() + test_model_M() + + +if __name__ == "__main__": + main() diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py new file mode 100755 index 000000000..ca35eba45 --- /dev/null +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/train.py @@ -0,0 +1,1115 @@ +#!/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_stateless5/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless5/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless5/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless5/exp \ + --full-libri 1 \ + --max-duration 550 + +""" + + +import argparse +import copy +import logging +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 TAL_CSASRAsrDataModule +from conformer import Conformer +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 local.text_normalize import text_normalize +from local.tokenize_with_bpe_model import tokenize_by_bpe_model +from model import Transducer +from optim import Eden, Eve +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 icefall import diagnostics +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +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.lexicon import Lexicon +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=int, + default=24, + help="Number of conformer encoder layers..", + ) + + parser.add_argument( + "--dim-feedforward", + type=int, + default=1536, + help="Feedforward dimension of the conformer encoder layer.", + ) + + parser.add_argument( + "--nhead", + type=int, + default=8, + help="Number of attention heads in the conformer encoder layer.", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=384, + help="Attention dimension in the conformer encoder layer.", + ) + + 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. + """, + ) + + +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=str, + default="pruned_transducer_stateless5/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--initial-lr", + type=float, + default=0.003, + help="The initial learning rate. This value should not need " + "to be changed.", + ) + + 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=6, + 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( + "--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 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=100, + 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 warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 100, + "valid_interval": 2000, + # parameters for conformer + "feature_dim": 80, + "subsampling_factor": 4, + # parameters for Noam + "model_warm_step": 1000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + ) + 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=params.encoder_dim, + 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=params.encoder_dim, + 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"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + 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], + graph_compiler: CharCtcTrainingGraphCompiler, + batch: dict, + is_training: bool, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + 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) + + texts = batch["supervisions"]["text"] + y = graph_compiler.texts_to_ids_with_bpe(texts) + if type(y) == list: + y = k2.RaggedTensor(y).to(device) + else: + y = 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, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.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], + graph_compiler: CharCtcTrainingGraphCompiler, + 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, + graph_compiler=graph_compiler, + 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, + graph_compiler: CharCtcTrainingGraphCompiler, + 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() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + # print(batch["supervisions"]) + + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # 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() + + 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 + ): + params.cur_batch_idx = batch_idx + 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, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[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}" + ) + + 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 batch_idx > 0 and batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + bpe_model = params.lang_dir + "/bpe.model" + import sentencepiece as spm + + sp = spm.SentencePieceProcessor() + sp.load(bpe_model) + + lexicon = Lexicon(params.lang_dir) + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + 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) + + 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]) + + optimizer = Eve(model.parameters(), lr=params.initial_lr) + + 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( + 2 ** 22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + tal_csasr = TAL_CSASRAsrDataModule(args) + train_cuts = tal_csasr.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 + return 1.0 <= c.duration <= 20.0 + + def text_normalize_for_cut(c: Cut): + # Text normalize for each sample + text = c.supervisions[0].text + text = text.strip("\n").strip("\t") + text = text_normalize(text) + text = tokenize_by_bpe_model(sp, text) + c.supervisions[0].text = text + return c + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + train_cuts = train_cuts.map(text_normalize_for_cut) + + 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 = tal_csasr.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = tal_csasr.valid_cuts() + valid_cuts = valid_cuts.map(text_normalize_for_cut) + valid_dl = tal_csasr.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + graph_compiler=graph_compiler, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16) + 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, + graph_compiler=graph_compiler, + 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 scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + graph_compiler: CharCtcTrainingGraphCompiler, + params: AttributeDict, +): + return + 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: + # warmup = 0.0 is so that the derivs for the pruned loss stay zero + # (i.e. are not remembered by the decaying-average in adam), because + # we want to avoid these params being subject to shrinkage in adam. + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + warmup=0.0, + ) + loss.backward() + optimizer.step() + 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]}) ..." + ) + raise + + +def main(): + parser = get_parser() + TAL_CSASRAsrDataModule.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/tal_csasr/ASR/shared b/egs/tal_csasr/ASR/shared new file mode 120000 index 000000000..e9461a6d7 --- /dev/null +++ b/egs/tal_csasr/ASR/shared @@ -0,0 +1 @@ +../../librispeech/ASR/shared \ No newline at end of file diff --git a/icefall/char_graph_compiler.py b/icefall/char_graph_compiler.py index a50b57d40..235160e14 100644 --- a/icefall/char_graph_compiler.py +++ b/icefall/char_graph_compiler.py @@ -79,6 +79,31 @@ class CharCtcTrainingGraphCompiler(object): ids.append(sub_ids) return ids + def texts_to_ids_with_bpe(self, texts: List[str]) -> List[List[int]]: + """Convert a list of texts (which include chars and bpes) + to a list-of-list of token IDs. + + Args: + texts: + It is a list of strings. + An example containing two strings is given below: + + [['你', '好', '▁C', 'hina'], ['北','京', '▁', 'welcome', '您'] + Returns: + Return a list-of-list of token IDs. + """ + ids: List[List[int]] = [] + for text in texts: + text = text.split("/") + sub_ids = [ + self.token_table[txt] + if txt in self.token_table + else self.oov_id + for txt in text + ] + ids.append(sub_ids) + return ids + def compile( self, token_ids: List[List[int]], diff --git a/icefall/utils.py b/icefall/utils.py index bd154dcec..360f29588 100644 --- a/icefall/utils.py +++ b/icefall/utils.py @@ -20,6 +20,7 @@ import argparse import collections import logging import os +import re import subprocess from collections import defaultdict from contextlib import contextmanager @@ -30,6 +31,7 @@ from typing import Dict, Iterable, List, TextIO, Tuple, Union import k2 import k2.version import kaldialign +import sentencepiece as spm import torch import torch.distributed as dist import torch.nn as nn @@ -876,3 +878,40 @@ def load_averaged_model( model.load_state_dict(average_checkpoints(filenames, device=device)) return model + + +def tokenize_by_bpe_model( + sp: spm.SentencePieceProcessor, + txt: str, +) -> str: + """ + Tokenize text with bpe model. This function is from + https://github1s.com/wenet-e2e/wenet/blob/main/wenet/dataset/processor.py#L322-L342. + Args: + sp: spm.SentencePieceProcessor. + txt: str + + Return: + A new string which includes chars and bpes. + """ + tokens = [] + # CJK(China Japan Korea) unicode range is [U+4E00, U+9FFF], ref: + # https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block) + pattern = re.compile(r"([\u4e00-\u9fff])") + # Example: + # txt = "你好 ITS'S OKAY 的" + # chars = ["你", "好", " ITS'S OKAY ", "的"] + chars = pattern.split(txt.upper()) + mix_chars = [w for w in chars if len(w.strip()) > 0] + for ch_or_w in mix_chars: + # ch_or_w is a single CJK charater(i.e., "你"), do nothing. + if pattern.fullmatch(ch_or_w) is not None: + tokens.append(ch_or_w) + # ch_or_w contains non-CJK charaters(i.e., " IT'S OKAY "), + # encode ch_or_w using bpe_model. + else: + for p in sp.encode_as_pieces(ch_or_w): + tokens.append(p) + txt_with_bpe = "/".join(tokens) + + return txt_with_bpe