diff --git a/egs/aishell/ASR/prepare.sh b/egs/aishell/ASR/prepare.sh index d36dc5ed30..9f73a2073c 100755 --- a/egs/aishell/ASR/prepare.sh +++ b/egs/aishell/ASR/prepare.sh @@ -261,10 +261,9 @@ if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then fi if [ ! -f $lang_char_dir/HLG.fst ]; then - lang_phone_dir=data/lang_phone ./local/prepare_lang_fst.py \ - --lang-dir $lang_phone_dir \ - --ngram-G ./data/lm/G_3_gram.fst.txt + --lang-dir $lang_char_dir \ + --ngram-G ./data/lm/G_3_gram_char.fst.txt fi fi diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7/decode.py index be58c4e43a..696eea9060 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7/decode.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7/decode.py @@ -641,7 +641,7 @@ def main(): contexts_text.append(line.strip()) contexts = graph_compiler.texts_to_ids(contexts_text) context_graph = ContextGraph(params.context_score) - context_graph.build(contexts) + context_graph.build([(c, 0.0) for c in contexts]) else: context_graph = None else: diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/train2.py b/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py similarity index 99% rename from egs/aishell/ASR/pruned_transducer_stateless7/train2.py rename to egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py index 057af297f0..6027273b2b 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7/train2.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py @@ -1234,6 +1234,7 @@ def scan_pessimistic_batches_for_oom( def main(): + raise RuntimeError("Please don't use this file directly!") parser = get_parser() AsrDataModule.add_arguments(parser) args = parser.parse_args() diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/export-onnx.py b/egs/aishell/ASR/pruned_transducer_stateless7/export-onnx.py index 2a9fc57d5f..39d988cd04 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7/export-onnx.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7/export-onnx.py @@ -56,7 +56,7 @@ from decoder2 import Decoder from onnxruntime.quantization import QuantType, quantize_dynamic from scaling_converter import convert_scaled_to_non_scaled -from train2 import add_model_arguments, get_params, get_transducer_model +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from zipformer import Zipformer from icefall.checkpoint import ( diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py index f5ae836fd7..99110d6b6e 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py @@ -686,7 +686,7 @@ def main(): contexts_text.append(line.strip()) contexts = graph_compiler.texts_to_ids(contexts_text) context_graph = ContextGraph(params.context_score) - context_graph.build(contexts) + context_graph.build([(c, 0.0) for c in contexts]) else: context_graph = None else: diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train2.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py similarity index 99% rename from egs/aishell/ASR/pruned_transducer_stateless7_streaming/train2.py rename to egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index 88eb341048..3c13c19c69 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train2.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -1233,6 +1233,7 @@ def scan_pessimistic_batches_for_oom( def main(): + raise RuntimeError("Please don't use this file directly!") parser = get_parser() AishellAsrDataModule.add_arguments(parser) args = parser.parse_args() diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/README.md b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/README.md index 991875aaa3..6c20bab2c0 100644 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/README.md +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/README.md @@ -4,6 +4,6 @@ See https://k2-fsa.github.io/icefall/recipes/Streaming-ASR/librispeech/zipformer [./emformer.py](./emformer.py) and [./train.py](./train.py) are basically the same as -[./emformer2.py](./emformer2.py) and [./train2.py](./train2.py). -The only purpose of [./emformer2.py](./emformer2.py) and [./train2.py](./train2.py) +[./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py). +The only purpose of [./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py) is for exporting to [sherpa-ncnn](https://github.com/k2-fsa/sherpa-ncnn). diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train2.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py similarity index 99% rename from egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train2.py rename to egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index c09c9537c5..61a3f27db6 100755 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train2.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -1237,6 +1237,7 @@ def scan_pessimistic_batches_for_oom( def main(): + raise RuntimeError("Please don't use this file directly!") parser = get_parser() CommonVoiceAsrDataModule.add_arguments(parser) args = parser.parse_args() diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train2.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py similarity index 99% rename from egs/csj/ASR/pruned_transducer_stateless7_streaming/train2.py rename to egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index 4c866ddd81..acde72d802 100755 --- a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train2.py +++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -1274,6 +1274,7 @@ def scan_pessimistic_batches_for_oom( def main(): + raise RuntimeError("Please don't use this file directly!") parser = get_parser() CSJAsrDataModule.add_arguments(parser) Tokenizer.add_arguments(parser) diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py index ebdb596a57..b210430c67 100755 --- a/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py +++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py @@ -72,7 +72,7 @@ import torch from scaling_converter import convert_scaled_to_non_scaled from tokenizer import Tokenizer -from train2 import add_model_arguments, get_params, get_transducer_model +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/libriheavy/ASR/README.md b/egs/libriheavy/ASR/README.md new file mode 100644 index 0000000000..2498d017fa --- /dev/null +++ b/egs/libriheavy/ASR/README.md @@ -0,0 +1,6 @@ +# Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context + +Libriheavy is a labeled version of [Librilight](https://arxiv.org/pdf/1912.07875.pdf). Please refer to our repository [k2-fsa/libriheavy](https://github.com/k2-fsa/libriheavy) for more details. We also have a paper: *Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context*, [Preprint available on arxiv](https://arxiv.org/abs/2309.08105). + + +See [RESULTS](./RESULTS.md) for the results for icefall recipes. diff --git a/egs/libriheavy/ASR/RESULTS.md b/egs/libriheavy/ASR/RESULTS.md index 4fbedad98d..513bbf72e8 100644 --- a/egs/libriheavy/ASR/RESULTS.md +++ b/egs/libriheavy/ASR/RESULTS.md @@ -1,6 +1,116 @@ -## Results +# Results -### Zipformer PromptASR (zipformer + PromptASR + BERT text encoder) +## zipformer (zipformer + pruned stateless transducer) + +See for more details. + +[zipformer](./zipformer) + +### Non-streaming + +#### Training on normalized text, i.e. Upper case without punctuation + +##### normal-scaled model, number of model parameters: 65805511, i.e., 65.81 M + +You can find a pretrained model, training logs at: + + +Note: The repository above contains three models trained on different subset of libriheavy exp(large set), exp_medium_subset(medium set), +exp_small_subset(small set). + +Results of models: + +| training set | decoding method | librispeech clean | librispeech other | libriheavy clean | libriheavy other | comment | +|---------------|---------------------|-------------------|-------------------|------------------|------------------|--------------------| +| small | greedy search | 4.19 | 9.99 | 4.75 | 10.25 |--epoch 90 --avg 20 | +| small | modified beam search| 4.05 | 9.89 | 4.68 | 10.01 |--epoch 90 --avg 20 | +| medium | greedy search | 2.39 | 4.85 | 2.90 | 6.6 |--epoch 60 --avg 20 | +| medium | modified beam search| 2.35 | 4.82 | 2.90 | 6.57 |--epoch 60 --avg 20 | +| large | greedy search | 1.67 | 3.32 | 2.24 | 5.61 |--epoch 16 --avg 3 | +| large | modified beam search| 1.62 | 3.36 | 2.20 | 5.57 |--epoch 16 --avg 3 | + +The training command is: +```bash +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +python ./zipformer/train.py \ + --world-size 4 \ + --master-port 12365 \ + --exp-dir zipformer/exp \ + --num-epochs 60 \ # 16 for large; 90 for small + --lr-hours 15000 \ # 20000 for large; 5000 for small + --use-fp16 1 \ + --start-epoch 1 \ + --bpe-model data/lang_bpe_500/bpe.model \ + --max-duration 1000 \ + --subset medium +``` + +The decoding command is: +```bash +export CUDA_VISIBLE_DEVICES="0" +for m in greedy_search modified_beam_search; do + ./zipformer/decode.py \ + --epoch 16 \ + --avg 3 \ + --exp-dir zipformer/exp \ + --max-duration 1000 \ + --causal 0 \ + --decoding-method $m +done +``` + +#### Training on full formatted text, i.e. with casing and punctuation + +##### normal-scaled model, number of model parameters: 66074067 , i.e., 66M + +You can find a pretrained model, training logs at: + + +Note: The repository above contains three models trained on different subset of libriheavy exp(large set), exp_medium_subset(medium set), +exp_small_subset(small set). + +Results of models: + +| training set | decoding method | libriheavy clean (WER) | libriheavy other (WER) | libriheavy clean (CER) | libriheavy other (CER) | comment | +|---------------|---------------------|-------------------|-------------------|------------------|------------------|--------------------| +| small | modified beam search| 13.04 | 19.54 | 4.51 | 7.90 |--epoch 88 --avg 41 | +| medium | modified beam search| 9.84 | 13.39 | 3.02 | 5.10 |--epoch 50 --avg 15 | +| large | modified beam search| 7.76 | 11.32 | 2.41 | 4.22 |--epoch 16 --avg 2 | + +The training command is: +```bash +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +python ./zipformer/train.py \ + --world-size 4 \ + --master-port 12365 \ + --exp-dir zipformer/exp \ + --num-epochs 60 \ # 16 for large; 90 for small + --lr-hours 15000 \ # 20000 for large; 10000 for small + --use-fp16 1 \ + --train-with-punctuation 1 \ + --start-epoch 1 \ + --bpe-model data/lang_punc_bpe_756/bpe.model \ + --max-duration 1000 \ + --subset medium +``` + +The decoding command is: +```bash +export CUDA_VISIBLE_DEVICES="0" +for m in greedy_search modified_beam_search; do + ./zipformer/decode.py \ + --epoch 16 \ + --avg 3 \ + --exp-dir zipformer/exp \ + --max-duration 1000 \ + --causal 0 \ + --decoding-method $m +done +``` + +## Zipformer PromptASR (zipformer + PromptASR + BERT text encoder) #### [zipformer_prompt_asr](./zipformer_prompt_asr) diff --git a/egs/libriheavy/ASR/local/compute_fbank_libriheavy.py b/egs/libriheavy/ASR/local/compute_fbank_libriheavy.py new file mode 100755 index 0000000000..010531db28 --- /dev/null +++ b/egs/libriheavy/ASR/local/compute_fbank_libriheavy.py @@ -0,0 +1,242 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang, +# 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 file computes fbank features of the Libriheavy 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 +from typing import Optional + +import torch +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + KaldifeatFbank, + KaldifeatFbankConfig, + LilcomChunkyWriter, +) + +from icefall.utils import get_executor, str2bool + +# 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 get_args(): + parser = argparse.ArgumentParser() + + parser.add_argument( + "--manifest-dir", + type=str, + help="""The source directory that contains raw manifests. + """, + default="data/manifests", + ) + + parser.add_argument( + "--fbank-dir", + type=str, + help="""Fbank output dir + """, + default="data/fbank", + ) + + parser.add_argument( + "--subset", + type=str, + help="""Dataset parts to compute fbank. If None, we will use all""", + ) + + parser.add_argument( + "--num-workers", + type=int, + default=20, + help="Number of dataloading workers used for reading the audio.", + ) + + parser.add_argument( + "--batch-duration", + type=float, + default=600.0, + help="The maximum number of audio seconds in a batch." + "Determines batch size dynamically.", + ) + + parser.add_argument( + "--perturb-speed", + type=str2bool, + default=False, + help="Whether to use speed perturbation.", + ) + + parser.add_argument( + "--use-splits", + type=str2bool, + default=False, + help="Whether to compute fbank on splits.", + ) + + parser.add_argument( + "--num-splits", + type=int, + help="""The number of splits of the medium and large subset. + Only needed when --use-splits is true.""", + ) + + parser.add_argument( + "--start", + type=int, + default=0, + help="""Process pieces starting from this number (inclusive). + Only needed when --use-splits is true.""", + ) + + parser.add_argument( + "--stop", + type=int, + default=-1, + help="""Stop processing pieces until this number (exclusive). + Only needed when --use-splits is true.""", + ) + + return parser.parse_args() + + +def compute_fbank_libriheavy(args): + src_dir = Path(args.manifest_dir) + output_dir = Path(args.fbank_dir) + num_jobs = min(15, os.cpu_count()) + num_mel_bins = 80 + subset = args.subset + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + output_cuts_path = output_dir / f"libriheavy_cuts_{subset}.jsonl.gz" + if output_cuts_path.exists(): + logging.info(f"{output_cuts_path} exists - skipping") + return + + input_cuts_path = src_dir / f"libriheavy_cuts_{subset}.jsonl.gz" + assert input_cuts_path.exists(), f"{input_cuts_path} does not exist!" + logging.info(f"Loading {input_cuts_path}") + cut_set = CutSet.from_file(input_cuts_path) + + logging.info("Computing features") + + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/libriheavy_feats_{subset}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + + logging.info(f"Saving to {output_cuts_path}") + cut_set.to_file(output_cuts_path) + + +def compute_fbank_libriheavy_splits(args): + num_splits = args.num_splits + subset = args.subset + src_dir = f"{args.manifest_dir}/libriheavy_{subset}_split" + src_dir = Path(src_dir) + output_dir = f"{args.fbank_dir}/libriheavy_{subset}_split" + output_dir = Path(output_dir) + output_dir.mkdir(parents=True, exist_ok=True) + + start = args.start + stop = args.stop + if stop < start: + stop = num_splits + + stop = min(stop, num_splits) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) + logging.info(f"device: {device}") + + num_digits = 8 # num_digits is fixed by lhotse split-lazy + for i in range(start, stop): + idx = f"{i + 1}".zfill(num_digits) + logging.info(f"Processing {idx}/{num_splits}") + + cuts_path = output_dir / f"libriheavy_cuts_{subset}.{idx}.jsonl.gz" + if cuts_path.is_file(): + logging.info(f"{cuts_path} exists - skipping") + continue + + raw_cuts_path = src_dir / f"libriheavy_cuts_{subset}.{idx}.jsonl.gz" + if not raw_cuts_path.is_file(): + logging.info(f"{raw_cuts_path} does not exist - skipping it") + continue + + logging.info(f"Loading {raw_cuts_path}") + cut_set = CutSet.from_file(raw_cuts_path) + + logging.info("Computing features") + if (output_dir / f"libriheavy_feats_{subset}_{idx}.lca").exists(): + logging.info(f"Removing {output_dir}/libriheavy_feats_{subset}_{idx}.lca") + os.remove(output_dir / f"libriheavy_feats_{subset}_{idx}.lca") + + cut_set = cut_set.compute_and_store_features_batch( + extractor=extractor, + storage_path=f"{output_dir}/libriheavy_feats_{subset}_{idx}", + num_workers=args.num_workers, + batch_duration=args.batch_duration, + overwrite=True, + ) + + logging.info("About to split cuts into smaller chunks.") + cut_set = cut_set.trim_to_supervisions( + keep_overlapping=False, min_duration=None + ) + + logging.info(f"Saving to {cuts_path}") + cut_set.to_file(cuts_path) + logging.info(f"Saved to {cuts_path}") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + args = get_args() + logging.info(vars(args)) + + if args.use_splits: + assert args.num_splits is not None, "Please provide num_splits" + compute_fbank_libriheavy_splits(args) + else: + compute_fbank_libriheavy(args) diff --git a/egs/libriheavy/ASR/local/compute_fbank_musan.py b/egs/libriheavy/ASR/local/compute_fbank_musan.py new file mode 120000 index 0000000000..5833f2484e --- /dev/null +++ b/egs/libriheavy/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/libriheavy/ASR/local/norm_text.py b/egs/libriheavy/ASR/local/norm_text.py new file mode 100755 index 0000000000..c2fc0d92d5 --- /dev/null +++ b/egs/libriheavy/ASR/local/norm_text.py @@ -0,0 +1,58 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import codecs +import sys + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--text", + type=str, + help="""Path to the input text. + """, + ) + return parser.parse_args() + + +def remove_punc_to_upper(text: str) -> str: + text = text.replace("‘", "'") + text = text.replace("’", "'") + tokens = set("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789'") + s_list = [x.upper() if x in tokens else " " for x in text] + s = " ".join("".join(s_list).split()).strip() + return s + + +def main(): + args = get_args() + if args.text: + f = codecs.open(args.text, encoding="utf-8") + else: + f = codecs.getreader("utf-8")(sys.stdin.buffer) + + sys.stdout = codecs.getwriter("utf-8")(sys.stdout.buffer) + line = f.readline() + while line: + print(remove_punc_to_upper(line)) + line = f.readline() + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/local/prepare_manifest.py b/egs/libriheavy/ASR/local/prepare_manifest.py new file mode 100755 index 0000000000..42f392cae4 --- /dev/null +++ b/egs/libriheavy/ASR/local/prepare_manifest.py @@ -0,0 +1,47 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import gzip +import json +import sys +from pathlib import Path + + +def simple_cleanup(text: str) -> str: + table = str.maketrans("’‘，。；？！（）：-《》、“”【】", "'',.;?!(): <>/\"\"[]") + text = text.translate(table) + return text.strip() + + +# Assign text of the supervisions and remove unnecessary entries. +def main(): + assert len(sys.argv) == 3, "Usage: ./local/prepare_manifest.py INPUT OUTPUT_DIR" + fname = Path(sys.argv[1]).name + oname = Path(sys.argv[2]) / fname + with gzip.open(sys.argv[1], "r") as fin, gzip.open(oname, "w") as fout: + for line in fin: + cut = json.loads(line) + cut["supervisions"][0]["text"] = simple_cleanup( + cut["supervisions"][0]["custom"]["texts"][0] + ) + del cut["supervisions"][0]["custom"] + del cut["custom"] + fout.write((json.dumps(cut) + "\n").encode()) + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/local/train_bpe_model.py b/egs/libriheavy/ASR/local/train_bpe_model.py new file mode 100755 index 0000000000..19caf43abd --- /dev/null +++ b/egs/libriheavy/ASR/local/train_bpe_model.py @@ -0,0 +1,113 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +# You can install sentencepiece via: +# +# pip install sentencepiece +# +# Due to an issue reported in +# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030 +# +# Please install a version >=0.1.96 + +import argparse +import shutil +from pathlib import Path + +import sentencepiece as spm + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + The generated bpe.model is saved to this directory. + """, + ) + + parser.add_argument( + "--byte-fallback", + action="store_true", + help="""Whether to enable byte_fallback when training bpe.""", + ) + + parser.add_argument( + "--character-coverage", + type=float, + default=1.0, + help="Character coverage in vocabulary.", + ) + + parser.add_argument( + "--transcript", + type=str, + help="Training transcript.", + ) + + parser.add_argument( + "--vocab-size", + type=int, + help="Vocabulary size for BPE training", + ) + + return parser.parse_args() + + +def main(): + args = get_args() + vocab_size = args.vocab_size + lang_dir = Path(args.lang_dir) + + model_type = "unigram" + + model_prefix = f"{lang_dir}/{model_type}_{vocab_size}" + train_text = args.transcript + input_sentence_size = 100000000 + + user_defined_symbols = ["", ""] + unk_id = len(user_defined_symbols) + # Note: unk_id is fixed to 2. + # If you change it, you should also change other + # places that are using it. + + model_file = Path(model_prefix + ".model") + if not model_file.is_file(): + spm.SentencePieceTrainer.train( + input=train_text, + vocab_size=vocab_size, + model_type=model_type, + model_prefix=model_prefix, + input_sentence_size=input_sentence_size, + character_coverage=args.character_coverage, + user_defined_symbols=user_defined_symbols, + byte_fallback=args.byte_fallback, + unk_id=unk_id, + bos_id=-1, + eos_id=-1, + ) + else: + print(f"{model_file} exists - skipping") + return + + shutil.copyfile(model_file, f"{lang_dir}/bpe.model") + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/prepare.sh b/egs/libriheavy/ASR/prepare.sh new file mode 100755 index 0000000000..af7e3c5b06 --- /dev/null +++ b/egs/libriheavy/ASR/prepare.sh @@ -0,0 +1,314 @@ +#!/usr/bin/env bash + +# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674 +export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python + +set -eou pipefail + +nj=15 +stage=-1 +stop_stage=100 +export CUDA_VISIBLE_DEVICES="" + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/librilight +# You can find small, medium, large, etc. inside it. +# +# - $dl_dir/libriheavy +# You can find libriheavy_cuts_small.jsonl.gz, libriheavy_cuts_medium.jsonl.gz, etc. inside it. +# +# - $dl_dir/musan +# This directory contains the following directories downloaded from +# http://www.openslr.org/17/ +# +# - music +# - noise +# - speech +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# vocab size for sentence piece models. +# It will generate data/lang_bpe_xxx, +# data/lang_bpe_yyy if the array contains xxx, yyy +vocab_sizes=( + # 5000 + # 2000 + # 1000 + 500 +) + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data +fbank_dir=data/fbank +manifests_dir=data/manifests + +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 -1 ] && [ $stop_stage -ge -1 ]; then + log "Stage -1: Download audio data." + # If you have pre-downloaded it to /path/to/librilight, + # you can create a symlink + # + # ln -sfv /path/to/librilight $dl_dir/librilight + # + mkdir -p $dl_dir/librilight + for subset in small medium large; do + log "Downloading ${subset} subset." + if [ ! -d $dl_dir/librilight/${subset} ]; then + wget -P $dl_dir/librilight -c https://dl.fbaipublicfiles.com/librilight/data/${subset}.tar + tar xf $dl_dir/librilight/${subset}.tar -C $dl_dir/librilight + else + log "Skipping download, ${subset} subset exists." + fi + done +fi + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "Stage 0: Download manifests from huggingface." + + # If you have pre-downloaded it to /path/to/libriheavy, + # you can create a symlink + # + # ln -sfv /path/to/libriheavy $dl_dir/libriheavy + # + mkdir -p $dl_dir/libriheavy + for subset in small medium large dev test_clean test_other; do + if [ ! -e $dl_dir/libriheavy/libriheavy_cuts_${subset}.jsonl.gz ]; then + log "Downloading ${subset} subset." + wget -P $dl_dir/libriheavy -c https://huggingface.co/datasets/pkufool/libriheavy/resolve/main/libriheavy_cuts_${subset}.jsonl.gz + else + log "Skipping download, ${subset} subset exists." + fi + done + + # If you have pre-downloaded it to /path/to/musan, + # you can create a symlink + # + # ln -sfv /path/to/musan $dl_dir/ + # + 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: Download manifests from modelscope" + mkdir -p $dl_dir/libriheavy + if [ ! -e $dl_dir/libriheavy/libriheavy_cuts_small.jsonl.gz ]; then + cd $dl_dir/libriheavy + GIT_LFS_SKIP_SMUDGE=1 git clone https://www.modelscope.cn/datasets/pkufool/Libriheavy.git + cd Libriheavy + git lfs pull --exclude "raw/*" + mv *.jsonl.gz ../ + cd .. + rm -rf Libriheavy + cd ../../ + 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 $dl_dir/musan + mkdir -p $manifests_dir + if [ ! -e $manifests_dir/.musan.done ]; then + lhotse prepare musan $dl_dir/musan $manifests_dir + touch $manifests_dir/.musan.done + fi +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Prepare Libriheavy manifests" + mkdir -p $manifests_dir + for subset in small medium large dev test_clean test_other; do + if [ ! -e $manifests_dir/libriheavy_cuts_${subset}.jsonl.gz ]; then + log "Prepare manifest for subset : ${subset}" + ./local/prepare_manifest.py $dl_dir/libriheavy/libriheavy_cuts_${subset}.jsonl.gz $manifests_dir + fi + done +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Compute fbank for musan" + mkdir -p $fbank_dir + if [ ! -e $fbank_dir/.musan.done ]; then + ./local/compute_fbank_musan.py + touch $fbank_dir/.musan.done + fi +fi + +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Compute fbank for small subset and validation subsets" + for subset in test_clean test_other dev small; do + log "Computing $subset subset." + if [ ! -e $fbank_dir/.libriheavy.${subset}.done ]; then + ./local/compute_fbank_libriheavy.py \ + --manifest-dir ${manifests_dir} \ + --subset ${subset} \ + --fbank-dir $fbank_dir \ + --num-workers $nj + fi + done +fi + +num_per_split=8000 +if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then + log "Stage 6: Split medium and large subsets." + for subset in medium large; do + log "Spliting subset : $subset" + split_dir=$manifests_dir/libriheavy_${subset}_split + mkdir -p $split_dir + if [ ! -e $split_dir/.split_completed ]; then + lhotse split-lazy $manifests_dir/libriheavy_cuts_${subset}.jsonl.gz $split_dir $num_per_split + touch $split_dir/.split_completed + fi + done +fi + +if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then + log "Stage 7: Compute fbank for medium and large subsets" + mkdir -p $fbank_dir + chunk_size=20 + for subset in medium large; do + if [ $subset == "large" ]; then + chunk_size=200 + fi + num_splits=$(find $manifests_dir/libriheavy_${subset}_split -name "libriheavy_cuts_${subset}.*.jsonl.gz" | wc -l) + if [ ! -e $fbank_dir/.libriheavy.${subset}.done ]; then + for i in $(seq 0 1 6); do + start=$(( i * $chunk_size )) + end=$(( (i+1) * $chunk_size )) + ./local/compute_fbank_libriheavy.py \ + --manifest-dir ${manifests_dir} \ + --use-splits 1 \ + --subset ${subset} \ + --fbank-dir $fbank_dir \ + --num-splits $num_splits \ + --num-workers $nj \ + --start $start \ + --stop $end & + done + wait + touch $fbank_dir/.libriheavy.${subset}.done + fi + done +fi + +if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then + log "Stage 8: Combine features for medium and large subsets." + for subset in medium large; do + log "Combining $subset subset." + if [ ! -f $fbank_dir/libriheavy_cuts_${subset}.jsonl.gz ]; then + pieces=$(find $fbank_dir/libriheavy_${subset}_split -name "libriheavy_cuts_${subset}.*.jsonl.gz") + lhotse combine $pieces $fbank_dir/libriheavy_cuts_${subset}.jsonl.gz + fi + done +fi + +if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then + log "Stage 9: Train BPE model for normalized text" + + if [ ! -f data/texts ]; then + gunzip -c $manifests_dir/libriheavy_cuts_medium.jsonl.gz \ + | jq '.supervisions[].text' | sed 's/"//;s/\\//g;s/"$//' \ + | ./local/norm_text.py > data/texts + fi + + for vocab_size in ${vocab_sizes[@]}; do + lang_dir=data/lang_bpe_${vocab_size} + mkdir -p $lang_dir + + cp data/texts $lang_dir/text + + if [ ! -f $lang_dir/bpe.model ]; then + ./local/train_bpe_model.py \ + --lang-dir $lang_dir \ + --vocab-size $vocab_size \ + --transcript $lang_dir/text + fi + done +fi + + +if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then + log "Stage 10: Train BPE model for unnormalized text" + if [ ! -f data/punc_texts ]; then + gunzip -c $manifests_dir/libriheavy_cuts_medium.jsonl.gz \ + | jq '.supervisions[].text' | sed 's/"//;s/\\//g;s/"$//' > data/punc_texts + fi + for vocab_size in ${vocab_sizes[@]}; do + new_vacab_size = $(($vocab_size + 256)) + lang_dir=data/lang_punc_bpe_${new_vocab_size} + mkdir -p $lang_dir + + cp data/punc_texts $lang_dir/text + + if [ ! -f $lang_dir/bpe.model ]; then + ./local/train_bpe_model.py \ + --lang-dir $lang_dir \ + --byte-fallback \ + --vocab-size ${new_vocab_size} \ + --byte-fallback \ + --character-coverage 0.99 \ + --transcript $lang_dir/text + fi + done +fi + +if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then + log "Stage 11: Prepare language model for normalized text" + + for subset in small medium large; do + if [ ! -f $manifests_dir/texts_${subset} ]; then + gunzip -c $manifests_dir/libriheavy_cuts_${subset}.jsonl.gz \ + | jq '.supervisions[].text' | sed 's/"//;s/\\//g;s/"$//' \ + | ./local/norm_text.py > $manifests_dir/texts_${subset} + fi + done + + mkdir -p data/lm + if [ ! -f data/lm/text ]; then + cat $manifests_dir/texts_small $manifests_dir/texts_medium $manifests_dir/texts_large > data/lm/text + fi + + (echo ' 0'; echo '!SIL 1'; echo ' 2'; echo ' 3';) \ + > data/lm/words.txt + + cat data/lm/text | sed 's/ /\n/g' | sort -u | sed '/^$/d' \ + | awk '{print $1" "NR+3}' >> data/lm/words.txt + + num_lines=$(< data/lm/words.txt wc -l) + (echo "#0 $num_lines"; echo " $(($num_lines + 1))"; echo " $(($num_lines + 2))";) \ + >> data/lm/words.txt + + # Train LM on transcripts + if [ ! -f data/lm/3-gram.unpruned.arpa ]; then + python3 ./shared/make_kn_lm.py \ + -ngram-order 3 \ + -text data/lm/text \ + -lm data/lm/3-gram.unpruned.arpa + fi + + # We assume you have install kaldilm, if not, please install + # it using: pip install kaldilm + if [ ! -f data/lm/G_3_gram_char.fst.txt ]; then + # It is used in building HLG + python3 -m kaldilm \ + --read-symbol-table=data/lm/words.txt \ + --disambig-symbol='#0' \ + --max-order=3 \ + data/lm/3-gram.unpruned.arpa > data/lm/G_3_gram.fst.txt + fi +fi + diff --git a/egs/libriheavy/ASR/zipformer/asr_datamodule.py b/egs/libriheavy/ASR/zipformer/asr_datamodule.py new file mode 100644 index 0000000000..df761c1b8d --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/asr_datamodule.py @@ -0,0 +1,443 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022-2023 Xiaomi Corporation (Authors: Mingshuang Luo, +# 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. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, 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, + SimpleCutSampler, + 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 LibriHeavyAsrDataModule: + """ + 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( + "--subset", + type=str, + default="S", + help="""The subset to be used. Should be S, M or L. Note: S subset + includes libriheavy_cuts_small.jsonl.gz, M subset includes + libriheavy_cuts_small.jsonl.gz and libriheavy_cuts_medium.jsonl.gz, + L subset includes libriheavy_cuts_small.jsonl.gz, + libriheavy_cuts_medium.jsonl.gz and libriheavy_cuts_large.jsonl.gz. + """, + ) + + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help="When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--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. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + transforms.append( + CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + if self.args.concatenate_cuts: + logging.info( + f"Using cut concatenation with duration factor " + f"{self.args.duration_factor} and gap {self.args.gap}." + ) + # Cut concatenation should be the first transform in the list, + # so that if we e.g. mix noise in, it will fill the gaps between + # different utterances. + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + 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, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # '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, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + if self.args.concatenate_cuts: + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + logging.info("About to create dev dataset") + if self.args.on_the_fly_feats: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + return_cuts=self.args.return_cuts, + ) + else: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.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, + shuffle=False, + ) + logging.debug("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_small_cuts(self) -> CutSet: + logging.info("About to get small subset cuts") + return load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_small.jsonl.gz" + ) + + @lru_cache() + def train_medium_cuts(self) -> CutSet: + logging.info("About to get medium subset cuts") + return load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_medium.jsonl.gz" + ) + + @lru_cache() + def train_large_cuts(self) -> CutSet: + logging.info("About to get large subset cuts") + return load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_large.jsonl.gz" + ) + + @lru_cache() + def dev_cuts(self) -> CutSet: + logging.info("About to get dev cuts") + return load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_dev.jsonl.gz" + ) + + @lru_cache() + def test_clean_cuts(self) -> CutSet: + logging.info("About to get the test-clean cuts") + return load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_test_clean.jsonl.gz" + ) + + @lru_cache() + def test_other_cuts(self) -> CutSet: + logging.info("About to get the test-other cuts") + return load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_test_other.jsonl.gz" + ) diff --git a/egs/libriheavy/ASR/zipformer/beam_search.py b/egs/libriheavy/ASR/zipformer/beam_search.py new file mode 120000 index 0000000000..e24eca39f2 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/decode.py b/egs/libriheavy/ASR/zipformer/decode.py new file mode 100644 index 0000000000..1928e26358 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/decode.py @@ -0,0 +1,794 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(6) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 +""" + + +import argparse +import logging +import math +import warnings +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 LibriHeavyAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.cut import Cut +from text_normalization import remove_punc_to_upper +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--train-with-punctuation", + type=str2bool, + default=False, + help="""Set to True, if the model was trained on texts with casing + and punctuation.""", + ) + + parser.add_argument( + "--post-normalization", + type=str2bool, + default=False, + help="""Upper case and remove all chars except ' and - + """, + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + this_batch = [] + if params.post_normalization and params.train_with_punctuation: + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = remove_punc_to_upper(ref_text).split() + hyp_words = remove_punc_to_upper(" ".join(hyp_words)).split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[f"{name}_norm"].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriHeavyAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + libriheavy = LibriHeavyAsrDataModule(args) + + def normalize_text(c: Cut): + text = remove_punc_to_upper(c.supervisions[0].text) + c.supervisions[0].text = text + return c + + test_clean_cuts = libriheavy.test_clean_cuts() + test_other_cuts = libriheavy.test_other_cuts() + + if not params.train_with_punctuation: + test_clean_cuts = test_clean_cuts.map(normalize_text) + test_other_cuts = test_other_cuts.map(normalize_text) + + test_clean_dl = libriheavy.test_dataloaders(test_clean_cuts) + test_other_dl = libriheavy.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer/decoder.py b/egs/libriheavy/ASR/zipformer/decoder.py new file mode 120000 index 0000000000..5a8018680d --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/encoder_interface.py b/egs/libriheavy/ASR/zipformer/encoder_interface.py new file mode 120000 index 0000000000..c2eaca6712 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/export-onnx.py b/egs/libriheavy/ASR/zipformer/export-onnx.py new file mode 120000 index 0000000000..70a15683c2 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/export-onnx.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/export.py b/egs/libriheavy/ASR/zipformer/export.py new file mode 120000 index 0000000000..dfc1bec080 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/export.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/jit_pretrained.py b/egs/libriheavy/ASR/zipformer/jit_pretrained.py new file mode 120000 index 0000000000..25108391fa --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/jit_pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/jit_pretrained.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/joiner.py b/egs/libriheavy/ASR/zipformer/joiner.py new file mode 120000 index 0000000000..5b8a36332e --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/model.py b/egs/libriheavy/ASR/zipformer/model.py new file mode 120000 index 0000000000..cd7e07d72b --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/onnx_decode.py b/egs/libriheavy/ASR/zipformer/onnx_decode.py new file mode 120000 index 0000000000..0573b88c5b --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/onnx_decode.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_decode.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/onnx_pretrained.py b/egs/libriheavy/ASR/zipformer/onnx_pretrained.py new file mode 120000 index 0000000000..8f32f4ee7a --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/onnx_pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/optim.py b/egs/libriheavy/ASR/zipformer/optim.py new file mode 120000 index 0000000000..5eaa3cffd4 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/pretrained.py b/egs/libriheavy/ASR/zipformer/pretrained.py new file mode 120000 index 0000000000..0bd71dde4d --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/pretrained.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/scaling.py b/egs/libriheavy/ASR/zipformer/scaling.py new file mode 120000 index 0000000000..6f398f431d --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/scaling_coverter.py b/egs/libriheavy/ASR/zipformer/scaling_coverter.py new file mode 120000 index 0000000000..b0ecee05e1 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/scaling_coverter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/subsampling.py b/egs/libriheavy/ASR/zipformer/subsampling.py new file mode 120000 index 0000000000..01ae9002c6 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer/text_normalization.py b/egs/libriheavy/ASR/zipformer/text_normalization.py new file mode 100644 index 0000000000..92590769ce --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/text_normalization.py @@ -0,0 +1,50 @@ +from num2words import num2words + + +def remove_punc_to_upper(text: str) -> str: + text = text.replace("‘", "'") + text = text.replace("’", "'") + tokens = set("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789'") + s_list = [x.upper() if x in tokens else " " for x in text] + s = " ".join("".join(s_list).split()).strip() + return s + + +def word_normalization(word: str) -> str: + # 1. Use full word for some abbreviation + # 2. Convert digits to english words + # 3. Convert ordinal number to english words + if word == "MRS": + return "MISSUS" + if word == "MR": + return "MISTER" + if word == "ST": + return "SAINT" + if word == "ECT": + return "ET CETERA" + + if word[-2:] in ("ST", "ND", "RD", "TH") and word[:-2].isnumeric(): # e.g 9TH, 6TH + word = num2words(word[:-2], to="ordinal") + word = word.replace("-", " ") + + if word.isnumeric(): + num = int(word) + if num > 1500 and num < 2030: + word = num2words(word, to="year") + else: + word = num2words(word) + word = word.replace("-", " ") + return word.upper() + + +def text_normalization(text: str) -> str: + text = text.upper() + return " ".join([word_normalization(x) for x in text.split()]) + + +if __name__ == "__main__": + assert remove_punc_to_upper("I like this 《book>") == "I LIKE THIS BOOK" + assert ( + text_normalization("Hello Mrs st 21st world 3rd she 99th MR") + == "HELLO MISSUS SAINT TWENTY FIRST WORLD THIRD SHE NINETY NINTH MISTER" + ) diff --git a/egs/libriheavy/ASR/zipformer/train.py b/egs/libriheavy/ASR/zipformer/train.py new file mode 100644 index 0000000000..c97da4a117 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/train.py @@ -0,0 +1,1415 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey, +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +# For non-streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --full-libri 1 \ + --max-duration 1000 + +# For streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --causal 1 \ + --full-libri 1 \ + --max-duration 1000 + +It supports training with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +import logging +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriHeavyAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from text_normalization import remove_punc_to_upper +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-hours", + type=float, + default=30000, + help="""Number of hours that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--train-with-punctuation", + type=str2bool, + default=False, + help="If True, the training text will include casing and punctuation.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + # Use the number of hours of speech to adjust the learning rate + scheduler.step_epoch( + params.batch_idx_train * params.max_duration * params.world_size / 3600 + ) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_hours) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + def normalize_text(c: Cut): + text = remove_punc_to_upper(c.supervisions[0].text) + c.supervisions[0].text = text + return c + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 2.0 or c.duration > 30.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = sp.encode(c.supervisions[0].text, out_type=str) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + libriheavy = LibriHeavyAsrDataModule(args) + + train_cuts = libriheavy.train_small_cuts() + if params.subset == "M" or params.subset == "L": + train_cuts += libriheavy.train_medium_cuts() + if params.subset == "L": + train_cuts += libriheavy.train_large_cuts() + + if not params.train_with_punctuation: + train_cuts = train_cuts.map(normalize_text) + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = libriheavy.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = libriheavy.dev_cuts() + + if not params.train_with_punctuation: + valid_cuts = valid_cuts.map(normalize_text) + + valid_dl = libriheavy.valid_dataloaders(valid_cuts) + + # if not params.print_diagnostics: + # scan_pessimistic_batches_for_oom( + # model=model, + # train_dl=train_dl, + # optimizer=optimizer, + # sp=sp, + # params=params, + # ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriHeavyAsrDataModule.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/libriheavy/ASR/zipformer/zipformer.py b/egs/libriheavy/ASR/zipformer/zipformer.py new file mode 120000 index 0000000000..23011dda71 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train2.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/do_not_use_it_directly.py similarity index 99% rename from egs/librispeech/ASR/conv_emformer_transducer_stateless2/train2.py rename to egs/librispeech/ASR/conv_emformer_transducer_stateless2/do_not_use_it_directly.py index 420dc1065a..d614f0914e 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train2.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/do_not_use_it_directly.py @@ -1099,6 +1099,7 @@ def scan_pessimistic_batches_for_oom( def main(): + raise RuntimeError("Please don't use this file directly!") parser = get_parser() LibriSpeechAsrDataModule.add_arguments(parser) args = parser.parse_args() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-for-ncnn.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-for-ncnn.py index 85dbd4661a..953f95c458 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-for-ncnn.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-for-ncnn.py @@ -39,8 +39,8 @@ import k2 import torch +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from scaling_converter import convert_scaled_to_non_scaled -from train2 import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py index ab046557fb..1e59e08583 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py @@ -61,7 +61,7 @@ from decoder import Decoder from emformer import Emformer from scaling_converter import convert_scaled_to_non_scaled -from train2 import add_model_arguments, get_params, get_transducer_model +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py index 5243660687..5195a4ef6c 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py @@ -927,9 +927,9 @@ def main(): if os.path.exists(params.context_file): contexts = [] for line in open(params.context_file).readlines(): - contexts.append(line.strip()) + contexts.append((sp.encode(line.strip()), 0.0)) context_graph = ContextGraph(params.context_score) - context_graph.build(sp.encode(contexts)) + context_graph.build(contexts) else: context_graph = None else: diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/README.md b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/README.md index d3691e6473..0f3c63e752 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/README.md +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/README.md @@ -4,7 +4,7 @@ See https://k2-fsa.github.io/icefall/recipes/Streaming-ASR/librispeech/zipformer [./emformer.py](./emformer.py) and [./train.py](./train.py) are basically the same as -[./emformer2.py](./emformer2.py) and [./train2.py](./train2.py). -The only purpose of [./emformer2.py](./emformer2.py) and [./train2.py](./train2.py) +[./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py). +The only purpose of [./emformer2.py](./emformer2.py) and [./do_not_use_it_directly.py](./do_not_use_it_directly.py) is for exporting to [sherpa-ncnn](https://github.com/k2-fsa/sherpa-ncnn). diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train2.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py similarity index 99% rename from egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train2.py rename to egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index aa6c0668a9..cd26db6f3b 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train2.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -1234,6 +1234,7 @@ def scan_pessimistic_batches_for_oom( def main(): + raise RuntimeError("Please don't use this file directly!") parser = get_parser() LibriSpeechAsrDataModule.add_arguments(parser) args = parser.parse_args() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py index 07de57a86f..a7d06a5ddb 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py @@ -68,8 +68,8 @@ import k2 import torch +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from scaling_converter import convert_scaled_to_non_scaled -from train2 import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py index 9a6b31268d..8f2178b1d6 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py @@ -66,8 +66,8 @@ import k2 import torch +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from scaling_converter import convert_scaled_to_non_scaled -from train2 import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/do_not_use_it_directly.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/do_not_use_it_directly.py new file mode 120000 index 0000000000..beeffaa03c --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/do_not_use_it_directly.py @@ -0,0 +1 @@ +../pruned_transducer_stateless7_streaming/do_not_use_it_directly.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn.py index 9a6b31268d..8f2178b1d6 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/export-for-ncnn.py @@ -66,8 +66,8 @@ import k2 import torch +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from scaling_converter import convert_scaled_to_non_scaled -from train2 import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train2.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train2.py deleted file mode 120000 index 3c3280b688..0000000000 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train2.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless7_streaming/train2.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer/decode.py b/egs/librispeech/ASR/zipformer/decode.py index 3531d657f5..339e253e6e 100755 --- a/egs/librispeech/ASR/zipformer/decode.py +++ b/egs/librispeech/ASR/zipformer/decode.py @@ -1001,9 +1001,9 @@ def main(): if os.path.exists(params.context_file): contexts = [] for line in open(params.context_file).readlines(): - contexts.append(line.strip()) + contexts.append((sp.encode(line.strip()), 0.0)) context_graph = ContextGraph(params.context_score) - context_graph.build(sp.encode(contexts)) + context_graph.build(contexts) else: context_graph = None else: diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py index 36b8a4b670..d665f33645 100755 --- a/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py +++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/decode.py @@ -868,7 +868,7 @@ def main(): contexts_text.append(line.strip()) contexts = graph_compiler.texts_to_ids(contexts_text) context_graph = ContextGraph(params.context_score) - context_graph.build(contexts) + context_graph.build([(c, 0.0) for c in contexts]) else: context_graph = None else: diff --git a/icefall/context_graph.py b/icefall/context_graph.py index 0b7c42c0b9..b3d7972a8e 100644 --- a/icefall/context_graph.py +++ b/icefall/context_graph.py @@ -84,6 +84,9 @@ def __init__(self, context_score: float): context_score: The bonus score for each token(note: NOT for each word/phrase, it means longer word/phrase will have larger bonus score, they have to be matched though). + Note: This is just the default score for each token, the users can manually + specify the context_score for each word/phrase (i.e. different phrase might + have different token score). """ self.context_score = context_score self.num_nodes = 0 @@ -133,7 +136,7 @@ def _fill_fail_output(self): node.output_score += 0 if output is None else output.output_score queue.append(node) - def build(self, token_ids: List[List[int]]): + def build(self, token_ids: List[Tuple[List[int], float]]): """Build the ContextGraph from a list of token list. It first build a trie from the given token lists, then fill the fail arc for each trie node. @@ -142,26 +145,46 @@ def build(self, token_ids: List[List[int]]): Args: token_ids: - The given token lists to build the ContextGraph, it is a list of token list, - each token list contains the token ids for a word/phrase. The token id - could be an id of a char (modeling with single Chinese char) or an id - of a BPE (modeling with BPEs). + The given token lists to build the ContextGraph, it is a list of tuple of + token list and its customized score, the token list contains the token ids + for a word/phrase. The token id could be an id of a char + (modeling with single Chinese char) or an id of a BPE + (modeling with BPEs). The score is the total score for current token list, + 0 means using the default value (i.e. self.context_score). + + Note: The phrases would have shared states, the score of the shared states is + the maximum value among all the tokens sharing this state. """ - for tokens in token_ids: + for (tokens, score) in token_ids: node = self.root + # If has customized score using the customized token score, otherwise + # using the default score + context_score = ( + self.context_score if score == 0.0 else round(score / len(tokens), 2) + ) for i, token in enumerate(tokens): + node_next = {} if token not in node.next: self.num_nodes += 1 + node_id = self.num_nodes + token_score = context_score is_end = i == len(tokens) - 1 - node_score = node.node_score + self.context_score - node.next[token] = ContextState( - id=self.num_nodes, - token=token, - token_score=self.context_score, - node_score=node_score, - output_score=node_score if is_end else 0, - is_end=is_end, - ) + else: + # node exists, get the score of shared state. + token_score = max(context_score, node.next[token].token_score) + node_id = node.next[token].id + node_next = node.next[token].next + is_end = i == len(tokens) - 1 or node.next[token].is_end + node_score = node.node_score + token_score + node.next[token] = ContextState( + id=node_id, + token=token, + token_score=token_score, + node_score=node_score, + output_score=node_score if is_end else 0, + is_end=is_end, + ) + node.next[token].next = node_next node = node.next[token] self._fill_fail_output() @@ -343,7 +366,7 @@ def draw( return dot -if __name__ == "__main__": +def _test(queries, score): contexts_str = [ "S", "HE", @@ -355,9 +378,11 @@ def draw( "THIS", "THEM", ] + + # test default score (1) contexts = [] for s in contexts_str: - contexts.append([ord(x) for x in s]) + contexts.append(([ord(x) for x in s], score)) context_graph = ContextGraph(context_score=1) context_graph.build(contexts) @@ -369,21 +394,10 @@ def draw( context_graph.draw( title="Graph for: " + " / ".join(contexts_str), - filename="context_graph.pdf", + filename=f"context_graph_{score}.pdf", symbol_table=symbol_table, ) - queries = { - "HEHERSHE": 14, # "HE", "HE", "HERS", "S", "SHE", "HE" - "HERSHE": 12, # "HE", "HERS", "S", "SHE", "HE" - "HISHE": 9, # "HIS", "S", "SHE", "HE" - "SHED": 6, # "S", "SHE", "HE" - "SHELF": 6, # "S", "SHE", "HE" - "HELL": 2, # "HE" - "HELLO": 7, # "HE", "HELLO" - "DHRHISQ": 4, # "HIS", "S" - "THEN": 2, # "HE" - } for query, expected_score in queries.items(): total_scores = 0 state = context_graph.root @@ -393,8 +407,47 @@ def draw( score, state = context_graph.finalize(state) assert state.token == -1, state.token total_scores += score - assert total_scores == expected_score, ( + assert round(total_scores, 2) == expected_score, ( total_scores, expected_score, query, ) + + +if __name__ == "__main__": + # test default score + queries = { + "HEHERSHE": 14, # "HE", "HE", "HERS", "S", "SHE", "HE" + "HERSHE": 12, # "HE", "HERS", "S", "SHE", "HE" + "HISHE": 9, # "HIS", "S", "SHE", "HE" + "SHED": 6, # "S", "SHE", "HE" + "SHELF": 6, # "S", "SHE", "HE" + "HELL": 2, # "HE" + "HELLO": 7, # "HE", "HELLO" + "DHRHISQ": 4, # "HIS", "S" + "THEN": 2, # "HE" + } + _test(queries, 0) + + # test custom score (5) + # S : 5 + # HE : 5 (2.5 + 2.5) + # SHE : 8.34 (5 + 1.67 + 1.67) + # SHELL : 10.34 (5 + 1.67 + 1.67 + 1 + 1) + # HIS : 5.84 (2.5 + 1.67 + 1.67) + # HERS : 7.5 (2.5 + 2.5 + 1.25 + 1.25) + # HELLO : 8 (2.5 + 2.5 + 1 + 1 + 1) + # THIS : 5 (1.25 + 1.25 + 1.25 + 1.25) + queries = { + "HEHERSHE": 35.84, # "HE", "HE", "HERS", "S", "SHE", "HE" + "HERSHE": 30.84, # "HE", "HERS", "S", "SHE", "HE" + "HISHE": 24.18, # "HIS", "S", "SHE", "HE" + "SHED": 18.34, # "S", "SHE", "HE" + "SHELF": 18.34, # "S", "SHE", "HE" + "HELL": 5, # "HE" + "HELLO": 13, # "HE", "HELLO" + "DHRHISQ": 10.84, # "HIS", "S" + "THEN": 5, # "HE" + } + + _test(queries, 5) diff --git a/requirements-ci.txt b/requirements-ci.txt index e1232a7689..6c74f688ce 100644 --- a/requirements-ci.txt +++ b/requirements-ci.txt @@ -17,6 +17,7 @@ six git+https://github.com/lhotse-speech/lhotse kaldilm==1.11 kaldialign==0.7.1 +num2words sentencepiece==0.1.96 tensorboard==2.8.0 typeguard==2.13.3 diff --git a/requirements.txt b/requirements.txt index 5a8326619d..9502fcbd26 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,6 +1,7 @@ kaldifst kaldilm kaldialign +num2words kaldi-decoder sentencepiece>=0.1.96 tensorboard