From 8e296b7047bc0cfae3cdd41cf6223e356730ab2f Mon Sep 17 00:00:00 2001 From: Your Name Date: Fri, 9 Aug 2024 18:54:10 +0000 Subject: [PATCH 01/24] add librilight ssl recipe update Update ssl_datamodule.py Update pretrain.py Update pretrain.sh Update pretrain.sh Update hubert_ce.py Update pretrain.py --- egs/librilight/SSL/local/analyze_codebook.py | 88 ++ .../local/extract_kmeans_from_hubert_base.py | 251 +++ .../SSL/local/preprocess_librilight.py | 107 ++ egs/librilight/SSL/prepare.sh | 87 + egs/librilight/SSL/pretrain.sh | 22 + egs/librilight/SSL/shared | 1 + egs/librilight/SSL/zipformer/decode.py | 2 - egs/librilight/SSL/zipformer/finetune.py | 19 +- egs/librilight/SSL/zipformer/pretrain.py | 71 +- .../SSL/zipformer/ssl_datamodule.py | 159 +- egs/librispeech/SSL/hubert/asr_datamodule.py | 17 +- egs/librispeech/SSL/hubert/decode.py | 0 egs/librispeech/SSL/hubert/decode_ce.py | 0 egs/librispeech/SSL/hubert/finetune.py | 4 + egs/librispeech/SSL/hubert/finetune_ce.py | 4 + egs/librispeech/SSL/hubert/pretrain.py | 0 egs/librispeech/SSL/hubert/pretrain_ce.py | 0 egs/librispeech/SSL/hubert/ssl_datamodule.py | 11 + egs/librispeech/SSL/pretrain.sh | 19 + egs/librispeech/SSL/zipformer/decode.py | 0 egs/librispeech/SSL/zipformer/finetune.py | 4 + egs/librispeech/SSL/zipformer/hubert_ce.py | 1 - egs/librispeech/SSL/zipformer/model.py | 6 +- egs/librispeech/SSL/zipformer/pretrain.py | 7 +- .../SSL/zipformer_ctc/asr_datamodule.py | 1 + .../SSL/zipformer_ctc/beam_search.py | 1 + .../SSL/zipformer_ctc/ctc_decode.py | 823 ++++++++++ egs/librispeech/SSL/zipformer_ctc/dataset.py | 1 + .../SSL/zipformer_ctc/encoder_interface.py | 1 + .../SSL/zipformer_ctc/finetune_ctc.py | 1399 +++++++++++++++++ .../SSL/zipformer_ctc/hubert_ce.py | 1 + egs/librispeech/SSL/zipformer_ctc/model.py | 153 ++ egs/librispeech/SSL/zipformer_ctc/optim.py | 1 + egs/librispeech/SSL/zipformer_ctc/scaling.py | 1 + egs/librispeech/SSL/zipformer_ctc/utils.py | 1 + .../SSL/zipformer_ctc/wav2vec2_module.py | 1 + .../SSL/zipformer_ctc/zipformer.py | 1 + icefall/char_graph_compiler.py | 5 +- 38 files changed, 3139 insertions(+), 131 deletions(-) create mode 100755 egs/librilight/SSL/local/analyze_codebook.py create mode 100755 egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py create mode 100755 egs/librilight/SSL/local/preprocess_librilight.py create mode 100755 egs/librilight/SSL/prepare.sh create mode 100755 egs/librilight/SSL/pretrain.sh create mode 120000 egs/librilight/SSL/shared mode change 100644 => 100755 egs/librilight/SSL/zipformer/decode.py mode change 100644 => 100755 egs/librilight/SSL/zipformer/finetune.py mode change 100644 => 100755 egs/librilight/SSL/zipformer/pretrain.py mode change 100644 => 100755 egs/librispeech/SSL/hubert/decode.py mode change 100644 => 100755 egs/librispeech/SSL/hubert/decode_ce.py mode change 100644 => 100755 egs/librispeech/SSL/hubert/finetune.py mode change 100644 => 100755 egs/librispeech/SSL/hubert/finetune_ce.py mode change 100644 => 100755 egs/librispeech/SSL/hubert/pretrain.py mode change 100644 => 100755 egs/librispeech/SSL/hubert/pretrain_ce.py create mode 100755 egs/librispeech/SSL/pretrain.sh mode change 100644 => 100755 egs/librispeech/SSL/zipformer/decode.py mode change 100644 => 100755 egs/librispeech/SSL/zipformer/finetune.py mode change 100644 => 100755 egs/librispeech/SSL/zipformer/pretrain.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/asr_datamodule.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/beam_search.py create mode 100755 egs/librispeech/SSL/zipformer_ctc/ctc_decode.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/dataset.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/encoder_interface.py create mode 100755 egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/hubert_ce.py create mode 100644 egs/librispeech/SSL/zipformer_ctc/model.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/optim.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/scaling.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/utils.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/wav2vec2_module.py create mode 120000 egs/librispeech/SSL/zipformer_ctc/zipformer.py diff --git a/egs/librilight/SSL/local/analyze_codebook.py b/egs/librilight/SSL/local/analyze_codebook.py new file mode 100755 index 0000000000..80c61a75bc --- /dev/null +++ b/egs/librilight/SSL/local/analyze_codebook.py @@ -0,0 +1,88 @@ +#!/usr/bin/env python3 +# Copyright 2024 Xiaomi Corp. (authors: Yifan 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. + +import argparse +import logging +from collections import Counter +from pathlib import Path + +import torch +from lhotse import CutSet +from tqdm import tqdm + +# 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( + "--cuts-path", + type=str, + default="data/kmeans/librispeech_cuts_dev-clean.jsonl.gz", + ) + + parser.add_argument( + "--num-clusters", + type=int, + default=500, + ) + + return parser.parse_args() + + +def analyze_codebook(args): + cuts_path = Path(args.cuts_path) + assert cuts_path.is_file(), f"{cuts_path} does not exist" + + logging.info(f"Loading {cuts_path}") + cut_set = CutSet.from_file(cuts_path) + + cluster_counts = Counter() + logging.info("Analyzing codebook") + for cut in tqdm(cut_set): + kmeans = map(int, cut.custom["kmeans"].split()) + cluster_counts.update(kmeans) + + utilized_clusters = len(cluster_counts) + + total_count = sum(cluster_counts.values()) + counts = torch.tensor([cluster_counts[i] for i in range(args.num_clusters)]) + normalized_counts = (counts / total_count).clamp(min=1e-10) + codebook_entropy = ( + -(normalized_counts * normalized_counts.log()).sum() + * torch.log2(torch.tensor(torch.e)) + ).item() + + logging.info( + f"Codebook utilization rate: {utilized_clusters / args.num_clusters:%}" + ) + logging.info(f"Codebook entropy: {codebook_entropy}") + + +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)) + analyze_codebook(args) diff --git a/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py b/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py new file mode 100755 index 0000000000..e85a629d50 --- /dev/null +++ b/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py @@ -0,0 +1,251 @@ +#!/usr/bin/env python3 +# Copyright 2024 Xiaomi Corp. (authors: Yifan 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. + +import argparse +import logging +from pathlib import Path +from typing import Optional + +import fairseq +import joblib +import numpy as np +import torch +from lhotse import CutSet, SupervisionSegment +from lhotse.utils import fastcopy +from silero_vad import get_speech_timestamps, load_silero_vad +from tqdm import tqdm + +# 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) + + +class ApplyKmeans(object): + def __init__(self, km_path): + self.km_model = joblib.load(km_path) + self.C_np = self.km_model.cluster_centers_.transpose() + self.Cnorm_np = (self.C_np**2).sum(0, keepdims=True) + + self.C = torch.from_numpy(self.C_np) + self.Cnorm = torch.from_numpy(self.Cnorm_np) + if torch.cuda.is_available(): + self.C = self.C.cuda() + self.Cnorm = self.Cnorm.cuda() + + def __call__(self, x): + if isinstance(x, torch.Tensor): + dist = ( + x.pow(2).sum(1, keepdim=True) - 2 * torch.matmul(x, self.C) + self.Cnorm + ) + return dist.argmin(dim=1).cpu().numpy() + else: + dist = ( + (x**2).sum(1, keepdims=True) + - 2 * np.matmul(x, self.C_np) + + self.Cnorm_np + ) + return np.argmin(dist, axis=1) + + +def get_args(): + parser = argparse.ArgumentParser() + + parser.add_argument( + "--subset", + type=str, + default="small", + ) + + parser.add_argument( + "--model-path", + type=str, + default="download/hubert_base_ls960.pt", + ) + + parser.add_argument( + "--kmeans-model-path", + type=str, + default="download/hubert_base_ls960_L9_km500.model", + ) + + parser.add_argument( + "--start", + type=int, + default=0, + help="Process pieces starting from this number (inclusive).", + ) + + parser.add_argument( + "--stop", + type=int, + default=-1, + help="Stop processing pieces until this number (exclusive).", + ) + + return parser.parse_args() + + +def extract_and_save_one_cuts( + raw_cuts_path, cuts_path, model, vad_model, apply_kmeans, do_normalize, device +): + logging.info(f"Loading {raw_cuts_path}") + cut_set = CutSet.from_file(raw_cuts_path) + + logging.info("Extracting kmeans") + cuts = [] + for cut in tqdm(cut_set): + assert cut.sampling_rate == 16000, f"{cut.sampling_rate}" + audio = cut.load_audio() + + if audio.shape[-1] > 64 * 16000: + timestamps = get_speech_timestamps(audio, vad_model) + offsets = [i["start"] for i in timestamps] + audios = [audio[:, i["start"] : i["end"]] for i in timestamps] + logging.info(f"Trim audio {cut.id} into {len(audios)} segments") + else: + offsets = [0] + audios = [audio] + + seq = 0 + for audio, offset in zip(audios, offsets): + x = torch.from_numpy(audio).float().to(device) + + with torch.no_grad(): + if do_normalize: + x = torch.nn.functional.layer_norm(x, x.shape) + + feature, _ = model.extract_features( + source=x, + padding_mask=None, + mask=False, + output_layer=9, + ) + feature = feature.squeeze(0) + + kmeans = " ".join(map(str, apply_kmeans(feature).tolist())) + + supervision_segment = fastcopy( + cut.supervisions[0], + id=f"{cut.id}-{seq}", + start=0.0, + duration=audio.shape[-1] / 16000, + ) + cut_with_kmeans = fastcopy( + cut, + id=f"{cut.id}-{seq}", + start=cut.start + offset / 16000, + duration=audio.shape[-1] / 16000, + supervisions=[supervision_segment], + custom={"kmeans": kmeans}, + ) + cuts.append(cut_with_kmeans) + + seq += 1 + + cuts = CutSet(cuts) + + logging.info(f"Saving to {cuts_path}") + cuts.to_file(cuts_path) + + +def extract_kmeans(args): + assert args.subset in ("small", "medium", "large"), f"{args.subset}" + + output_dir = ( + f"data/kmeans/{args.subset}_split" if args.subset != "small" else "data/kmeans" + ) + output_dir = Path(output_dir) + assert output_dir.exists(), f"{output_dir} does not exist!" + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + logging.info(f"device: {device}") + + prefix = "librilight" + + vad_model = load_silero_vad() + apply_kmeans = ApplyKmeans(args.kmeans_model_path) + model, _, task = fairseq.checkpoint_utils.load_model_ensemble_and_task( + [args.model_path] + ) + model = model[0].eval().to(device) + do_normalize = task.cfg.normalize + + if args.subset == "small": + cuts_path = output_dir / f"{prefix}_cuts_{args.subset}.jsonl.gz" + if cuts_path.is_file(): + logging.info(f"{cuts_path} exists - skipping") + return + + raw_cuts_path = output_dir / f"{prefix}_cuts_{args.subset}_raw.jsonl.gz" + if not raw_cuts_path.is_file(): + logging.info(f"{raw_cuts_path} does not exist - skipping it") + return + + extract_and_save_one_cuts( + raw_cuts_path, + cuts_path, + model, + vad_model, + apply_kmeans, + do_normalize, + device, + ) + else: + num_digits = 8 # num_digits is fixed by lhotse split-lazy + start = args.start + stop = args.stop + assert stop > start, "stop must be larger than start!" + + for i in range(start, stop): + idx = f"{i}".zfill(num_digits) + logging.info(f"Processing {idx}/{stop - 1}") + + cuts_path = output_dir / f"{prefix}_cuts_{args.subset}.{idx}.jsonl.gz" + if cuts_path.is_file(): + logging.info(f"{cuts_path} exists - skipping") + continue + + raw_cuts_path = ( + output_dir / f"{prefix}_cuts_{args.subset}_raw.{idx}.jsonl.gz" + ) + if not raw_cuts_path.is_file(): + logging.info(f"{raw_cuts_path} does not exist - skipping it") + continue + + extract_and_save_one_cuts( + raw_cuts_path, + cuts_path, + model, + vad_model, + apply_kmeans, + do_normalize, + device, + ) + + +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)) + extract_kmeans(args) diff --git a/egs/librilight/SSL/local/preprocess_librilight.py b/egs/librilight/SSL/local/preprocess_librilight.py new file mode 100755 index 0000000000..4fb97cbde1 --- /dev/null +++ b/egs/librilight/SSL/local/preprocess_librilight.py @@ -0,0 +1,107 @@ +#!/usr/bin/env python3 +# Copyright 2024 Xiaomi Corp. (authors: Yifan 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. + +import argparse +import logging +import os +from pathlib import Path +from typing import Optional + +import torch +from lhotse import CutSet +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import 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( + "--dataset", + type=str, + help="""Dataset parts to compute fbank. If None, we will use all""", + ) + + return parser.parse_args() + + +def preprocess_librilight( + dataset: Optional[str] = None, +): + src_dir = Path("data/manifests") + output_dir = Path("data/kmeans") + + if dataset is None: + dataset_parts = ( + "small", + "medium", + "large", + ) + else: + dataset_parts = dataset.split(" ", -1) + + prefix = "librilight" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + for partition, m in manifests.items(): + cuts_filename = f"{prefix}_cuts_{partition}_raw.{suffix}" + if (output_dir / cuts_filename).is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + cut_set = cut_set.trim_to_supervisions( + keep_overlapping=False, min_duration=None + ) + logging.info(f"Saving to {output_dir / cuts_filename}") + cut_set.to_file(output_dir / cuts_filename) + + +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)) + preprocess_librilight( + dataset=args.dataset, + ) diff --git a/egs/librilight/SSL/prepare.sh b/egs/librilight/SSL/prepare.sh new file mode 100755 index 0000000000..804da9d6eb --- /dev/null +++ b/egs/librilight/SSL/prepare.sh @@ -0,0 +1,87 @@ +#!/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 +# run step 0 to step 5 by default +stage=0 +stop_stage=5 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/LibriLight +# - small +# - medium +# - large +# +# You can download them from +# - https://dl.fbaipublicfiles.com/librilight/data/small.tar +# - https://dl.fbaipublicfiles.com/librilight/data/medium.tar +# - https://dl.fbaipublicfiles.com/librilight/data/large.tar + +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "Running prepare.sh" + +log "dl_dir: $dl_dir" + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare Libri-Light manifest" + # We assume that you have downloaded the Libri-Light corpus + # to $dl_dir/LibriLight + mkdir -p data/manifests + if [ ! -e data/manifests/.librilight.done ]; then + lhotse prepare librilight -j $nj $dl_dir/LibriLight data/manifests + touch data/manifests/.librilight.done + fi +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Preprocess Libri-Light manifest" + mkdir -p data/kmeans + if [ ! -f data/kmeans/.preprocess_complete ]; then + python3 ./local/preprocess_librilight.py + touch data/fbank/.preprocess_complete + fi +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Split medium and large subset into pieces" + num_per_split=200000 + split_dir=data/kmeans/medium_split + if [ ! -f $split_dir/.split_completed ]; then + lhotse split-lazy ./data/kmeans/librilight_cuts_medium_raw.jsonl.gz $split_dir $num_per_split + touch $split_dir/.split_completed + fi + split_dir=data/kmeans/large_split + if [ ! -f $split_dir/.split_completed ]; then + lhotse split-lazy ./data/kmeans/librilight_cuts_large_raw.jsonl.gz $split_dir $num_per_split + touch $split_dir/.split_completed + fi +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Extract SSL target for librilight" + mkdir -p data/fbank + if [ ! -e data/fbank/.librispeech.done ]; then + ./local/compute_fbank_librispeech.py + touch data/fbank/.librispeech.done + fi +fi diff --git a/egs/librilight/SSL/pretrain.sh b/egs/librilight/SSL/pretrain.sh new file mode 100755 index 0000000000..c282967f9a --- /dev/null +++ b/egs/librilight/SSL/pretrain.sh @@ -0,0 +1,22 @@ +export PYTHONPATH=$(pwd)/../../.. + +./zipformer/pretrain.py \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp_pretrain \ + --max-duration 650 \ + --quadratic-duration 512 \ + --accum-grad 1 \ + --do-normalize 1 \ + --mask-prob 0.8 \ + --extractor-mode "layer_norm" \ + --dropout-input 0.0 \ + --dropout-features 0.0 \ + --feature-grad-mult 1.0 \ + --num-encoder-layers 2,2,3,4,3,2 \ + --feedforward-dim 512,768,1024,1536,1024,768 \ + --encoder-dim 192,256,448,768,448,192 \ + --encoder-unmasked-dim 192,192,256,256,256,192 \ + --base-lr 0.045 diff --git a/egs/librilight/SSL/shared b/egs/librilight/SSL/shared new file mode 120000 index 0000000000..4c5e91438c --- /dev/null +++ b/egs/librilight/SSL/shared @@ -0,0 +1 @@ +../../../icefall/shared/ \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/decode.py b/egs/librilight/SSL/zipformer/decode.py old mode 100644 new mode 100755 index 95643c5e14..1562c28b8b --- a/egs/librilight/SSL/zipformer/decode.py +++ b/egs/librilight/SSL/zipformer/decode.py @@ -1015,8 +1015,6 @@ def main(): test_sets = ["dev-clean", "dev-other", "test-clean", "test-other"] test_dl = [dev_clean_dl, dev_other_dl, test_clean_dl, test_other_dl] - # test_sets = ["dev-clean", "dev-other"] - # test_dl = [dev_clean_dl, dev_other_dl] for test_set, test_dl in zip(test_sets, test_dl): results_dict = decode_dataset( diff --git a/egs/librilight/SSL/zipformer/finetune.py b/egs/librilight/SSL/zipformer/finetune.py old mode 100644 new mode 100755 index 50dbd5f2d7..2e521f1772 --- a/egs/librilight/SSL/zipformer/finetune.py +++ b/egs/librilight/SSL/zipformer/finetune.py @@ -1,11 +1,10 @@ #!/usr/bin/env python3 -# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, -# Wei Kang, -# Mingshuang Luo, -# Zengwei Yao, -# Yifan Yang, -# Daniel Povey) -# +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) # Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) # # See ../../../../LICENSE for clarification regarding multiple authors @@ -1246,7 +1245,7 @@ def save_bad_model(suffix: str = ""): tb_writer, "train/valid_", params.batch_idx_train ) - if batch_idx % params.accum_grad != params.accum_grad - 1: + if sub_batch_idx % params.accum_grad != params.accum_grad - 1: optimizer.zero_grad() loss_value = tot_loss["loss"] / tot_loss["frames"] params.train_loss = loss_value @@ -1388,6 +1387,8 @@ def remove_short_and_long_utt(c: Cut): train_cuts, do_normalize=params.do_normalize, sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, ) valid_cuts = librispeech.dev_clean_cuts() @@ -1396,6 +1397,8 @@ def remove_short_and_long_utt(c: Cut): valid_dl = librispeech.valid_dataloaders( valid_cuts, do_normalize=params.do_normalize, + world_size=world_size, + rank=rank, ) if params.sanity_check and not params.print_diagnostics: diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py old mode 100644 new mode 100755 index 5728dbe752..251184b69f --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -1,11 +1,10 @@ #!/usr/bin/env python3 -# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, -# Wei Kang, -# Mingshuang Luo, -# Zengwei Yao, -# Yifan Yang, -# Daniel Povey) -# +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) # Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) # # See ../../../../LICENSE for clarification regarding multiple authors @@ -32,7 +31,8 @@ --num-epochs 400 \ --start-epoch 1 \ --use-fp16 1 \ - --exp-dir zipformer/exp \ + --exp-dir hubert/exp \ + --full-libri 1 \ --max-duration 87.5 \ --accum-grad 4 """ @@ -46,6 +46,7 @@ from shutil import copyfile from typing import Any, Dict, Optional, Tuple, Union +import lhotse import optim import torch import torch.multiprocessing as mp @@ -398,13 +399,6 @@ def add_model_arguments(parser: argparse.ArgumentParser): and the value should be the multiple of 4, for faster computation""", ) - parser.add_argument( - "--untie-final-proj", - type=bool, - default=False, - help="use separate projection for each target", - ) - def get_parser(): parser = argparse.ArgumentParser( @@ -483,7 +477,7 @@ def get_parser(): parser.add_argument( "--lr-epochs", type=float, - default=10.5, + default=0.2, help="""Number of epochs that affects how rapidly the learning rate decreases. """, ) @@ -541,7 +535,7 @@ def get_parser(): parser.add_argument( "--save-every-n", type=int, - default=100000, + default=10000, 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 @@ -554,7 +548,7 @@ def get_parser(): parser.add_argument( "--keep-last-k", type=int, - default=30, + default=100000, 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`. @@ -591,17 +585,24 @@ def get_parser(): ) parser.add_argument( - "--max-sample-size", - type=float, - default=250000, - help="max sample size", + "--max-keep-size", + type=int, + default=1024000, + help="exclude sample longer than this.", ) parser.add_argument( - "--min-sample-size", + "--min-keep-size", type=float, default=32000, - help="min sample size", + help="exclude sample longer less than this.", + ) + + parser.add_argument( + "--max-sample-size", + type=float, + default=1024000, + help="max sample size to crop to for batching.", ) add_model_arguments(parser) @@ -960,10 +961,10 @@ def save_bad_model(suffix: str = ""): else: continue - except: # noqa + except Exception as e: # noqa save_bad_model() display_and_save_batch(batch, params=params) - raise + raise e if params.print_diagnostics and batch_idx == 5: return @@ -1064,7 +1065,7 @@ def save_bad_model(suffix: str = ""): tb_writer, "train/valid_", params.batch_idx_train ) - if batch_idx % params.accum_grad != params.accum_grad - 1: + if sub_batch_idx % params.accum_grad != params.accum_grad - 1: optimizer.zero_grad() loss_value = tot_loss["loss"] / tot_loss["frames"] params.train_loss = loss_value @@ -1165,7 +1166,7 @@ def run(rank, world_size, args): librilight = LibriLightDataModule(args) - train_cuts = librilight.train_all_shuf_cuts() + train_cuts = librilight.all_shuf_cuts() def remove_short_and_long_utt(c: Cut): # Keep only utterances with duration between 1 second and 20 seconds @@ -1177,11 +1178,11 @@ def remove_short_and_long_utt(c: Cut): # an utterance duration distribution for your dataset to select # the threshold if ( - c.duration < params.min_sample_size / params.sample_rate - or c.duration > params.max_sample_size / params.sample_rate + c.duration < params.min_keep_size / params.sample_rate + or c.duration > params.max_keep_size / params.sample_rate ): # logging.warning( - # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" # ) return False @@ -1198,6 +1199,7 @@ def remove_short_and_long_utt(c: Cut): train_dl = librilight.train_dataloaders( train_cuts, + max_sample_size=params.max_sample_size, sample_rate=params.sample_rate, label_rate=params.label_rate, random_crop=params.random_crop, @@ -1205,6 +1207,8 @@ def remove_short_and_long_utt(c: Cut): num_classes=params.num_classes, do_normalize=params.do_normalize, sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, ) valid_cuts = librilight.dev_clean_cuts() @@ -1213,12 +1217,15 @@ def remove_short_and_long_utt(c: Cut): valid_dl = librilight.valid_dataloaders( valid_cuts, + max_sample_size=params.max_sample_size, sample_rate=params.sample_rate, label_rate=params.label_rate, random_crop=params.random_crop, pad_audio=False, num_classes=params.num_classes, do_normalize=params.do_normalize, + world_size=world_size, + rank=rank, ) if params.sanity_check and not params.print_diagnostics: @@ -1339,7 +1346,7 @@ def scan_pessimistic_batches_for_oom( f"(={crit_values[criterion]}) ..." ) display_and_save_batch(batch, params=params) - raise + raise e logging.info( f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" ) diff --git a/egs/librilight/SSL/zipformer/ssl_datamodule.py b/egs/librilight/SSL/zipformer/ssl_datamodule.py index dc0dbec6ca..3cf7c2f73c 100644 --- a/egs/librilight/SSL/zipformer/ssl_datamodule.py +++ b/egs/librilight/SSL/zipformer/ssl_datamodule.py @@ -1,5 +1,4 @@ -# Copyright 2021 Piotr Żelasko -# Copyright 2023 Xiaomi Corporation (Author: Yifan Yang) +# Copyright 2024 Xiaomi Corporation (Author: Yifan Yang) # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -25,8 +24,9 @@ from typing import Any, Dict, Optional import torch +import lhotse from dataset import HubertDataset -from lhotse import CutSet, combine, load_manifest_lazy +from lhotse import CutSet, load_manifest_lazy from lhotse.dataset import DynamicBucketingSampler, SimpleCutSampler from lhotse.utils import fix_random_seed from torch.utils.data import DataLoader @@ -46,7 +46,7 @@ class LibriLightDataModule: """ DataModule for SSL experiments. It assumes there is always one train and valid dataloader, - but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + but there can be multiple test dataloaders (e.g. LibriLight test-clean and test-other). It contains all the common data pipeline modules used in SSL @@ -63,7 +63,7 @@ def __init__(self, args: argparse.Namespace): @classmethod def add_arguments(cls, parser: argparse.ArgumentParser): group = parser.add_argument_group( - title="ASR SSL related options", + title="SSL 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.", @@ -92,10 +92,29 @@ def add_arguments(cls, parser: argparse.ArgumentParser): group.add_argument( "--num-buckets", type=int, - default=30, + default=1000, help="The number of buckets for the DynamicBucketingSampler" "(you might want to increase it for larger datasets).", ) + group.add_argument( + "--num-cuts-for-bins-estimate", + type=float, + default=1000000, + help="We will draw this many cuts to estimate the duration" + "bins for creating similar-duration buckets. Larger number" + "means a better estimate to the data distribution, possibly" + "at a longer init cost." + ) + group.add_argument( + "--quadratic-duration", + type=float, + default=None, + help="When set, it adds an extra penalty that's quadratic" + "in size w.r.t. a cuts duration. This helps get a more" + "even GPU utilization across different input lengths when" + "models have quadratic input complexity. Set between 15" + "and 40 for transformers.", + ) group.add_argument( "--shuffle", type=str2bool, @@ -112,7 +131,7 @@ def add_arguments(cls, parser: argparse.ArgumentParser): group.add_argument( "--num-workers", type=int, - default=2, + default=8, help="The number of training dataloader workers that " "collect the batches.", ) @@ -126,12 +145,13 @@ def add_arguments(cls, parser: argparse.ArgumentParser): "--random-crop", type=str2bool, default=True, - help="audio sample rate", + help="always crop from the beginning if false", ) def train_dataloaders( self, cuts_train: CutSet, + max_sample_size: Optional[int] = None, sample_rate: float = 16000, label_rate: float = 50, random_crop: bool = True, @@ -139,6 +159,8 @@ def train_dataloaders( num_classes: list = [504], do_normalize: bool = True, sampler_state_dict: Optional[Dict[str, Any]] = None, + world_size: Optional[int] = None, + rank: Optional[int] = None, ) -> DataLoader: """ Args: @@ -149,6 +171,7 @@ def train_dataloaders( """ logging.info("About to create train dataset") train = HubertDataset( + max_sample_size=max_sample_size, sample_rate=sample_rate, label_rate=label_rate, random_crop=random_crop, @@ -162,9 +185,14 @@ def train_dataloaders( train_sampler = DynamicBucketingSampler( cuts_train, max_duration=self.args.max_duration, + quadratic_duration=self.args.quadratic_duration, shuffle=self.args.shuffle, num_buckets=self.args.num_buckets, + buffer_size=self.args.num_buckets * 2000, + num_cuts_for_bins_estimate=self.args.num_cuts_for_bins_estimate, drop_last=self.args.drop_last, + world_size=world_size, + rank=rank, ) else: logging.info("Using SimpleCutSampler.") @@ -172,6 +200,8 @@ def train_dataloaders( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, + world_size=world_size, + rank=rank, ) logging.info("About to create train dataloader") @@ -198,15 +228,19 @@ def train_dataloaders( def valid_dataloaders( self, cuts_valid: CutSet, + max_sample_size: Optional[int] = None, sample_rate: float = 16000, label_rate: float = 50, random_crop: bool = True, pad_audio: bool = False, num_classes: list = [504], do_normalize: bool = True, + world_size: Optional[int] = None, + rank: Optional[int] = None, ) -> DataLoader: logging.info("About to create dev dataset") validate = HubertDataset( + max_sample_size=max_sample_size, sample_rate=sample_rate, label_rate=label_rate, random_crop=random_crop, @@ -217,7 +251,10 @@ def valid_dataloaders( valid_sampler = DynamicBucketingSampler( cuts_valid, max_duration=self.args.max_duration, + quadratic_duration=self.args.quadratic_duration, shuffle=False, + world_size=world_size, + rank=rank, ) logging.info("About to create dev dataloader") valid_dl = DataLoader( @@ -230,105 +267,65 @@ def valid_dataloaders( return valid_dl - def test_dataloaders( - self, - cuts: CutSet, - sample_rate: float = 16000, - label_rate: float = 50, - random_crop: bool = True, - pad_audio: bool = False, - num_classes: list = [504], - do_normalize: bool = True, - ) -> DataLoader: - logging.debug("About to create test dataset") - test = HubertDataset( - sample_rate=sample_rate, - label_rate=label_rate, - random_crop=random_crop, - pad_audio=pad_audio, - num_classes=num_classes, - do_normalize=do_normalize, + @lru_cache() + def all_shuf_cuts(self) -> CutSet: + logging.info( + "About to get the shuffled librilight small, medium and large cuts" ) - sampler = DynamicBucketingSampler( - cuts, - max_duration=self.args.max_duration, - shuffle=False, + small_cuts = self.small_cuts() + medium_cuts = self.medium_cuts() + large_cuts = self.large_cuts() + return CutSet.mux( + small_cuts, + medium_cuts, + large_cuts, + weights=[ + 229051, # len(small_cuts) + 2022949, # len(medium_cuts) + 19883414, # len(large_cuts) + ], ) - logging.debug("About to create test dataloader") - test_dl = DataLoader( - test, - batch_size=None, - sampler=sampler, - num_workers=self.args.num_workers, + + @lru_cache() + def dev_clean_cuts(self) -> CutSet: + logging.info("About to get librispeech dev-clean cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz" ) - return test_dl @lru_cache() def small_cuts(self) -> CutSet: - logging.info("About to get small cuts") + logging.info("About to get librilight small cuts") return load_manifest_lazy( self.args.manifest_dir / "librilight_cuts_small.jsonl.gz" ) @lru_cache() def medium_cuts(self) -> CutSet: - logging.info("About to get medium cuts") + logging.info("About to get librilight medium cuts") filenames = glob.glob( - f"{self.args.manifest_dir}/medium_splits/librilight_cuts_medium.*.jsonl.gz" + str(self.args.manifest_dir / "medium_split" / "librilight_cuts_medium.*.jsonl.gz") ) pattern = re.compile(r"librilight_cuts_medium.([0-9]+).jsonl.gz") idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames) idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) sorted_filenames = [f[1] for f in idx_filenames] - logging.info( - f"Loading LibriLight medium {len(sorted_filenames)} splits in lazy mode" + logging.info(f"Loading Libri-Light medium {len(sorted_filenames)} splits in lazy mode") + return lhotse.combine( + lhotse.load_manifest_lazy(p) for p in sorted_filenames ) - return combine(load_manifest_lazy(p) for p in sorted_filenames) - @lru_cache() def large_cuts(self) -> CutSet: - logging.info("About to get large cuts") + logging.info("About to get librilight large cuts") filenames = glob.glob( - f"{self.args.manifest_dir}/large_splits/librilight_cuts_large.*.jsonl.gz" + str(self.args.manifest_dir / "large_split" / "librilight_cuts_large.*.jsonl.gz") ) pattern = re.compile(r"librilight_cuts_large.([0-9]+).jsonl.gz") idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames) idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) sorted_filenames = [f[1] for f in idx_filenames] - logging.info( - f"Loading LibriLight large {len(sorted_filenames)} splits in lazy mode" - ) - - return combine(load_manifest_lazy(p) for p in sorted_filenames) - - @lru_cache() - def train_all_shuf_cuts(self) -> CutSet: - logging.info("About to get the shuffled small, medium and large cuts") - small_cuts = self.small_cuts() - medium_cuts = self.medium_cuts() - large_cuts = self.large_cuts() - return CutSet.mux( - small_cuts, - medium_cuts, - large_cuts, - weights=[ - 122867, # len(small_cuts) - 1104071, # len(medium_cuts) - 11012085, # len(large_cuts) - ], - ) - - @lru_cache() - def dev_clean_cuts(self) -> CutSet: - logging.info("About to get dev-clean cuts") - return load_manifest_lazy( - self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz" - ) - - @lru_cache() - def dev_other_cuts(self) -> CutSet: - logging.info("About to get dev-other cuts") - return load_manifest_lazy( - self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz" + logging.info(f"Loading Libri-Light large {len(sorted_filenames)} splits in lazy mode") + return lhotse.combine( + lhotse.load_manifest_lazy(p) for p in sorted_filenames ) diff --git a/egs/librispeech/SSL/hubert/asr_datamodule.py b/egs/librispeech/SSL/hubert/asr_datamodule.py index 3746d8a3a3..60ecba8d76 100644 --- a/egs/librispeech/SSL/hubert/asr_datamodule.py +++ b/egs/librispeech/SSL/hubert/asr_datamodule.py @@ -132,6 +132,8 @@ def train_dataloaders( cuts_train: CutSet, do_normalize: bool, sampler_state_dict: Optional[Dict[str, Any]] = None, + world_size: Optional[int] = None, + rank: Optional[int] = None, ) -> DataLoader: """ Args: @@ -150,7 +152,10 @@ def train_dataloaders( max_duration=self.args.max_duration, shuffle=self.args.shuffle, num_buckets=self.args.num_buckets, + buffer_size=self.args.num_buckets * 2000, drop_last=self.args.drop_last, + world_size=world_size, + rank=rank, ) else: logging.info("Using SimpleCutSampler.") @@ -158,6 +163,8 @@ def train_dataloaders( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, + world_size=world_size, + rank=rank, ) logging.info("About to create train dataloader") @@ -181,13 +188,21 @@ def train_dataloaders( return train_dl - def valid_dataloaders(self, cuts_valid: CutSet, do_normalize: bool) -> DataLoader: + def valid_dataloaders( + self, + cuts_valid: CutSet, + do_normalize: bool, + world_size: Optional[int] = None, + rank: Optional[int] = None, + ) -> DataLoader: logging.info("About to create dev dataset") validate = HubertAsrDataset(do_normalize=do_normalize) valid_sampler = DynamicBucketingSampler( cuts_valid, max_duration=self.args.max_duration, shuffle=False, + world_size=world_size, + rank=rank, ) logging.info("About to create dev dataloader") valid_dl = DataLoader( diff --git a/egs/librispeech/SSL/hubert/decode.py b/egs/librispeech/SSL/hubert/decode.py old mode 100644 new mode 100755 diff --git a/egs/librispeech/SSL/hubert/decode_ce.py b/egs/librispeech/SSL/hubert/decode_ce.py old mode 100644 new mode 100755 diff --git a/egs/librispeech/SSL/hubert/finetune.py b/egs/librispeech/SSL/hubert/finetune.py old mode 100644 new mode 100755 index 17daa3c9d3..05b942f632 --- a/egs/librispeech/SSL/hubert/finetune.py +++ b/egs/librispeech/SSL/hubert/finetune.py @@ -1090,6 +1090,8 @@ def remove_short_and_long_utt(c: Cut): train_cuts, do_normalize=params.do_normalize, sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, ) valid_cuts = librispeech.dev_clean_cuts() @@ -1098,6 +1100,8 @@ def remove_short_and_long_utt(c: Cut): valid_dl = librispeech.valid_dataloaders( valid_cuts, do_normalize=params.do_normalize, + world_size=world_size, + rank=rank, ) if params.sanity_check and not params.print_diagnostics: diff --git a/egs/librispeech/SSL/hubert/finetune_ce.py b/egs/librispeech/SSL/hubert/finetune_ce.py old mode 100644 new mode 100755 index 2723cc770e..1081313f1d --- a/egs/librispeech/SSL/hubert/finetune_ce.py +++ b/egs/librispeech/SSL/hubert/finetune_ce.py @@ -1090,6 +1090,8 @@ def remove_short_and_long_utt(c: Cut): train_cuts, do_normalize=params.do_normalize, sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, ) valid_cuts = librispeech.dev_clean_cuts() @@ -1098,6 +1100,8 @@ def remove_short_and_long_utt(c: Cut): valid_dl = librispeech.valid_dataloaders( valid_cuts, do_normalize=params.do_normalize, + world_size=world_size, + rank=rank, ) if params.sanity_check and not params.print_diagnostics: diff --git a/egs/librispeech/SSL/hubert/pretrain.py b/egs/librispeech/SSL/hubert/pretrain.py old mode 100644 new mode 100755 diff --git a/egs/librispeech/SSL/hubert/pretrain_ce.py b/egs/librispeech/SSL/hubert/pretrain_ce.py old mode 100644 new mode 100755 diff --git a/egs/librispeech/SSL/hubert/ssl_datamodule.py b/egs/librispeech/SSL/hubert/ssl_datamodule.py index ac1a0997da..55f06c6206 100644 --- a/egs/librispeech/SSL/hubert/ssl_datamodule.py +++ b/egs/librispeech/SSL/hubert/ssl_datamodule.py @@ -144,6 +144,8 @@ def train_dataloaders( num_classes: list = [504], do_normalize: bool = True, sampler_state_dict: Optional[Dict[str, Any]] = None, + world_size: Optional[int] = None, + rank: Optional[int] = None, ) -> DataLoader: """ Args: @@ -170,7 +172,10 @@ def train_dataloaders( max_duration=self.args.max_duration, shuffle=self.args.shuffle, num_buckets=self.args.num_buckets, + buffer_size=self.args.num_buckets * 2000, drop_last=self.args.drop_last, + world_size=world_size, + rank=rank, ) else: logging.info("Using SimpleCutSampler.") @@ -178,6 +183,8 @@ def train_dataloaders( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, + world_size=world_size, + rank=rank, ) logging.info("About to create train dataloader") @@ -211,6 +218,8 @@ def valid_dataloaders( pad_audio: bool = False, num_classes: list = [504], do_normalize: bool = True, + world_size: Optional[int] = None, + rank: Optional[int] = None, ) -> DataLoader: logging.info("About to create dev dataset") validate = HubertDataset( @@ -226,6 +235,8 @@ def valid_dataloaders( cuts_valid, max_duration=self.args.max_duration, shuffle=False, + world_size=world_size, + rank=rank, ) logging.info("About to create dev dataloader") valid_dl = DataLoader( diff --git a/egs/librispeech/SSL/pretrain.sh b/egs/librispeech/SSL/pretrain.sh new file mode 100755 index 0000000000..70401aeb82 --- /dev/null +++ b/egs/librispeech/SSL/pretrain.sh @@ -0,0 +1,19 @@ +./zipformer/pretrain.py \ + --world-size 8 \ + --num-epochs 300 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp_pretrain \ + --full-libri 1 \ + --max-duration 600 \ + --accum-grad 1 \ + --do-normalize 0 \ + --mask-prob 0.8 \ + --dropout-input 0.1 \ + --dropout-features 0.1 \ + --feature-grad-mult 0.1 \ + --num-encoder-layers 2,2,3,4,3,2 \ + --feedforward-dim 512,768,1024,1536,1024,768 \ + --encoder-dim 192,256,448,768,448,192 \ + --encoder-unmasked-dim 192,192,256,256,256,192 \ + --base-lr 0.045 diff --git a/egs/librispeech/SSL/zipformer/decode.py b/egs/librispeech/SSL/zipformer/decode.py old mode 100644 new mode 100755 diff --git a/egs/librispeech/SSL/zipformer/finetune.py b/egs/librispeech/SSL/zipformer/finetune.py old mode 100644 new mode 100755 index c907b41c55..2e521f1772 --- a/egs/librispeech/SSL/zipformer/finetune.py +++ b/egs/librispeech/SSL/zipformer/finetune.py @@ -1387,6 +1387,8 @@ def remove_short_and_long_utt(c: Cut): train_cuts, do_normalize=params.do_normalize, sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, ) valid_cuts = librispeech.dev_clean_cuts() @@ -1395,6 +1397,8 @@ def remove_short_and_long_utt(c: Cut): valid_dl = librispeech.valid_dataloaders( valid_cuts, do_normalize=params.do_normalize, + world_size=world_size, + rank=rank, ) if params.sanity_check and not params.print_diagnostics: diff --git a/egs/librispeech/SSL/zipformer/hubert_ce.py b/egs/librispeech/SSL/zipformer/hubert_ce.py index 1ac368a1d0..43b41c3af3 100644 --- a/egs/librispeech/SSL/zipformer/hubert_ce.py +++ b/egs/librispeech/SSL/zipformer/hubert_ce.py @@ -296,7 +296,6 @@ def __init__( self.layer_norm = LayerNorm(self.embed) - self.untie_final_proj = cfg.untie_final_proj self.final_proj = nn.Linear(encoder_output_dim, sum(cfg.num_classes)) # modules below are not needed during fine-tuning diff --git a/egs/librispeech/SSL/zipformer/model.py b/egs/librispeech/SSL/zipformer/model.py index 46a968b69e..948e811c98 100644 --- a/egs/librispeech/SSL/zipformer/model.py +++ b/egs/librispeech/SSL/zipformer/model.py @@ -154,9 +154,9 @@ def forward_ctc( ctc_loss = torch.nn.functional.ctc_loss( log_probs=ctc_output.permute(1, 0, 2), # (T, N, C) - targets=targets, - input_lengths=encoder_out_lens, - target_lengths=target_lengths, + targets=targets.cpu(), + input_lengths=encoder_out_lens.cpu(), + target_lengths=target_lengths.cpu(), reduction="sum", ) return ctc_loss diff --git a/egs/librispeech/SSL/zipformer/pretrain.py b/egs/librispeech/SSL/zipformer/pretrain.py old mode 100644 new mode 100755 index 937fb382ed..49c5921a30 --- a/egs/librispeech/SSL/zipformer/pretrain.py +++ b/egs/librispeech/SSL/zipformer/pretrain.py @@ -41,7 +41,6 @@ import argparse import copy import logging -import sys import warnings from pathlib import Path from shutil import copyfile @@ -594,7 +593,7 @@ def get_parser(): parser.add_argument( "--max-keep-size", type=int, - default=sys.maxsize, + default=320000, help="exclude sample longer than this.", ) @@ -1218,6 +1217,8 @@ def remove_short_and_long_utt(c: Cut): num_classes=params.num_classes, do_normalize=params.do_normalize, sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, ) valid_cuts = librispeech.dev_clean_cuts() @@ -1233,6 +1234,8 @@ def remove_short_and_long_utt(c: Cut): pad_audio=False, num_classes=params.num_classes, do_normalize=params.do_normalize, + world_size=world_size, + rank=rank, ) if params.sanity_check and not params.print_diagnostics: diff --git a/egs/librispeech/SSL/zipformer_ctc/asr_datamodule.py b/egs/librispeech/SSL/zipformer_ctc/asr_datamodule.py new file mode 120000 index 0000000000..3c8b7f2d45 --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/asr_datamodule.py @@ -0,0 +1 @@ +../zipformer/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/beam_search.py b/egs/librispeech/SSL/zipformer_ctc/beam_search.py new file mode 120000 index 0000000000..586c34b2f5 --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/beam_search.py @@ -0,0 +1 @@ +../zipformer/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/ctc_decode.py b/egs/librispeech/SSL/zipformer_ctc/ctc_decode.py new file mode 100755 index 0000000000..68a4772e2a --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/ctc_decode.py @@ -0,0 +1,823 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Liyong Guo, +# Quandong Wang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +(1) ctc-decoding +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method ctc-decoding + +(2) 1best +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --decoding-method 1best + +(3) nbest +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --decoding-method nbest + +(4) nbest-rescoring +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --nbest-scale 1.0 \ + --lm-dir data/lm \ + --decoding-method nbest-rescoring + +(5) whole-lattice-rescoring +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --nbest-scale 1.0 \ + --lm-dir data/lm \ + --decoding-method whole-lattice-rescoring +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from finetune_ctc import add_model_arguments, get_model, get_params + +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import ( + get_lattice, + nbest_decoding, + nbest_oracle, + one_best_decoding, + rescore_with_n_best_list, + rescore_with_whole_lattice, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + get_texts, + 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( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="ctc-decoding", + help="""Decoding method. + Supported values are: + - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece + model, i.e., lang_dir/bpe.model, to convert word pieces to words. + It needs neither a lexicon nor an n-gram LM. + - (2) 1best. Extract the best path from the decoding lattice as the + decoding result. + - (3) nbest. Extract n paths from the decoding lattice; the path + with the highest score is the decoding result. + - (4) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an n-gram LM (e.g., a 4-gram LM), the path with + the highest score is the decoding result. + - (5) whole-lattice-rescoring. Rescore the decoding lattice with an + n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice + is the decoding result. + you have trained an RNN LM using ./rnn_lm/train.py + - (6) nbest-oracle. Its WER is the lower bound of any n-best + rescoring method can achieve. Useful for debugging n-best + rescoring method. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for n-best based decoding method. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=1.0, + help="""The scale to be applied to `lattice.scores`. + It's needed if you use any kinds of n-best based rescoring. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + A smaller value results in more unique paths. + """, + ) + + parser.add_argument( + "--hlg-scale", + type=float, + default=0.6, + help="""The scale to be applied to `hlg.scores`. + """, + ) + + parser.add_argument( + "--lm-dir", + type=str, + default="data/lm", + help="""The n-gram LM dir. + It should contain either G_4_gram.pt or G_4_gram.fst.txt + """, + ) + + add_model_arguments(parser) + + return parser + + +def get_decoding_params() -> AttributeDict: + """Parameters for decoding.""" + params = AttributeDict( + { + "frame_shift_ms": 10, + "search_beam": 20, + "output_beam": 8, + "min_active_states": 30, + "max_active_states": 10000, + "use_double_scores": True, + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + lexicon: Lexicon, + graph_compiler: CharCtcTrainingGraphCompiler, + batch: dict, + word_table: k2.SymbolTable, + G: 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 no rescoring is used, the key is the string `no_rescore`. + If LM rescoring is used, the key is the string `lm_scale_xxx`, + where `xxx` is the value of `lm_scale`. An example key is + `lm_scale_0.7` + - 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`. + + - params.decoding_method is "1best", it uses 1best decoding without LM rescoring. + - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring. + - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring. + - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM + rescoring. + + model: + The neural model. + HLG: + The decoding graph. Used only when params.decoding_method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return the decoding result. See above description for the format of + the returned dict. Note: If it decodes to nothing, then return None. + """ + if HLG is not None: + device = HLG.device + else: + device = H.device + + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + encoder_out, encoder_out_lens = model.forward_encoder(audio, padding_mask) + ctc_output = model.ctc_output(encoder_out) + + num_frames = encoder_out_lens.cpu() + + supervision_segments = torch.stack( + ( + torch.arange(audio.shape[0], dtype=torch.int32), + torch.zeros_like(num_frames, dtype=torch.int32), + num_frames, + ), + 1, + ).to(torch.int32) + + if H is None: + assert HLG is not None + decoding_graph = HLG + else: + assert HLG is None + decoding_graph = H + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=decoding_graph, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + ) + + if params.decoding_method == "ctc-decoding": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + # Note: `best_path.aux_labels` contains token IDs, not word IDs + # since we are using H, not HLG here. + # + # token_ids is a lit-of-list of IDs + token_ids = get_texts(best_path) + + # hyps is a list of str, e.g., ['xxx yyy zzz', ...] + hyps = [ + "".join(lexicon.token_table[idx].replace("|", " ") for idx in token_id) + for token_id in token_ids + ] + + # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] + hyps = [s.split() for s in hyps] + key = "ctc-decoding" + return {key: hyps} + + if params.decoding_method == "nbest-oracle": + # Note: You can also pass rescored lattices to it. + # We choose the HLG decoded lattice for speed reasons + # as HLG decoding is faster and the oracle WER + # is only slightly worse than that of rescored lattices. + best_path = nbest_oracle( + lattice=lattice, + num_paths=params.num_paths, + ref_texts=batch["supervisions"]["text"], + word_table=word_table, + nbest_scale=params.nbest_scale, + oov="", + ) + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa + return {key: hyps} + + if params.decoding_method in ["1best", "nbest"]: + if params.decoding_method == "1best": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + key = "no_rescore" + else: + best_path = nbest_decoding( + lattice=lattice, + num_paths=params.num_paths, + use_double_scores=params.use_double_scores, + nbest_scale=params.nbest_scale, + ) + key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa + + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + return {key: hyps} + + assert params.decoding_method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + ] + + lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7] + lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3] + lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0] + + if params.decoding_method == "nbest-rescoring": + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=lm_scale_list, + nbest_scale=params.nbest_scale, + ) + elif params.decoding_method == "whole-lattice-rescoring": + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=lm_scale_list, + ) + else: + assert False, f"Unsupported decoding method: {params.decoding_method}" + + ans = dict() + if best_path_dict is not None: + for lm_scale_str, best_path in best_path_dict.items(): + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + ans[lm_scale_str] = hyps + else: + ans = None + return ans + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + lexicon: Lexicon, + graph_compiler: CharCtcTrainingGraphCompiler, + word_table: k2.SymbolTable, + G: 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. + HLG: + The decoding graph. Used only when params.decoding_method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + word_table: + It is the word symbol table. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return a dict, whose key may be "no-rescore" if no LM rescoring + is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["cuts"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + HLG=HLG, + H=H, + lexicon=lexicon, + graph_compiler=graph_compiler, + batch=batch, + word_table=word_table, + G=G, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % 100 == 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}-{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}-{params.suffix}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}-{key}", results) + 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}-{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() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + args.lm_dir = Path(args.lm_dir) + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + assert params.decoding_method in ( + "ctc-decoding", + "1best", + "nbest", + "nbest-rescoring", + "whole-lattice-rescoring", + "nbest-oracle", + ) + 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.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}") + logging.info(params) + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + if params.decoding_method == "ctc-decoding": + HLG = None + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + else: + H = None + HLG = k2.Fsa.from_dict( + torch.load(f"{params.lang_dir}/HLG.pt", map_location=device) + ) + assert HLG.requires_grad is False + + HLG.scores *= params.hlg_scale + if not hasattr(HLG, "lm_scores"): + HLG.lm_scores = HLG.scores.clone() + + if params.decoding_method in ( + "nbest-rescoring", + "whole-lattice-rescoring", + ): + if not (params.lm_dir / "G_4_gram.pt").is_file(): + logging.info("Loading G_4_gram.fst.txt") + logging.warning("It may take 8 minutes.") + with open(params.lm_dir / "G_4_gram.fst.txt") as f: + first_word_disambig_id = lexicon.word_table["#0"] + + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + # G.aux_labels is not needed in later computations, so + # remove it here. + del G.aux_labels + # CAUTION: The following line is crucial. + # Arcs entering the back-off state have label equal to #0. + # We have to change it to 0 here. + G.labels[G.labels >= first_word_disambig_id] = 0 + # See https://github.com/k2-fsa/k2/issues/874 + # for why we need to set G.properties to None + G.__dict__["_properties"] = None + G = k2.Fsa.from_fsas([G]).to(device) + G = k2.arc_sort(G) + # Save a dummy value so that it can be loaded in C++. + # See https://github.com/pytorch/pytorch/issues/67902 + # for why we need to do this. + G.dummy = 1 + + torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt") + else: + logging.info("Loading pre-compiled G_4_gram.pt") + d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device) + G = k2.Fsa.from_dict(d) + + if params.decoding_method == "whole-lattice-rescoring": + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + G = G.to(device) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + else: + G = None + + 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() + + 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 + librispeech = LibriSpeechAsrDataModule(args) + + dev_clean_cuts = librispeech.dev_clean_cuts() + dev_other_cuts = librispeech.dev_other_cuts() + + dev_clean_dl = librispeech.test_dataloaders( + dev_clean_cuts, + do_normalize=params.do_normalize, + ) + dev_other_dl = librispeech.test_dataloaders( + dev_other_cuts, + do_normalize=params.do_normalize, + ) + + test_sets = ["dev-clean", "dev-other"] + test_dl = [dev_clean_dl, dev_other_dl] + + # test_clean_cuts = librispeech.test_clean_cuts() + # test_other_cuts = librispeech.test_other_cuts() + + # test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + # test_other_dl = librispeech.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, + HLG=HLG, + H=H, + lexicon=lexicon, + graph_compiler=graph_compiler, + word_table=lexicon.word_table, + G=G, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/zipformer_ctc/dataset.py b/egs/librispeech/SSL/zipformer_ctc/dataset.py new file mode 120000 index 0000000000..942f79d30b --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/dataset.py @@ -0,0 +1 @@ +../zipformer/dataset.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/encoder_interface.py b/egs/librispeech/SSL/zipformer_ctc/encoder_interface.py new file mode 120000 index 0000000000..1f1d40cf90 --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/encoder_interface.py @@ -0,0 +1 @@ +../zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py new file mode 100755 index 0000000000..e35fad39a1 --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py @@ -0,0 +1,1399 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +# For HuBERT model finetuning: +./zipformer_ctc/finetune_ctc.py \ + --world-size 8 \ + --num-epochs 222 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer_ctc/exp \ + --full-libri 0 \ + --max-duration 600 +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from hubert_ce import HubertModel +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 torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +from icefall import diagnostics +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.lexicon import Lexicon +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.accum_grad + * (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.", + ) + + # hubert parameters + parser.add_argument( + "--label-rate", + type=float, + default=50, + ) + + parser.add_argument( + "--sample-rate", + type=float, + default=16000, + ) + + parser.add_argument( + "--extractor-mode", + type=str, + default="default", + help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group + norm with d groups in the first conv block, whereas layer_norm + has layer norms in every block (meant to use with normalize=True)""", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--conv-bias", type=bool, default=False, help="include bias in conv encoder" + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + default=1.0, + help="multiply feature extractor var grads by this", + ) + + # masking + parser.add_argument("--mask-length", type=int, default=10, help="mask_length") + + parser.add_argument( + "--mask-prob", + type=float, + default=0.65, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length", + ) + + parser.add_argument( + "--mask-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-overlap", + type=bool, + default=False, + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # channel masking + parser.add_argument( + "--mask-channel-length", + type=int, + default=10, + help="length of the mask for features (channels)", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + default=0.0, + help="probability of replacing a feature with 0", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length for channel masking", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + type=bool, + default=False, + help="whether to allow channel masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # loss computation + parser.add_argument( + "--skip-masked", + type=bool, + default=False, + help="skip computing losses over masked frames", + ) + + parser.add_argument( + "--skip-nomask", + type=bool, + default=False, + help="skip computing losses over unmasked frames", + ) + + parser.add_argument( + "--checkpoint-activations", + type=bool, + default=False, + help="recompute activations and save memory for extra compute", + ) + + parser.add_argument( + "--pred-masked-weight", + type=float, + default=1, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--pred-nomask-weight", + type=float, + default=0, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--loss-weights", + type=float, + nargs="*", + default=[10], + help="weight for masked part in ssl loss", + ) + + # FP16 optimization + parser.add_argument( + "--required-seq-len-multiple", + type=int, + default=2, + help="pad the input to encoder such that the sequence length is divisible by multiple", + ) + + parser.add_argument( + "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" + ) + + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Positional encoding type to use in conformer", + ) + + parser.add_argument( + "--logit-temp", type=float, default=0.1, help="temperature to divide logits by" + ) + + parser.add_argument( + "--dropout-input", + type=float, + default=0.0, + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + default=0.0, + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--num-classes", + type=int, + nargs="*", + default=[504], + help="""num class, a little larger than the number of cluster, + the largest is for padding, + and the value should be the multiple of 4, for faster computation""", + ) + + parser.add_argument( + "--untie-final-proj", + type=bool, + default=False, + help="use separate projection for each target", + ) + + +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=222, + 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="hubert/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--pretrained-dir", + type=str, + help="""The pretrained model dir. + It specifies the directory where the pretrained checkpoint is saved.""", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.001, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=100000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=100, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--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( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + 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=100000, + 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( + "--accum-grad", + type=int, + default=1, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + + contains number of updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + if hasattr(params, "pretrained_dir"): + logging.info(f"Loading {params.pretrained_dir}") + pretrained = torch.load(params.pretrained_dir) + encoder = HubertModel(params) + encoder.load_state_dict(pretrained["model"]) + else: + encoder = HubertModel(params) + return encoder + + +def get_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + model = AsrModel( + encoder=encoder, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + graph_compiler: CharCtcTrainingGraphCompiler, + 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 `dataset.HubertAsrDataset()` + 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 + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + batch_idx_train = params.batch_idx_train + + texts = batch["supervisions"]["text"] + y = graph_compiler.texts_to_ids(texts, sep="|") + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + ctc_loss, num_frames = model( + x=audio, + padding_mask=padding_mask, + y=y, + ) + + assert ctc_loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_frames.sum().item() + + # Note: We use reduction=sum while computing the loss. + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return ctc_loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + graph_compiler: CharCtcTrainingGraphCompiler, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + graph_compiler: CharCtcTrainingGraphCompiler, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + 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 sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + 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, + graph_compiler=graph_compiler, + 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 / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, graph_compiler=graph_compiler) + 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, + graph_compiler=graph_compiler, + 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 + ) + + if sub_batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + 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}") + + lexicon = Lexicon(params.lang_dir) + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + delimiter="|", + ) + + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs, warmup_batches=0) + + 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) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + return True + + 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 = librispeech.train_dataloaders( + train_cuts, + do_normalize=params.do_normalize, + sampler_state_dict=sampler_state_dict, + world_size=world_size, + rank=rank, + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + + valid_dl = librispeech.valid_dataloaders( + valid_cuts, + do_normalize=params.do_normalize, + world_size=world_size, + rank=rank, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + graph_compiler=graph_compiler, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + graph_compiler=graph_compiler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + graph_compiler: CharCtcTrainingGraphCompiler, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertAsrDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + y = graph_compiler.texts_to_ids(batch["supervisions"]["text"], sep="|") + 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, + graph_compiler: CharCtcTrainingGraphCompiler, + 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, + graph_compiler=graph_compiler, + 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, graph_compiler=graph_compiler) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.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/librispeech/SSL/zipformer_ctc/hubert_ce.py b/egs/librispeech/SSL/zipformer_ctc/hubert_ce.py new file mode 120000 index 0000000000..ef944a9e78 --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/hubert_ce.py @@ -0,0 +1 @@ +../zipformer/hubert_ce.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/model.py b/egs/librispeech/SSL/zipformer_ctc/model.py new file mode 100644 index 0000000000..1a6df71d05 --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/model.py @@ -0,0 +1,153 @@ +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Zengwei Yao, +# Yifan 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. + +from typing import Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from scaling import ScaledLinear + +from icefall.utils import add_sos + + +class AsrModel(nn.Module): + def __init__( + self, + encoder, + encoder_dim: int = 768, + vocab_size: int = 500, + ): + """CTC ASR model. + + - Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks (http://imagine.enpc.fr/~obozinsg/teaching/mva_gm/papers/ctc.pdf) + + Args: + encoder: + It is the transcription network in the paper. Its accepts + inputs: `x` of (N, T, encoder_dim). + It returns two tensors: `logits` of shape (N, T, encoder_dim) and + `logit_lens` of shape (N,). + """ + super().__init__() + + self.encoder = encoder + + # Modules for CTC head + self.ctc_output = nn.Sequential( + nn.Dropout(p=0.1), + nn.Linear(encoder_dim, vocab_size), + nn.LogSoftmax(dim=-1), + ) + + def forward_encoder( + self, + x: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute encoder outputs. + Args: + x: + A 2-D tensor of shape (N, T). + + Returns: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + """ + if padding_mask is None: + padding_mask = torch.zeros_like(x, dtype=torch.bool) + + encoder_out, padding_mask = self.encoder.extract_features( + source=x, + padding_mask=padding_mask, + mask=self.encoder.training, + ) + encoder_out_lens = torch.sum(~padding_mask, dim=1) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + + return encoder_out, encoder_out_lens + + def forward_ctc( + self, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + targets: torch.Tensor, + target_lengths: torch.Tensor, + ) -> torch.Tensor: + """Compute CTC loss. + Args: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + targets: + Target Tensor of shape (sum(target_lengths)). The targets are assumed + to be un-padded and concatenated within 1 dimension. + """ + # Compute CTC log-prob + ctc_output = self.ctc_output(encoder_out) # (N, T, C) + + ctc_loss = torch.nn.functional.ctc_loss( + log_probs=ctc_output.permute(1, 0, 2), # (T, N, C) + targets=targets.cpu(), + input_lengths=encoder_out_lens.cpu(), + target_lengths=target_lengths.cpu(), + reduction="sum", + ) + return ctc_loss + + def forward( + self, + x: torch.Tensor, + y: k2.RaggedTensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 2-D tensor of shape (N, T). + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + Returns: + Return the CTC loss, + """ + assert x.ndim == 2, x.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == y.dim0, (x.shape, y.dim0) + + # Compute encoder outputs + encoder_out, encoder_out_lens = self.forward_encoder(x, padding_mask) + + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + # Compute CTC loss + targets = y.values + ctc_loss = self.forward_ctc( + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + targets=targets, + target_lengths=y_lens, + ) + + return ctc_loss, encoder_out_lens diff --git a/egs/librispeech/SSL/zipformer_ctc/optim.py b/egs/librispeech/SSL/zipformer_ctc/optim.py new file mode 120000 index 0000000000..207eecfcda --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/optim.py @@ -0,0 +1 @@ +../zipformer/optim.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/scaling.py b/egs/librispeech/SSL/zipformer_ctc/scaling.py new file mode 120000 index 0000000000..58e4b0a0fe --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/scaling.py @@ -0,0 +1 @@ +../zipformer/scaling.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/utils.py b/egs/librispeech/SSL/zipformer_ctc/utils.py new file mode 120000 index 0000000000..3f60b6eedc --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/utils.py @@ -0,0 +1 @@ +../zipformer/utils.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/wav2vec2_module.py b/egs/librispeech/SSL/zipformer_ctc/wav2vec2_module.py new file mode 120000 index 0000000000..2fac5c458f --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/wav2vec2_module.py @@ -0,0 +1 @@ +../zipformer/wav2vec2_module.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer_ctc/zipformer.py b/egs/librispeech/SSL/zipformer_ctc/zipformer.py new file mode 120000 index 0000000000..a064749a48 --- /dev/null +++ b/egs/librispeech/SSL/zipformer_ctc/zipformer.py @@ -0,0 +1 @@ +../zipformer/zipformer.py \ No newline at end of file diff --git a/icefall/char_graph_compiler.py b/icefall/char_graph_compiler.py index 8c2355c873..8c51ce1253 100644 --- a/icefall/char_graph_compiler.py +++ b/icefall/char_graph_compiler.py @@ -70,12 +70,15 @@ def texts_to_ids(self, texts: List[str], sep: str = "") -> List[List[int]]: Returns: Return a list-of-list of token IDs. """ - assert sep in ("", "/"), sep + assert sep in ("", "/", "|"), sep ids: List[List[int]] = [] whitespace = re.compile(r"([ \t])") for text in texts: if sep == "": text = re.sub(whitespace, "", text) + elif sep == "|": + text = re.sub(r"\s+", " ", text) + text = re.sub(" ", "|", text) else: text = text.split(sep) sub_ids = [ From f26dd3ba17d5d87b4618c9c127d0fb892bbe92f7 Mon Sep 17 00:00:00 2001 From: Yifan Yeung Date: Sat, 10 Aug 2024 22:07:58 +0800 Subject: [PATCH 02/24] support multinode multigpu update --- egs/librilight/SSL/pretrain.sh | 22 ---- .../SSL/run_multi_node_multi_gpu.sh | 116 ++++++++++++++++++ egs/librilight/SSL/zipformer/pretrain.py | 68 +++++----- .../SSL/zipformer/ssl_datamodule.py | 26 ++-- 4 files changed, 172 insertions(+), 60 deletions(-) delete mode 100755 egs/librilight/SSL/pretrain.sh create mode 100755 egs/librilight/SSL/run_multi_node_multi_gpu.sh diff --git a/egs/librilight/SSL/pretrain.sh b/egs/librilight/SSL/pretrain.sh deleted file mode 100755 index c282967f9a..0000000000 --- a/egs/librilight/SSL/pretrain.sh +++ /dev/null @@ -1,22 +0,0 @@ -export PYTHONPATH=$(pwd)/../../.. - -./zipformer/pretrain.py \ - --world-size 8 \ - --num-epochs 30 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir zipformer/exp_pretrain \ - --max-duration 650 \ - --quadratic-duration 512 \ - --accum-grad 1 \ - --do-normalize 1 \ - --mask-prob 0.8 \ - --extractor-mode "layer_norm" \ - --dropout-input 0.0 \ - --dropout-features 0.0 \ - --feature-grad-mult 1.0 \ - --num-encoder-layers 2,2,3,4,3,2 \ - --feedforward-dim 512,768,1024,1536,1024,768 \ - --encoder-dim 192,256,448,768,448,192 \ - --encoder-unmasked-dim 192,192,256,256,256,192 \ - --base-lr 0.045 diff --git a/egs/librilight/SSL/run_multi_node_multi_gpu.sh b/egs/librilight/SSL/run_multi_node_multi_gpu.sh new file mode 100755 index 0000000000..03685d204c --- /dev/null +++ b/egs/librilight/SSL/run_multi_node_multi_gpu.sh @@ -0,0 +1,116 @@ +#!/usr/bin/env bash +# +# Copyright 2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Yifan Yang) +# +# This script is the entry point to start model training +# with multi-node multi-GPU. +# +# Read the usage instructions below for how to run this script. + +set -e + +# DDP related parameters +master_addr= +node_rank= +num_nodes=4 +master_port=12354 + +. shared/parse_options.sh + +function usage() { + echo "Usage: " + echo "" + echo " $0 \\" + echo " --master-addr \\" + echo " --master-port \\" + echo " --node-rank \\" + echo " --num-nodes " + echo "" + echo " --master-addr The ip address of the master node." + echo " --master-port The port of the master node." + echo " --node-rank Rank of this node." + echo " --num-nodes Number of nodes in DDP training." + echo "" + echo "Usage example:" + echo "Suppose you want to use DDP with two machines:" + echo " (1) Machine 1 has 4 GPUs. You want to use" + echo " GPU 0, 1, and 3 for training" + echo " IP of machine 1 is: 10.177.41.71" + echo " (2) Machine 2 has 4 GPUs. You want to use" + echo " GPU 0, 2, and 3 for training" + echo " IP of machine 2 is: 10.177.41.72" + echo "You want to select machine 1 as the master node and" + echo "assume that the port 1234 is free on machine 1." + echo "" + echo "On machine 1, you run:" + echo "" + echo " export CUDA_VISIBLE_DEVICES=\"0,1,3\"" + echo " ./run_multi_node_multi_gpu.sh --master-addr 10.177.41.71 --master-port 1234 --node-rank 0 --num-nodes 2" + echo "" + echo "On machine 2, you run:" + echo "" + echo " export CUDA_VISIBLE_DEVICES=\"0,2,3\"" + echo " ./run_multi_node_multi_gpu.sh --master-addr 10.177.41.71 --master-port 1234 --node-rank 1 --num-nodes 2" + echo "" + echo "Note 1:" + echo " You use CUDA_VISIBLE_DEVICES to decide which GPUs are used for training." + echo "" + echo "Note 2:" + echo " If you use torch < 1.9.0, then every node has to use the same number of GPUs for training." + echo " If you use torch >= 1.9.0, different nodes can have a different number of GPUs for training." + exit 1 +} + +default='\033[0m' +bold='\033[1m' +red='\033[31m' + +function error() { + printf "${bold}${red}[ERROR]${default} $1\n" +} + +[ ! -z $CUDA_VISIBLE_DEVICES ] || ( echo; error "Please set CUDA_VISIBLE_DEVICES"; echo; usage ) +[ ! -z $master_addr ] || ( echo; error "Please set --master-addr"; echo; usage ) +[ ! -z $master_port ] || ( echo; error "Please set --master-port"; echo; usage ) +[ ! -z $node_rank ] || ( echo; error "Please set --node-rank"; echo; usage ) +[ ! -z $num_nodes ] || ( echo; error "Please set --num-nodes"; echo; usage ) + +# Number of GPUs this node has +num_gpus=$(python3 -c "s=\"$CUDA_VISIBLE_DEVICES\"; print(len(s.split(',')))") + +echo "CUDA_VISIBLE_DEVICES: $CUDA_VISIBLE_DEVICES" +echo "num_gpus: $num_gpus" +echo "master_addr: $master_addr" + +export MASTER_ADDR=$master_addr +export MASTER_PORT=$master_port + +set -x + +torchrun \ + --nproc_per_node $num_gpus \ + --nnodes $num_nodes \ + --node_rank $node_rank \ + --master_addr $master_addr \ + --master_port $master_port \ + zipformer/pretrain.py \ + --use-multi-node 1 \ + --master-port $master_port \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp_pretrain \ + --max-duration 600 \ + --quadratic-duration 1024 \ + --accum-grad 1 \ + --do-normalize 1 \ + --mask-prob 0.8 \ + --dropout-input 0.0 \ + --dropout-features 0.0 \ + --feature-grad-mult 1.0 \ + --num-encoder-layers 2,2,3,4,3,2 \ + --feedforward-dim 512,768,1024,1536,1024,768 \ + --encoder-dim 192,256,448,768,448,192 \ + --encoder-unmasked-dim 192,192,256,256,256,192 \ + --base-lr 0.045 \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py index 251184b69f..27e41ed3f0 100755 --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -20,23 +20,6 @@ # 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="0,1,2,3,4,5,6,7" - -# For hubert model pretraining: -./zipformer/pretrain.py \ - --world-size 8 \ - --num-epochs 400 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir hubert/exp \ - --full-libri 1 \ - --max-duration 87.5 \ - --accum-grad 4 -""" - import argparse import copy @@ -46,7 +29,6 @@ from shutil import copyfile from typing import Any, Dict, Optional, Tuple, Union -import lhotse import optim import torch import torch.multiprocessing as mp @@ -69,7 +51,13 @@ save_checkpoint_with_global_batch_idx, update_averaged_model, ) -from icefall.dist import cleanup_dist, setup_dist +from icefall.dist import ( + cleanup_dist, + get_local_rank, + get_rank, + get_world_size, + setup_dist, +) from icefall.env import get_env_info from icefall.hooks import register_inf_check_hooks from icefall.utils import ( @@ -405,6 +393,15 @@ def get_parser(): formatter_class=argparse.ArgumentDefaultsHelpFormatter ) + parser.add_argument( + "--use-multi-node", + type=str2bool, + default=False, + help="""True if using multi-node multi-GPU. + You are not supposed to set it directly. + """, + ) + parser.add_argument( "--world-size", type=int, @@ -572,7 +569,7 @@ def get_parser(): parser.add_argument( "--accum-grad", type=int, - default=4, + default=1, help="""update gradient when batch_idx_train % accum_grad == 0. """, ) @@ -1090,8 +1087,15 @@ def run(rank, world_size, args): params.update(vars(args)) fix_random_seed(params.seed) + + if params.use_multi_node: + local_rank = get_local_rank() + else: + local_rank = rank + logging.info(f"rank: {rank}, world_size: {world_size}, local_rank: {local_rank}") + if world_size > 1: - setup_dist(rank, world_size, params.master_port) + setup_dist(rank, world_size, params.master_port, params.use_multi_node) setup_logger(f"{params.exp_dir}/log/log-train") logging.info("Training started") @@ -1103,8 +1107,8 @@ def run(rank, world_size, args): device = torch.device("cpu") if torch.cuda.is_available(): - device = torch.device("cuda", rank) - logging.info(f"Device: {device}") + device = torch.device("cuda", local_rank) + logging.info(f"Device: {device}, rank: {rank}, local_rank: {local_rank}") logging.info(params) logging.info("About to create model") @@ -1127,7 +1131,7 @@ def run(rank, world_size, args): model.to(device) if world_size > 1: logging.info("Using DDP") - model = DDP(model, device_ids=[rank], find_unused_parameters=True) + model = DDP(model, device_ids=[local_rank], find_unused_parameters=True) optimizer = ScaledAdam( get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), @@ -1358,12 +1362,18 @@ def main(): 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) + if args.use_multi_node: + rank = get_rank() + world_size = get_world_size() + args.world_size = world_size + run(rank=rank, world_size=world_size, args=args) else: - run(rank=0, world_size=1, args=args) + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) torch.set_num_threads(1) diff --git a/egs/librilight/SSL/zipformer/ssl_datamodule.py b/egs/librilight/SSL/zipformer/ssl_datamodule.py index 3cf7c2f73c..1c6d74a4a6 100644 --- a/egs/librilight/SSL/zipformer/ssl_datamodule.py +++ b/egs/librilight/SSL/zipformer/ssl_datamodule.py @@ -103,7 +103,7 @@ def add_arguments(cls, parser: argparse.ArgumentParser): help="We will draw this many cuts to estimate the duration" "bins for creating similar-duration buckets. Larger number" "means a better estimate to the data distribution, possibly" - "at a longer init cost." + "at a longer init cost.", ) group.add_argument( "--quadratic-duration", @@ -304,28 +304,36 @@ def small_cuts(self) -> CutSet: def medium_cuts(self) -> CutSet: logging.info("About to get librilight medium cuts") filenames = glob.glob( - str(self.args.manifest_dir / "medium_split" / "librilight_cuts_medium.*.jsonl.gz") + str( + self.args.manifest_dir + / "medium_split" + / "librilight_cuts_medium.*.jsonl.gz" + ) ) pattern = re.compile(r"librilight_cuts_medium.([0-9]+).jsonl.gz") idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames) idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) sorted_filenames = [f[1] for f in idx_filenames] - logging.info(f"Loading Libri-Light medium {len(sorted_filenames)} splits in lazy mode") - return lhotse.combine( - lhotse.load_manifest_lazy(p) for p in sorted_filenames + logging.info( + f"Loading Libri-Light medium {len(sorted_filenames)} splits in lazy mode" ) + return lhotse.combine(lhotse.load_manifest_lazy(p) for p in sorted_filenames) @lru_cache() def large_cuts(self) -> CutSet: logging.info("About to get librilight large cuts") filenames = glob.glob( - str(self.args.manifest_dir / "large_split" / "librilight_cuts_large.*.jsonl.gz") + str( + self.args.manifest_dir + / "large_split" + / "librilight_cuts_large.*.jsonl.gz" + ) ) pattern = re.compile(r"librilight_cuts_large.([0-9]+).jsonl.gz") idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames) idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) sorted_filenames = [f[1] for f in idx_filenames] - logging.info(f"Loading Libri-Light large {len(sorted_filenames)} splits in lazy mode") - return lhotse.combine( - lhotse.load_manifest_lazy(p) for p in sorted_filenames + logging.info( + f"Loading Libri-Light large {len(sorted_filenames)} splits in lazy mode" ) + return lhotse.combine(lhotse.load_manifest_lazy(p) for p in sorted_filenames) From d025ce11ff8590c137136b981c6d7fc682cf1912 Mon Sep 17 00:00:00 2001 From: yifanyeung Date: Mon, 19 Aug 2024 23:14:26 +0800 Subject: [PATCH 03/24] use lr hours in librilight ssl --- egs/librilight/SSL/run_multi_node_multi_gpu.sh | 18 +++++++++--------- egs/librilight/SSL/zipformer/pretrain.py | 13 ++++++++----- 2 files changed, 17 insertions(+), 14 deletions(-) diff --git a/egs/librilight/SSL/run_multi_node_multi_gpu.sh b/egs/librilight/SSL/run_multi_node_multi_gpu.sh index 03685d204c..d9d839b3f9 100755 --- a/egs/librilight/SSL/run_multi_node_multi_gpu.sh +++ b/egs/librilight/SSL/run_multi_node_multi_gpu.sh @@ -97,20 +97,20 @@ torchrun \ zipformer/pretrain.py \ --use-multi-node 1 \ --master-port $master_port \ - --num-epochs 30 \ + --num-epochs 20 \ --start-epoch 1 \ --use-fp16 1 \ --exp-dir zipformer/exp_pretrain \ - --max-duration 600 \ + --max-duration 350 \ --quadratic-duration 1024 \ --accum-grad 1 \ --do-normalize 1 \ --mask-prob 0.8 \ - --dropout-input 0.0 \ - --dropout-features 0.0 \ + --dropout-input 0.1 \ + --dropout-features 0.1 \ --feature-grad-mult 1.0 \ - --num-encoder-layers 2,2,3,4,3,2 \ - --feedforward-dim 512,768,1024,1536,1024,768 \ - --encoder-dim 192,256,448,768,448,192 \ - --encoder-unmasked-dim 192,192,256,256,256,192 \ - --base-lr 0.045 \ No newline at end of file + --num-encoder-layers 2,2,4,5,4,2 \ + --feedforward-dim 768,1536,2048,3072,2048,1536 \ + --encoder-dim 256,512,768,1024,768,512 \ + --encoder-unmasked-dim 192,192,256,320,256,192 \ + --base-lr 0.045 diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py index 27e41ed3f0..6fe5a59908 100755 --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -472,10 +472,10 @@ def get_parser(): ) parser.add_argument( - "--lr-epochs", + "--lr-hours", type=float, - default=0.2, - help="""Number of epochs that affects how rapidly the learning rate decreases. + default=30000, + help="""Number of hours that affects how rapidly the learning rate decreases. """, ) @@ -951,6 +951,10 @@ def save_bad_model(suffix: str = ""): if sub_batch_idx % params.accum_grad == params.accum_grad - 1: params.batch_idx_train += 1 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() @@ -1142,7 +1146,7 @@ def run(rank, world_size, args): scheduler = Eden( optimizer, params.lr_batches, - params.lr_epochs, + params.lr_hours, params.warmup_batches, params.warmup_start, ) @@ -1246,7 +1250,6 @@ def remove_short_and_long_utt(c: Cut): scaler.load_state_dict(checkpoints["grad_scaler"]) for epoch in range(params.start_epoch, params.num_epochs + 1): - scheduler.step_epoch(epoch - 1) fix_random_seed(params.seed + epoch - 1) train_dl.sampler.set_epoch(epoch - 1) From d0a96a601cd2664177689db002aea4063f2d9aca Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Wed, 21 Aug 2024 12:42:06 +0800 Subject: [PATCH 04/24] Update run_multi_node_multi_gpu.sh --- egs/librilight/SSL/run_multi_node_multi_gpu.sh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/egs/librilight/SSL/run_multi_node_multi_gpu.sh b/egs/librilight/SSL/run_multi_node_multi_gpu.sh index d9d839b3f9..a3ade79667 100755 --- a/egs/librilight/SSL/run_multi_node_multi_gpu.sh +++ b/egs/librilight/SSL/run_multi_node_multi_gpu.sh @@ -106,8 +106,8 @@ torchrun \ --accum-grad 1 \ --do-normalize 1 \ --mask-prob 0.8 \ - --dropout-input 0.1 \ - --dropout-features 0.1 \ + --dropout-input 0.0 \ + --dropout-features 0.0 \ --feature-grad-mult 1.0 \ --num-encoder-layers 2,2,4,5,4,2 \ --feedforward-dim 768,1536,2048,3072,2048,1536 \ From eca8afcb83cc2f8e1f3cdc69245701d671eeb277 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Wed, 21 Aug 2024 12:47:45 +0800 Subject: [PATCH 05/24] Update pretrain.py --- egs/librilight/SSL/zipformer/pretrain.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py index 6fe5a59908..f5c98f7eb7 100755 --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -474,7 +474,7 @@ def get_parser(): parser.add_argument( "--lr-hours", type=float, - default=30000, + default=10000, help="""Number of hours that affects how rapidly the learning rate decreases. """, ) From 8fe6713d1022fcdee78570bb2ba5a21e761fdc08 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Wed, 21 Aug 2024 12:53:19 +0800 Subject: [PATCH 06/24] Update pretrain.py --- egs/librilight/SSL/zipformer/pretrain.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py index f5c98f7eb7..93613be5a5 100755 --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -474,7 +474,7 @@ def get_parser(): parser.add_argument( "--lr-hours", type=float, - default=10000, + default=30000, help="""Number of hours that affects how rapidly the learning rate decreases. """, ) @@ -482,7 +482,7 @@ def get_parser(): parser.add_argument( "--warmup-batches", type=float, - default=5000, + default=10000, help="Eden warmup steps", ) From 6dbcdbac5d57a32ca392139a917d076f6bba0aff Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Thu, 22 Aug 2024 10:55:23 +0800 Subject: [PATCH 07/24] Update pretrain.py --- egs/librilight/SSL/zipformer/pretrain.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py index 93613be5a5..f5c98f7eb7 100755 --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -474,7 +474,7 @@ def get_parser(): parser.add_argument( "--lr-hours", type=float, - default=30000, + default=10000, help="""Number of hours that affects how rapidly the learning rate decreases. """, ) @@ -482,7 +482,7 @@ def get_parser(): parser.add_argument( "--warmup-batches", type=float, - default=10000, + default=5000, help="Eden warmup steps", ) From f672df2a636a1b131a41dd5ea977073c27fa0c3e Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Thu, 22 Aug 2024 10:57:26 +0800 Subject: [PATCH 08/24] Update run_multi_node_multi_gpu.sh --- egs/librilight/SSL/run_multi_node_multi_gpu.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librilight/SSL/run_multi_node_multi_gpu.sh b/egs/librilight/SSL/run_multi_node_multi_gpu.sh index a3ade79667..f8023095a5 100755 --- a/egs/librilight/SSL/run_multi_node_multi_gpu.sh +++ b/egs/librilight/SSL/run_multi_node_multi_gpu.sh @@ -112,5 +112,5 @@ torchrun \ --num-encoder-layers 2,2,4,5,4,2 \ --feedforward-dim 768,1536,2048,3072,2048,1536 \ --encoder-dim 256,512,768,1024,768,512 \ - --encoder-unmasked-dim 192,192,256,320,256,192 \ + --encoder-unmasked-dim 256,256,256,320,256,256 \ --base-lr 0.045 From ad61d72dfad73e868a01e7e5ceac057303d95d11 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Thu, 22 Aug 2024 18:32:33 +0800 Subject: [PATCH 09/24] Update pretrain.py --- egs/librilight/SSL/zipformer/pretrain.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py index f5c98f7eb7..10dd68ea91 100755 --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -482,7 +482,7 @@ def get_parser(): parser.add_argument( "--warmup-batches", type=float, - default=5000, + default=1000, help="Eden warmup steps", ) From ef5cf0250d437884eece93ec12ca596935af782f Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Mon, 2 Sep 2024 14:11:22 +0800 Subject: [PATCH 10/24] Update finetune_ctc.py --- egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py | 1 - 1 file changed, 1 deletion(-) diff --git a/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py index e35fad39a1..d5dd8d71f4 100755 --- a/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py +++ b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py @@ -1136,7 +1136,6 @@ def run(rank, world_size, args): graph_compiler = CharCtcTrainingGraphCompiler( lexicon=lexicon, device=device, - delimiter="|", ) params.blank_id = lexicon.token_table[""] From 13c6e81189b6102c404f012729e125f75c30278a Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Thu, 5 Sep 2024 12:57:42 +0800 Subject: [PATCH 11/24] Delete egs/librispeech/SSL/pretrain.sh --- egs/librispeech/SSL/pretrain.sh | 19 ------------------- 1 file changed, 19 deletions(-) delete mode 100755 egs/librispeech/SSL/pretrain.sh diff --git a/egs/librispeech/SSL/pretrain.sh b/egs/librispeech/SSL/pretrain.sh deleted file mode 100755 index 70401aeb82..0000000000 --- a/egs/librispeech/SSL/pretrain.sh +++ /dev/null @@ -1,19 +0,0 @@ -./zipformer/pretrain.py \ - --world-size 8 \ - --num-epochs 300 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir zipformer/exp_pretrain \ - --full-libri 1 \ - --max-duration 600 \ - --accum-grad 1 \ - --do-normalize 0 \ - --mask-prob 0.8 \ - --dropout-input 0.1 \ - --dropout-features 0.1 \ - --feature-grad-mult 0.1 \ - --num-encoder-layers 2,2,3,4,3,2 \ - --feedforward-dim 512,768,1024,1536,1024,768 \ - --encoder-dim 192,256,448,768,448,192 \ - --encoder-unmasked-dim 192,192,256,256,256,192 \ - --base-lr 0.045 From 673ca14e7f0d0f77681796c84ea539d60f7a0eb9 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 7 Sep 2024 23:10:33 +0800 Subject: [PATCH 12/24] Delete egs/librilight/SSL/zipformer/finetune.py --- egs/librilight/SSL/zipformer/finetune.py | 1555 ---------------------- 1 file changed, 1555 deletions(-) delete mode 100755 egs/librilight/SSL/zipformer/finetune.py diff --git a/egs/librilight/SSL/zipformer/finetune.py b/egs/librilight/SSL/zipformer/finetune.py deleted file mode 100755 index 2e521f1772..0000000000 --- a/egs/librilight/SSL/zipformer/finetune.py +++ /dev/null @@ -1,1555 +0,0 @@ -#!/usr/bin/env python3 -# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, -# Wei Kang, -# Mingshuang Luo, -# Zengwei Yao, -# Yifan Yang, -# Daniel Povey) -# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) -# -# 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="0,1,2,3,4,5,6,7" - -# For HuBERT model finetuning: -./hubert/finetune.py \ - --world-size 8 \ - --num-epochs 200 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir hubert/exp \ - --full-libri 0 \ - --max-duration 1000 - -It supports finetuning 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 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 LibriSpeechAsrDataModule -from decoder import Decoder -from hubert_ce import HubertModel -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 torch import Tensor -from torch.cuda.amp import GradScaler -from torch.nn.parallel import DistributedDataParallel as DDP -from torch.utils.tensorboard import SummaryWriter - -from icefall import diagnostics -from icefall.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.accum_grad - * (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.", - ) - - # hubert parameters - parser.add_argument( - "--label-rate", - type=float, - default=50, - ) - - parser.add_argument( - "--sample-rate", - type=float, - default=16000, - ) - - parser.add_argument( - "--extractor-mode", - type=str, - default="default", - help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group - norm with d groups in the first conv block, whereas layer_norm - has layer norms in every block (meant to use with normalize=True)""", - ) - - parser.add_argument( - "--conv-feature-layers", - type=str, - default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", - help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", - ) - - parser.add_argument( - "--conv-bias", type=bool, default=False, help="include bias in conv encoder" - ) - - parser.add_argument( - "--feature-grad-mult", - type=float, - default=1.0, - help="multiply feature extractor var grads by this", - ) - - # masking - parser.add_argument("--mask-length", type=int, default=10, help="mask_length") - - parser.add_argument( - "--mask-prob", - type=float, - default=0.65, - help="probability of replacing a token with mask", - ) - - parser.add_argument( - "--mask-selection", - type=str, - choices=["static", "uniform", "normal", "poisson"], - default="static", - help="how to choose mask length", - ) - - parser.add_argument( - "--mask-other", - type=float, - default=0, - help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", - ) - - parser.add_argument( - "--no-mask-overlap", - type=bool, - default=False, - help="whether to allow masks to overlap", - ) - - parser.add_argument( - "--mask-min-space", - type=int, - default=1, - help="min space between spans (if no overlap is enabled)", - ) - - # channel masking - parser.add_argument( - "--mask-channel-length", - type=int, - default=10, - help="length of the mask for features (channels)", - ) - - parser.add_argument( - "--mask-channel-prob", - type=float, - default=0.0, - help="probability of replacing a feature with 0", - ) - - parser.add_argument( - "--mask-channel-selection", - type=str, - choices=["static", "uniform", "normal", "poisson"], - default="static", - help="how to choose mask length for channel masking", - ) - - parser.add_argument( - "--mask-channel-other", - type=float, - default=0, - help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", - ) - - parser.add_argument( - "--no-mask-channel-overlap", - type=bool, - default=False, - help="whether to allow channel masks to overlap", - ) - - parser.add_argument( - "--mask-channel-min-space", - type=int, - default=1, - help="min space between spans (if no overlap is enabled)", - ) - - # loss computation - parser.add_argument( - "--skip-masked", - type=bool, - default=False, - help="skip computing losses over masked frames", - ) - - parser.add_argument( - "--skip-nomask", - type=bool, - default=False, - help="skip computing losses over unmasked frames", - ) - - parser.add_argument( - "--checkpoint-activations", - type=bool, - default=False, - help="recompute activations and save memory for extra compute", - ) - - parser.add_argument( - "--pred-masked-weight", - type=float, - default=1, - help="weight for masked part in ssl loss", - ) - - parser.add_argument( - "--pred-nomask-weight", - type=float, - default=0, - help="weight for masked part in ssl loss", - ) - - parser.add_argument( - "--loss-weights", - type=float, - nargs="*", - default=[10], - help="weight for masked part in ssl loss", - ) - - # FP16 optimization - parser.add_argument( - "--required-seq-len-multiple", - type=int, - default=2, - help="pad the input to encoder such that the sequence length is divisible by multiple", - ) - - parser.add_argument( - "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" - ) - - parser.add_argument( - "--pos-enc-type", - type=str, - default="abs", - help="Positional encoding type to use in conformer", - ) - - parser.add_argument( - "--logit-temp", type=float, default=0.1, help="temperature to divide logits by" - ) - - parser.add_argument( - "--dropout-input", - type=float, - default=0.0, - help="dropout to apply to the input (after feat extr)", - ) - - parser.add_argument( - "--dropout-features", - type=float, - default=0.0, - help="dropout to apply to the features (after feat extr)", - ) - - parser.add_argument( - "--num-classes", - type=int, - nargs="*", - default=[504], - help="""num class, a little larger than the number of cluster, - the largest is for padding, - and the value should be the multiple of 4, for faster computation""", - ) - - parser.add_argument( - "--untie-final-proj", - type=bool, - default=False, - help="use separate projection for each target", - ) - - 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( - "--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=222, - 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="hubert/exp", - help="""The experiment dir. - It specifies the directory where all training related - files, e.g., checkpoints, log, etc, are saved - """, - ) - - parser.add_argument( - "--pretrained-dir", - type=str, - help="""The pretrained model dir. - It specifies the directory where the pretrained checkpoint is 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=0.001, help="The base learning rate." - ) - - parser.add_argument( - "--lr-batches", - type=float, - default=100000, - help="""Number of steps that affects how rapidly the learning rate - decreases. We suggest not to change this.""", - ) - - parser.add_argument( - "--lr-epochs", - type=float, - default=100, - help="""Number of epochs that affects how rapidly the learning rate decreases. - """, - ) - - parser.add_argument( - "--ref-duration", - type=float, - default=600, - help="Reference batch duration for purposes of adjusting batch counts for setting various " - "schedules inside the model", - ) - - parser.add_argument( - "--context-size", - type=int, - default=2, - help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", - ) - - parser.add_argument( - "--prune-range", - type=int, - default=5, - help="The prune range for rnnt loss, it means how many symbols(context)" - "we are using to compute the loss", - ) - - parser.add_argument( - "--lm-scale", - type=float, - default=0.25, - help="The scale to smooth the loss with lm " - "(output of prediction network) part.", - ) - - parser.add_argument( - "--am-scale", - type=float, - default=0.0, - help="The scale to smooth the loss with am (output of encoder network)" "part.", - ) - - parser.add_argument( - "--simple-loss-scale", - type=float, - default=0.5, - help="To get pruning ranges, we will calculate a simple version" - "loss(joiner is just addition), this simple loss also uses for" - "training (as a regularization item). We will scale the simple loss" - "with this parameter before adding to the final loss.", - ) - - parser.add_argument( - "--ctc-loss-scale", - type=float, - default=0.2, - help="Scale for CTC loss.", - ) - - parser.add_argument( - "--seed", - type=int, - default=42, - help="The seed for random generators intended for reproducibility", - ) - - parser.add_argument( - "--print-diagnostics", - type=str2bool, - default=False, - help="Accumulate stats on activations, print them and exit.", - ) - - parser.add_argument( - "--sanity-check", - type=str2bool, - default=False, - help="Check if any of the batches in epoch 1 would cause OOM.", - ) - - 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=100000, - 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( - "--accum-grad", - type=int, - default=1, - help="""update gradient when batch_idx_train % accum_grad == 0. - """, - ) - - parser.add_argument( - "--use-fp16", - type=str2bool, - default=False, - help="Whether to use half precision training.", - ) - - add_model_arguments(parser) - - return parser - - -def get_params() -> AttributeDict: - """Return a dict containing training parameters. - - All training related parameters that are not passed from the commandline - are saved in the variable `params`. - - Commandline options are merged into `params` after they are parsed, so - you can also access them via `params`. - - Explanation of options saved in `params`: - - - best_train_loss: Best training loss so far. It is used to select - the model that has the lowest training loss. It is - updated during the training. - - - best_valid_loss: Best validation loss so far. It is used to select - the model that has the lowest validation loss. It is - updated during the training. - - - best_train_epoch: It is the epoch that has the best training loss. - - - best_valid_epoch: It is the epoch that has the best validation loss. - - - batch_idx_train: Used to writing statistics to tensorboard. It - - contains number of updates happen to the model so far across - epochs. - - - sub_batch_idx_train: It contains number of batch 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 - - - 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, - "sub_batch_idx_train": 0, - "log_interval": 50, - "reset_interval": 200, - "valid_interval": 3000, # For the 100h subset, use 800 - # parameters for pruned RNN-T loss - "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_model(params: AttributeDict) -> nn.Module: - if hasattr(params, "pretrained_dir"): - logging.info(f"Loading {params.pretrained_dir}") - pretrained = torch.load(params.pretrained_dir) - encoder = HubertModel(params) - encoder.load_state_dict(pretrained["model"]) - else: - encoder = HubertModel(params) - 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 = 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=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 `dataset.HubertAsrDataset()` - 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 - audio = batch["audio"].to(device) - padding_mask = batch["padding_mask"].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, num_frames = model( - x=audio, - padding_mask=padding_mask, - 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"] = num_frames.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 sub_batch_idx, batch in enumerate(train_dl): - params.sub_batch_idx_train += 1 - batch_idx = sub_batch_idx // params.accum_grad - - if batch_idx % 10 == 0: - set_batch_count(model, get_adjusted_batch_count(params)) - - 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 / params.accum_grad).backward() - - if sub_batch_idx % params.accum_grad == params.accum_grad - 1: - params.batch_idx_train += 1 - scheduler.step_batch(params.batch_idx_train) - - scaler.step(optimizer) - scaler.update() - optimizer.zero_grad() - else: - continue - - 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 - ) - - if sub_batch_idx % params.accum_grad != params.accum_grad - 1: - optimizer.zero_grad() - 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_epochs, warmup_batches=0) - - 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) - - librispeech = LibriSpeechAsrDataModule(args) - - train_cuts = ( - librispeech.train_all_shuf_cuts() - if params.full_libri - else librispeech.train_clean_100_cuts() - ) - - def remove_short_and_long_utt(c: Cut): - # Keep only utterances with duration between 1 second and 20 seconds - # - # Caution: There is a reason to select 20.0 here. Please see - # ../local/display_manifest_statistics.py - # - # You should use ../local/display_manifest_statistics.py to get - # an utterance duration distribution for your dataset to select - # the threshold - if c.duration < 1.0 or c.duration > 20.0: - # logging.warning( - # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" - # ) - return False - - return True - - 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 = librispeech.train_dataloaders( - train_cuts, - do_normalize=params.do_normalize, - sampler_state_dict=sampler_state_dict, - world_size=world_size, - rank=rank, - ) - - valid_cuts = librispeech.dev_clean_cuts() - valid_cuts += librispeech.dev_other_cuts() - - valid_dl = librispeech.valid_dataloaders( - valid_cuts, - do_normalize=params.do_normalize, - world_size=world_size, - rank=rank, - ) - - if params.sanity_check and not params.print_diagnostics: - scan_pessimistic_batches_for_oom( - model=model, - train_dl=train_dl, - optimizer=optimizer, - sp=sp, - params=params, - ) - - scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) - if checkpoints and "grad_scaler" in checkpoints: - logging.info("Loading grad scaler state dict") - scaler.load_state_dict(checkpoints["grad_scaler"]) - - for epoch in range(params.start_epoch, params.num_epochs + 1): - scheduler.step_epoch(epoch - 1) - fix_random_seed(params.seed + epoch - 1) - train_dl.sampler.set_epoch(epoch - 1) - - if tb_writer is not None: - tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) - - params.cur_epoch = epoch - - train_one_epoch( - params=params, - model=model, - model_avg=model_avg, - optimizer=optimizer, - scheduler=scheduler, - sp=sp, - train_dl=train_dl, - valid_dl=valid_dl, - scaler=scaler, - tb_writer=tb_writer, - world_size=world_size, - rank=rank, - ) - - if params.print_diagnostics: - diagnostic.print_diagnostics() - break - - save_checkpoint( - params=params, - model=model, - model_avg=model_avg, - optimizer=optimizer, - scheduler=scheduler, - sampler=train_dl.sampler, - scaler=scaler, - rank=rank, - ) - - logging.info("Done!") - - if world_size > 1: - torch.distributed.barrier() - cleanup_dist() - - -def display_and_save_batch( - batch: dict, - params: AttributeDict, - sp: spm.SentencePieceProcessor, -) -> None: - """Display the batch statistics and save the batch into disk. - - Args: - batch: - A batch of data. See `dataset.HubertAsrDataset()` - 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) - - audio = batch["audio"] - logging.info(f"audio shape: {audio.shape}") - - y = sp.encode(batch["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() - LibriSpeechAsrDataModule.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() From affc43bc958934c9d8aa3d0f96bbde268f8e984f Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 7 Sep 2024 23:10:47 +0800 Subject: [PATCH 13/24] Delete egs/librilight/SSL/zipformer/decode.py --- egs/librilight/SSL/zipformer/decode.py | 1043 ------------------------ 1 file changed, 1043 deletions(-) delete mode 100755 egs/librilight/SSL/zipformer/decode.py diff --git a/egs/librilight/SSL/zipformer/decode.py b/egs/librilight/SSL/zipformer/decode.py deleted file mode 100755 index 1562c28b8b..0000000000 --- a/egs/librilight/SSL/zipformer/decode.py +++ /dev/null @@ -1,1043 +0,0 @@ -#!/usr/bin/env python3 -# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, -# Zengwei Yao, -# Yifan 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 - -(7) fast beam search (with LG) -./zipformer/decode.py \ - --epoch 28 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --max-duration 600 \ - --decoding-method fast_beam_search_nbest_LG \ - --beam 20.0 \ - --max-contexts 8 \ - --max-states 64 -""" - - -import argparse -import logging -import math -import os -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 LibriSpeechAsrDataModule -from beam_search import ( - beam_search, - fast_beam_search_nbest, - fast_beam_search_nbest_LG, - fast_beam_search_nbest_oracle, - fast_beam_search_one_best, - greedy_search, - greedy_search_batch, - modified_beam_search, - modified_beam_search_lm_rescore, - modified_beam_search_lm_rescore_LODR, - modified_beam_search_lm_shallow_fusion, - modified_beam_search_LODR, -) -from finetune import add_model_arguments, get_model, get_params - -from icefall import ContextGraph, LmScorer, NgramLm -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, -) - - -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 - - modified_beam_search_LODR - - fast_beam_search - - fast_beam_search_nbest - - fast_beam_search_nbest_oracle - - fast_beam_search_nbest_LG - If you use fast_beam_search_nbest_LG, you have to specify - `--lang-dir`, which should contain `LG.pt`. - """, - ) - - parser.add_argument( - "--beam-size", - type=int, - default=4, - help="""An integer indicating how many candidates we will keep for each - frame. Used only when --decoding-method is beam_search or - modified_beam_search.""", - ) - - parser.add_argument( - "--beam", - type=float, - default=20.0, - help="""A floating point value to calculate the cutoff score during beam - search (i.e., `cutoff = max-score - beam`), which is the same as the - `beam` in Kaldi. - Used only when --decoding-method is fast_beam_search, - fast_beam_search_nbest, fast_beam_search_nbest_LG, - and fast_beam_search_nbest_oracle - """, - ) - - parser.add_argument( - "--ngram-lm-scale", - type=float, - default=0.01, - help=""" - Used only when --decoding-method is fast_beam_search_nbest_LG. - It specifies the scale for n-gram LM scores. - """, - ) - - parser.add_argument( - "--max-contexts", - type=int, - default=8, - help="""Used only when --decoding-method is - fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, - and fast_beam_search_nbest_oracle""", - ) - - parser.add_argument( - "--max-states", - type=int, - default=64, - help="""Used only when --decoding-method is - fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, - and fast_beam_search_nbest_oracle""", - ) - - parser.add_argument( - "--context-size", - type=int, - default=2, - help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", - ) - parser.add_argument( - "--max-sym-per-frame", - type=int, - default=1, - help="""Maximum number of symbols per frame. - Used only when --decoding-method is greedy_search""", - ) - - parser.add_argument( - "--num-paths", - type=int, - default=200, - help="""Number of paths for nbest decoding. - Used only when the decoding method is fast_beam_search_nbest, - fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", - ) - - parser.add_argument( - "--nbest-scale", - type=float, - default=0.5, - help="""Scale applied to lattice scores when computing nbest paths. - Used only when the decoding method is fast_beam_search_nbest, - fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", - ) - - parser.add_argument( - "--use-shallow-fusion", - type=str2bool, - default=False, - help="""Use neural network LM for shallow fusion. - If you want to use LODR, you will also need to set this to true - """, - ) - - parser.add_argument( - "--lm-type", - type=str, - default="rnn", - help="Type of NN lm", - choices=["rnn", "transformer"], - ) - - parser.add_argument( - "--lm-scale", - type=float, - default=0.3, - help="""The scale of the neural network LM - Used only when `--use-shallow-fusion` is set to True. - """, - ) - - parser.add_argument( - "--tokens-ngram", - type=int, - default=2, - help="""The order of the ngram lm. - """, - ) - - parser.add_argument( - "--backoff-id", - type=int, - default=500, - help="ID of the backoff symbol in the ngram LM", - ) - - parser.add_argument( - "--context-score", - type=float, - default=2, - help=""" - The bonus score of each token for the context biasing words/phrases. - Used only when --decoding-method is modified_beam_search and - modified_beam_search_LODR. - """, - ) - - parser.add_argument( - "--context-file", - type=str, - default="", - help=""" - The path of the context biasing lists, one word/phrase each line - Used only when --decoding-method is modified_beam_search and - modified_beam_search_LODR. - """, - ) - - add_model_arguments(parser) - - return parser - - -def decode_one_batch( - params: AttributeDict, - model: nn.Module, - sp: spm.SentencePieceProcessor, - batch: dict, - word_table: Optional[k2.SymbolTable] = None, - decoding_graph: Optional[k2.Fsa] = None, - context_graph: Optional[ContextGraph] = None, - LM: Optional[LmScorer] = None, - ngram_lm=None, - ngram_lm_scale: float = 0.0, -) -> Dict[str, List[List[str]]]: - """Decode one batch and return the result in a dict. The dict has the - following format: - - - key: It indicates the setting used for decoding. For example, - if greedy_search is used, it would be "greedy_search" - If beam search with a beam size of 7 is used, it would be - "beam_7" - - value: It contains the decoding result. `len(value)` equals to - batch size. `value[i]` is the decoding result for the i-th - utterance in the given batch. - Args: - params: - It's the return value of :func:`get_params`. - model: - The neural model. - sp: - The BPE model. - batch: - It is the return value from iterating - `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation - for the format of the `batch`. - word_table: - The word symbol table. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used - only when --decoding-method is fast_beam_search, fast_beam_search_nbest, - fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. - LM: - A neural network language model. - ngram_lm: - A ngram language model - ngram_lm_scale: - The scale for the ngram language model. - Returns: - Return the decoding result. See above description for the format of - the returned dict. - """ - device = next(model.parameters()).device - audio = batch["audio"].to(device) - padding_mask = batch["padding_mask"].to(device) - - encoder_out, encoder_out_lens = model.forward_encoder(audio, padding_mask) - - 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_LG": - hyp_tokens = fast_beam_search_nbest_LG( - model=model, - decoding_graph=decoding_graph, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam, - max_contexts=params.max_contexts, - max_states=params.max_states, - num_paths=params.num_paths, - nbest_scale=params.nbest_scale, - ) - for hyp in hyp_tokens: - hyps.append([word_table[i] for i in hyp]) - elif params.decoding_method == "fast_beam_search_nbest": - hyp_tokens = fast_beam_search_nbest( - model=model, - decoding_graph=decoding_graph, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam, - max_contexts=params.max_contexts, - max_states=params.max_states, - num_paths=params.num_paths, - nbest_scale=params.nbest_scale, - ) - for hyp in sp.decode(hyp_tokens): - hyps.append(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(batch["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, - context_graph=context_graph, - ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) - elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": - hyp_tokens = modified_beam_search_lm_shallow_fusion( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LM=LM, - ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) - elif params.decoding_method == "modified_beam_search_LODR": - hyp_tokens = modified_beam_search_LODR( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LODR_lm=ngram_lm, - LODR_lm_scale=ngram_lm_scale, - LM=LM, - context_graph=context_graph, - ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) - elif params.decoding_method == "modified_beam_search_lm_rescore": - lm_scale_list = [0.01 * i for i in range(10, 50)] - ans_dict = modified_beam_search_lm_rescore( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LM=LM, - lm_scale_list=lm_scale_list, - ) - elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": - lm_scale_list = [0.02 * i for i in range(2, 30)] - ans_dict = modified_beam_search_lm_rescore_LODR( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LM=LM, - LODR_lm=ngram_lm, - sp=sp, - lm_scale_list=lm_scale_list, - ) - 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}" - if "LG" in params.decoding_method: - key += f"_ngram_lm_scale_{params.ngram_lm_scale}" - - return {key: hyps} - elif "modified_beam_search" in params.decoding_method: - prefix = f"beam_size_{params.beam_size}" - if params.decoding_method in ( - "modified_beam_search_lm_rescore", - "modified_beam_search_lm_rescore_LODR", - ): - ans = dict() - assert ans_dict is not None - for key, hyps in ans_dict.items(): - hyps = [sp.decode(hyp).split() for hyp in hyps] - ans[f"{prefix}_{key}"] = hyps - return ans - else: - if params.has_contexts: - prefix += f"-context-score-{params.context_score}" - return {prefix: hyps} - else: - return {f"beam_size_{params.beam_size}": hyps} - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - params: AttributeDict, - model: nn.Module, - sp: spm.SentencePieceProcessor, - word_table: Optional[k2.SymbolTable] = None, - decoding_graph: Optional[k2.Fsa] = None, - context_graph: Optional[ContextGraph] = None, - LM: Optional[LmScorer] = None, - ngram_lm=None, - ngram_lm_scale: float = 0.0, -) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: - """Decode dataset. - - Args: - dl: - PyTorch's dataloader containing the dataset to decode. - params: - It is returned by :func:`get_params`. - model: - The neural model. - sp: - The BPE model. - word_table: - The word symbol table. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used - only when --decoding-method is fast_beam_search, fast_beam_search_nbest, - fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. - Returns: - Return a dict, whose key may be "greedy_search" if greedy search - is used, or it may be "beam_7" if beam size of 7 is used. - Its value is a list of tuples. Each tuple contains two elements: - The first is the reference transcript, and the second is the - predicted result. - """ - num_cuts = 0 - - try: - num_batches = len(dl) - except TypeError: - num_batches = "?" - - if params.decoding_method == "greedy_search": - log_interval = 50 - else: - log_interval = 20 - - results = defaultdict(list) - for batch_idx, batch in enumerate(dl): - texts = batch["supervisions"]["text"] - cut_ids = [cut.id for cut in batch["cuts"]] - - hyps_dict = decode_one_batch( - params=params, - model=model, - sp=sp, - decoding_graph=decoding_graph, - context_graph=context_graph, - word_table=word_table, - batch=batch, - LM=LM, - ngram_lm=ngram_lm, - ngram_lm_scale=ngram_lm_scale, - ) - - for name, hyps in hyps_dict.items(): - this_batch = [] - assert len(hyps) == len(texts) - for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): - ref_words = ref_text.split() - this_batch.append((cut_id, ref_words, hyp_words)) - - results[name].extend(this_batch) - - num_cuts += len(texts) - - if batch_idx % log_interval == 0: - batch_str = f"{batch_idx}/{num_batches}" - - logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") - return results - - -def save_results( - params: AttributeDict, - test_set_name: str, - results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], -): - test_set_wers = dict() - for key, results in results_dict.items(): - recog_path = ( - params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" - ) - results = sorted(results) - store_transcripts(filename=recog_path, texts=results) - logging.info(f"The transcripts are stored in {recog_path}") - - # The following prints out WERs, per-word error statistics and aligned - # ref/hyp pairs. - errs_filename = ( - params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" - ) - with open(errs_filename, "w") as f: - wer = write_error_stats( - f, f"{test_set_name}-{key}", results, enable_log=True - ) - test_set_wers[key] = wer - - logging.info("Wrote detailed error stats to {}".format(errs_filename)) - - test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) - errs_info = ( - params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" - ) - with open(errs_info, "w") as f: - print("settings\tWER", file=f) - for key, val in test_set_wers: - print("{}\t{}".format(key, val), file=f) - - s = "\nFor {}, WER of different settings are:\n".format(test_set_name) - note = "\tbest for {}".format(test_set_name) - for key, val in test_set_wers: - s += "{}\t{}{}\n".format(key, val, note) - note = "" - logging.info(s) - - -@torch.no_grad() -def main(): - parser = get_parser() - LibriSpeechAsrDataModule.add_arguments(parser) - LmScorer.add_arguments(parser) - args = parser.parse_args() - args.exp_dir = Path(args.exp_dir) - - params = get_params() - params.update(vars(args)) - - assert params.decoding_method in ( - "greedy_search", - "beam_search", - "fast_beam_search", - "fast_beam_search_nbest", - "fast_beam_search_nbest_LG", - "fast_beam_search_nbest_oracle", - "modified_beam_search", - "modified_beam_search_LODR", - "modified_beam_search_lm_shallow_fusion", - "modified_beam_search_lm_rescore", - "modified_beam_search_lm_rescore_LODR", - ) - params.res_dir = params.exp_dir / params.decoding_method - - if os.path.exists(params.context_file): - params.has_contexts = True - else: - params.has_contexts = False - - if params.iter > 0: - params.suffix = f"iter-{params.iter}-avg-{params.avg}" - else: - params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" - - if "fast_beam_search" in params.decoding_method: - params.suffix += f"-beam-{params.beam}" - params.suffix += f"-max-contexts-{params.max_contexts}" - params.suffix += f"-max-states-{params.max_states}" - if "nbest" in params.decoding_method: - params.suffix += f"-nbest-scale-{params.nbest_scale}" - params.suffix += f"-num-paths-{params.num_paths}" - if "LG" in params.decoding_method: - params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" - elif "beam_search" in params.decoding_method: - params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" - if params.decoding_method in ( - "modified_beam_search", - "modified_beam_search_LODR", - ): - if params.has_contexts: - params.suffix += f"-context-score-{params.context_score}" - else: - params.suffix += f"-context-{params.context_size}" - params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" - - if params.use_shallow_fusion: - params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" - - if "LODR" in params.decoding_method: - params.suffix += ( - f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" - ) - - 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() - - # only load the neural network LM if required - if params.use_shallow_fusion or params.decoding_method in ( - "modified_beam_search_lm_rescore", - "modified_beam_search_lm_rescore_LODR", - "modified_beam_search_lm_shallow_fusion", - "modified_beam_search_LODR", - ): - LM = LmScorer( - lm_type=params.lm_type, - params=params, - device=device, - lm_scale=params.lm_scale, - ) - LM.to(device) - LM.eval() - else: - LM = None - - # only load N-gram LM when needed - if params.decoding_method == "modified_beam_search_lm_rescore_LODR": - try: - import kenlm - except ImportError: - print("Please install kenlm first. You can use") - print(" pip install https://github.com/kpu/kenlm/archive/master.zip") - print("to install it") - import sys - - sys.exit(-1) - ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") - logging.info(f"lm filename: {ngram_file_name}") - ngram_lm = kenlm.Model(ngram_file_name) - ngram_lm_scale = None # use a list to search - - elif params.decoding_method == "modified_beam_search_LODR": - lm_filename = f"{params.tokens_ngram}gram.fst.txt" - logging.info(f"Loading token level lm: {lm_filename}") - ngram_lm = NgramLm( - str(params.lang_dir / lm_filename), - backoff_id=params.backoff_id, - is_binary=False, - ) - logging.info(f"num states: {ngram_lm.lm.num_states}") - ngram_lm_scale = params.ngram_lm_scale - else: - ngram_lm = None - ngram_lm_scale = None - - if "fast_beam_search" in params.decoding_method: - if params.decoding_method == "fast_beam_search_nbest_LG": - lexicon = Lexicon(params.lang_dir) - word_table = lexicon.word_table - lg_filename = params.lang_dir / "LG.pt" - logging.info(f"Loading {lg_filename}") - decoding_graph = k2.Fsa.from_dict( - torch.load(lg_filename, map_location=device) - ) - decoding_graph.scores *= params.ngram_lm_scale - else: - word_table = None - decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) - else: - decoding_graph = None - word_table = None - - if "modified_beam_search" in params.decoding_method: - if os.path.exists(params.context_file): - contexts = [] - for line in open(params.context_file).readlines(): - contexts.append((sp.encode(line.strip()), 0.0)) - context_graph = ContextGraph(params.context_score) - context_graph.build(contexts) - else: - context_graph = None - else: - context_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 - librispeech = LibriSpeechAsrDataModule(args) - - dev_clean_cuts = librispeech.dev_clean_cuts() - dev_other_cuts = librispeech.dev_other_cuts() - - dev_clean_dl = librispeech.test_dataloaders( - dev_clean_cuts, - do_normalize=params.do_normalize, - ) - dev_other_dl = librispeech.test_dataloaders( - dev_other_cuts, - do_normalize=params.do_normalize, - ) - - test_clean_cuts = librispeech.test_clean_cuts() - test_other_cuts = librispeech.test_other_cuts() - - test_clean_dl = librispeech.test_dataloaders( - test_clean_cuts, - do_normalize=params.do_normalize, - ) - test_other_dl = librispeech.test_dataloaders( - test_other_cuts, - do_normalize=params.do_normalize, - ) - - test_sets = ["dev-clean", "dev-other", "test-clean", "test-other"] - test_dl = [dev_clean_dl, dev_other_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, - word_table=word_table, - decoding_graph=decoding_graph, - context_graph=context_graph, - LM=LM, - ngram_lm=ngram_lm, - ngram_lm_scale=ngram_lm_scale, - ) - - save_results( - params=params, - test_set_name=test_set, - results_dict=results_dict, - ) - - logging.info("Done!") - - -if __name__ == "__main__": - main() From d4a5c404bff18b235e7123d1cf2c0974b7ecc25a Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 7 Sep 2024 23:11:00 +0800 Subject: [PATCH 14/24] Delete egs/librilight/SSL/zipformer/asr_datamodule.py --- egs/librilight/SSL/zipformer/asr_datamodule.py | 1 - 1 file changed, 1 deletion(-) delete mode 120000 egs/librilight/SSL/zipformer/asr_datamodule.py diff --git a/egs/librilight/SSL/zipformer/asr_datamodule.py b/egs/librilight/SSL/zipformer/asr_datamodule.py deleted file mode 120000 index b9313bffce..0000000000 --- a/egs/librilight/SSL/zipformer/asr_datamodule.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/SSL/zipformer/asr_datamodule.py \ No newline at end of file From 2e52cbf1fec364d12c6bbd019ac10c969e767a03 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 7 Sep 2024 23:12:07 +0800 Subject: [PATCH 15/24] Update finetune_ctc.py --- egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py index d5dd8d71f4..572e040bd9 100755 --- a/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py +++ b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py @@ -93,7 +93,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + ) + 100000 def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: From 2d3452fa320f6fbb9241ba4f2623497d633e4250 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 7 Sep 2024 23:12:38 +0800 Subject: [PATCH 16/24] Update finetune.py --- egs/librispeech/SSL/zipformer/finetune.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librispeech/SSL/zipformer/finetune.py b/egs/librispeech/SSL/zipformer/finetune.py index 2e521f1772..336c358136 100755 --- a/egs/librispeech/SSL/zipformer/finetune.py +++ b/egs/librispeech/SSL/zipformer/finetune.py @@ -99,7 +99,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + ) + 100000 def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: From 19cc5bab65497daf7d2cfe590fe9f7613c8ee4c6 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 7 Sep 2024 23:13:15 +0800 Subject: [PATCH 17/24] Update finetune_ce.py --- egs/librispeech/SSL/hubert/finetune_ce.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librispeech/SSL/hubert/finetune_ce.py b/egs/librispeech/SSL/hubert/finetune_ce.py index 1081313f1d..cec42ea123 100755 --- a/egs/librispeech/SSL/hubert/finetune_ce.py +++ b/egs/librispeech/SSL/hubert/finetune_ce.py @@ -99,7 +99,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + ) + 100000 def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: From f05b3b17018f55fe1652d16a8f4d86fa75452369 Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Sat, 7 Sep 2024 23:13:41 +0800 Subject: [PATCH 18/24] Update finetune.py --- egs/librispeech/SSL/hubert/finetune.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librispeech/SSL/hubert/finetune.py b/egs/librispeech/SSL/hubert/finetune.py index 05b942f632..582771dee8 100755 --- a/egs/librispeech/SSL/hubert/finetune.py +++ b/egs/librispeech/SSL/hubert/finetune.py @@ -99,7 +99,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + ) + 100000 def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: From b35924f361477e782feed83a048015c4b72360d6 Mon Sep 17 00:00:00 2001 From: Yifan Yang Date: Sat, 7 Sep 2024 23:41:15 +0800 Subject: [PATCH 19/24] small fix --- egs/librilight/SSL/prepare.sh | 19 +++++++++++++------ egs/librilight/SSL/zipformer/pretrain.py | 5 ++++- 2 files changed, 17 insertions(+), 7 deletions(-) diff --git a/egs/librilight/SSL/prepare.sh b/egs/librilight/SSL/prepare.sh index 804da9d6eb..4e3d00364c 100755 --- a/egs/librilight/SSL/prepare.sh +++ b/egs/librilight/SSL/prepare.sh @@ -6,9 +6,9 @@ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python set -eou pipefail nj=15 -# run step 0 to step 5 by default +# run step 0 to step 4 by default stage=0 -stop_stage=5 +stop_stage=4 # We assume dl_dir (download dir) contains the following # directories and files. If not, they will be downloaded @@ -79,9 +79,16 @@ fi if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then log "Stage 4: Extract SSL target for librilight" - mkdir -p data/fbank - if [ ! -e data/fbank/.librispeech.done ]; then - ./local/compute_fbank_librispeech.py - touch data/fbank/.librispeech.done + if [ ! -e data/kmeans/.extract_small.done ]; then + ./local/extract_kmeans_from_hubert_base.py --subset small + touch data/kmeans/.extract_small.done + fi + if [ ! -e data/kmeans/.extract_medium.done ]; then + ./local/extract_kmeans_from_hubert_base.py --subset medium + touch data/kmeans/.extract_medium.done + fi + if [ ! -e data/kmeans/.extract_large.done ]; then + ./local/extract_kmeans_from_hubert_base.py --subset large + touch data/kmeans/.extract_large.done fi fi diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py index 10dd68ea91..55096e4d33 100755 --- a/egs/librilight/SSL/zipformer/pretrain.py +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -953,7 +953,10 @@ def save_bad_model(suffix: str = ""): 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 + params.batch_idx_train + * params.max_duration + * params.world_size + / 3600 ) scaler.step(optimizer) From 8c257a38dc0c0730fa46880fef90d9a7cc9909fd Mon Sep 17 00:00:00 2001 From: Yifan Yang Date: Sun, 8 Sep 2024 13:28:13 +0800 Subject: [PATCH 20/24] fix isort --- egs/librilight/SSL/zipformer/ssl_datamodule.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librilight/SSL/zipformer/ssl_datamodule.py b/egs/librilight/SSL/zipformer/ssl_datamodule.py index 1c6d74a4a6..e32bedaf61 100644 --- a/egs/librilight/SSL/zipformer/ssl_datamodule.py +++ b/egs/librilight/SSL/zipformer/ssl_datamodule.py @@ -23,8 +23,8 @@ from pathlib import Path from typing import Any, Dict, Optional -import torch import lhotse +import torch from dataset import HubertDataset from lhotse import CutSet, load_manifest_lazy from lhotse.dataset import DynamicBucketingSampler, SimpleCutSampler From 84f8adff32e15e8612b6d767b9b49c9fcaa055ff Mon Sep 17 00:00:00 2001 From: Your Name Date: Wed, 23 Oct 2024 00:10:24 -0700 Subject: [PATCH 21/24] update prepare.sh --- .../local/extract_kmeans_from_hubert_base.py | 35 +++---------------- egs/librilight/SSL/prepare.sh | 12 +++++-- 2 files changed, 14 insertions(+), 33 deletions(-) diff --git a/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py b/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py index e85a629d50..a3b01b2311 100755 --- a/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py +++ b/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py @@ -26,7 +26,6 @@ import torch from lhotse import CutSet, SupervisionSegment from lhotse.utils import fastcopy -from silero_vad import get_speech_timestamps, load_silero_vad from tqdm import tqdm # Torch's multithreaded behavior needs to be disabled or @@ -82,7 +81,7 @@ def get_args(): parser.add_argument( "--kmeans-model-path", type=str, - default="download/hubert_base_ls960_L9_km500.model", + default="download/hubert_base_ls960_L9_km500.bin", ) parser.add_argument( @@ -103,7 +102,7 @@ def get_args(): def extract_and_save_one_cuts( - raw_cuts_path, cuts_path, model, vad_model, apply_kmeans, do_normalize, device + raw_cuts_path, cuts_path, model, apply_kmeans, do_normalize, device ): logging.info(f"Loading {raw_cuts_path}") cut_set = CutSet.from_file(raw_cuts_path) @@ -111,20 +110,11 @@ def extract_and_save_one_cuts( logging.info("Extracting kmeans") cuts = [] for cut in tqdm(cut_set): - assert cut.sampling_rate == 16000, f"{cut.sampling_rate}" + assert cut.sampling_rate == 16000, f"Sampling rate: {cut.sampling_rate}" audio = cut.load_audio() - if audio.shape[-1] > 64 * 16000: - timestamps = get_speech_timestamps(audio, vad_model) - offsets = [i["start"] for i in timestamps] - audios = [audio[:, i["start"] : i["end"]] for i in timestamps] - logging.info(f"Trim audio {cut.id} into {len(audios)} segments") - else: - offsets = [0] - audios = [audio] - - seq = 0 - for audio, offset in zip(audios, offsets): + offsets = 0 + if True: x = torch.from_numpy(audio).float().to(device) with torch.no_grad(): @@ -141,24 +131,12 @@ def extract_and_save_one_cuts( kmeans = " ".join(map(str, apply_kmeans(feature).tolist())) - supervision_segment = fastcopy( - cut.supervisions[0], - id=f"{cut.id}-{seq}", - start=0.0, - duration=audio.shape[-1] / 16000, - ) cut_with_kmeans = fastcopy( cut, - id=f"{cut.id}-{seq}", - start=cut.start + offset / 16000, - duration=audio.shape[-1] / 16000, - supervisions=[supervision_segment], custom={"kmeans": kmeans}, ) cuts.append(cut_with_kmeans) - seq += 1 - cuts = CutSet(cuts) logging.info(f"Saving to {cuts_path}") @@ -181,7 +159,6 @@ def extract_kmeans(args): prefix = "librilight" - vad_model = load_silero_vad() apply_kmeans = ApplyKmeans(args.kmeans_model_path) model, _, task = fairseq.checkpoint_utils.load_model_ensemble_and_task( [args.model_path] @@ -204,7 +181,6 @@ def extract_kmeans(args): raw_cuts_path, cuts_path, model, - vad_model, apply_kmeans, do_normalize, device, @@ -235,7 +211,6 @@ def extract_kmeans(args): raw_cuts_path, cuts_path, model, - vad_model, apply_kmeans, do_normalize, device, diff --git a/egs/librilight/SSL/prepare.sh b/egs/librilight/SSL/prepare.sh index 4e3d00364c..24bb9b8dee 100755 --- a/egs/librilight/SSL/prepare.sh +++ b/egs/librilight/SSL/prepare.sh @@ -5,7 +5,7 @@ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python set -eou pipefail -nj=15 +nj=32 # run step 0 to step 4 by default stage=0 stop_stage=4 @@ -58,13 +58,13 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then mkdir -p data/kmeans if [ ! -f data/kmeans/.preprocess_complete ]; then python3 ./local/preprocess_librilight.py - touch data/fbank/.preprocess_complete + touch data/kmeans/.preprocess_complete fi fi if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then log "Stage 3: Split medium and large subset into pieces" - num_per_split=200000 + num_per_split=2500 split_dir=data/kmeans/medium_split if [ ! -f $split_dir/.split_completed ]; then lhotse split-lazy ./data/kmeans/librilight_cuts_medium_raw.jsonl.gz $split_dir $num_per_split @@ -79,6 +79,12 @@ fi if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then log "Stage 4: Extract SSL target for librilight" + if [ ! -e download/hubert_base_ls960.pt ]; then + wget https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960.pt -P download + fi + if [ ! -e download/hubert_base_ls960_L9_km500.bin ]; then + wget https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960_L9_km500.bin -P download + fi if [ ! -e data/kmeans/.extract_small.done ]; then ./local/extract_kmeans_from_hubert_base.py --subset small touch data/kmeans/.extract_small.done From a6a80896d5a491639d8443429ed906a486fa3038 Mon Sep 17 00:00:00 2001 From: Your Name Date: Wed, 23 Oct 2024 06:03:12 -0700 Subject: [PATCH 22/24] add sliding window --- ..._from_hubert_base.py => extract_kmeans.py} | 104 ++++++++++++++---- egs/librilight/SSL/prepare.sh | 6 +- 2 files changed, 85 insertions(+), 25 deletions(-) rename egs/librilight/SSL/local/{extract_kmeans_from_hubert_base.py => extract_kmeans.py} (72%) diff --git a/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py b/egs/librilight/SSL/local/extract_kmeans.py similarity index 72% rename from egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py rename to egs/librilight/SSL/local/extract_kmeans.py index a3b01b2311..b8f6fcba9e 100755 --- a/egs/librilight/SSL/local/extract_kmeans_from_hubert_base.py +++ b/egs/librilight/SSL/local/extract_kmeans.py @@ -17,6 +17,7 @@ import argparse import logging +import math from pathlib import Path from typing import Optional @@ -98,44 +99,90 @@ def get_args(): help="Stop processing pieces until this number (exclusive).", ) + parser.add_argument( + "--window-duration", + type=float, + default=300.0, + ) + + parser.add_argument( + "--shift-duration", + type=float, + default=250.0, + ) + return parser.parse_args() +@torch.no_grad() def extract_and_save_one_cuts( - raw_cuts_path, cuts_path, model, apply_kmeans, do_normalize, device + raw_cuts_path, + cuts_path, + model, + apply_kmeans, + do_normalize, + window_duration, + shift_duration, ): logging.info(f"Loading {raw_cuts_path}") cut_set = CutSet.from_file(raw_cuts_path) logging.info("Extracting kmeans") cuts = [] + + assert window_duration >= shift_duration + window_size = int(window_duration * 16000) + shift_size = int(shift_duration * 16000) + overlap_size = window_size - shift_size + out_overlap_size = get_out_length(overlap_size) + for cut in tqdm(cut_set): assert cut.sampling_rate == 16000, f"Sampling rate: {cut.sampling_rate}" + audio = cut.load_audio() - offsets = 0 - if True: - x = torch.from_numpy(audio).float().to(device) + T = audio.shape[1] + start = 0 + kmeans = [] + while start < T: + real_window_size = min(window_size, T - start) + audio_window = audio[:, start : start + real_window_size] + + x = ( + torch.from_numpy(audio_window) + .float() + .to(next(model.parameters()).device) + ) + if do_normalize: + x = torch.nn.functional.layer_norm(x, x.shape) + + feature, _ = model.extract_features( + source=x, + padding_mask=None, + mask=False, + output_layer=9, + ) + feature = feature.squeeze(0) - with torch.no_grad(): - if do_normalize: - x = torch.nn.functional.layer_norm(x, x.shape) + current_kmeans = apply_kmeans(feature).tolist() - feature, _ = model.extract_features( - source=x, - padding_mask=None, - mask=False, - output_layer=9, - ) - feature = feature.squeeze(0) + if start == 0: + kmeans.extend(current_kmeans) + else: + kmeans.extend(current_kmeans[out_overlap_size:]) - kmeans = " ".join(map(str, apply_kmeans(feature).tolist())) + if T - start <= window_size: + break - cut_with_kmeans = fastcopy( - cut, - custom={"kmeans": kmeans}, - ) - cuts.append(cut_with_kmeans) + start += shift_size + + kmeans = " ".join(map(str, kmeans)) + + cut_with_kmeans = fastcopy( + cut, + custom={"kmeans": kmeans}, + ) + cuts.append(cut_with_kmeans) cuts = CutSet(cuts) @@ -166,6 +213,9 @@ def extract_kmeans(args): model = model[0].eval().to(device) do_normalize = task.cfg.normalize + window_duration = args.window_duration + shift_duration = args.shift_duration + if args.subset == "small": cuts_path = output_dir / f"{prefix}_cuts_{args.subset}.jsonl.gz" if cuts_path.is_file(): @@ -183,7 +233,8 @@ def extract_kmeans(args): model, apply_kmeans, do_normalize, - device, + window_duration, + shift_duration, ) else: num_digits = 8 # num_digits is fixed by lhotse split-lazy @@ -213,10 +264,19 @@ def extract_kmeans(args): model, apply_kmeans, do_normalize, - device, + window_duration, + shift_duration, ) +def get_out_length(T): + conv_layers = [(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512, 2, 2)] * 2 + for i, (out_channels, kernel_size, stride) in enumerate(conv_layers): + T = math.floor((T - kernel_size) / stride) + 1 + + return max(0, T) + + if __name__ == "__main__": formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" diff --git a/egs/librilight/SSL/prepare.sh b/egs/librilight/SSL/prepare.sh index 24bb9b8dee..e0a293b2c9 100755 --- a/egs/librilight/SSL/prepare.sh +++ b/egs/librilight/SSL/prepare.sh @@ -86,15 +86,15 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then wget https://dl.fbaipublicfiles.com/hubert/hubert_base_ls960_L9_km500.bin -P download fi if [ ! -e data/kmeans/.extract_small.done ]; then - ./local/extract_kmeans_from_hubert_base.py --subset small + ./local/extract_kmeans.py --subset small touch data/kmeans/.extract_small.done fi if [ ! -e data/kmeans/.extract_medium.done ]; then - ./local/extract_kmeans_from_hubert_base.py --subset medium + ./local/extract_kmeans.py --subset medium touch data/kmeans/.extract_medium.done fi if [ ! -e data/kmeans/.extract_large.done ]; then - ./local/extract_kmeans_from_hubert_base.py --subset large + ./local/extract_kmeans.py --subset large touch data/kmeans/.extract_large.done fi fi From ce72b34c30ed9fb6d2b8c5323f72b1392f61c27a Mon Sep 17 00:00:00 2001 From: Your Name Date: Thu, 24 Oct 2024 00:43:16 -0700 Subject: [PATCH 23/24] update --- egs/librilight/SSL/local/extract_kmeans.py | 3 +++ egs/librilight/SSL/prepare.sh | 2 +- 2 files changed, 4 insertions(+), 1 deletion(-) diff --git a/egs/librilight/SSL/local/extract_kmeans.py b/egs/librilight/SSL/local/extract_kmeans.py index b8f6fcba9e..9b101a4822 100755 --- a/egs/librilight/SSL/local/extract_kmeans.py +++ b/egs/librilight/SSL/local/extract_kmeans.py @@ -18,6 +18,7 @@ import argparse import logging import math +import os from pathlib import Path from typing import Optional @@ -36,6 +37,8 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) +os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:512" + class ApplyKmeans(object): def __init__(self, km_path): diff --git a/egs/librilight/SSL/prepare.sh b/egs/librilight/SSL/prepare.sh index e0a293b2c9..160c66c7da 100755 --- a/egs/librilight/SSL/prepare.sh +++ b/egs/librilight/SSL/prepare.sh @@ -64,7 +64,7 @@ fi if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then log "Stage 3: Split medium and large subset into pieces" - num_per_split=2500 + num_per_split=10000 split_dir=data/kmeans/medium_split if [ ! -f $split_dir/.split_completed ]; then lhotse split-lazy ./data/kmeans/librilight_cuts_medium_raw.jsonl.gz $split_dir $num_per_split From 1b89c6dac4ed4ad472917e0cd873733ba1ed4fdd Mon Sep 17 00:00:00 2001 From: Your Name <> Date: Mon, 28 Oct 2024 22:58:40 -0700 Subject: [PATCH 24/24] skipping batch counts hurts performance --- egs/librispeech/SSL/hubert/finetune.py | 2 +- egs/librispeech/SSL/hubert/finetune_ce.py | 2 +- egs/librispeech/SSL/zipformer/finetune.py | 2 +- egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py | 2 +- 4 files changed, 4 insertions(+), 4 deletions(-) diff --git a/egs/librispeech/SSL/hubert/finetune.py b/egs/librispeech/SSL/hubert/finetune.py index 582771dee8..05b942f632 100755 --- a/egs/librispeech/SSL/hubert/finetune.py +++ b/egs/librispeech/SSL/hubert/finetune.py @@ -99,7 +99,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + 100000 + ) def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: diff --git a/egs/librispeech/SSL/hubert/finetune_ce.py b/egs/librispeech/SSL/hubert/finetune_ce.py index cec42ea123..1081313f1d 100755 --- a/egs/librispeech/SSL/hubert/finetune_ce.py +++ b/egs/librispeech/SSL/hubert/finetune_ce.py @@ -99,7 +99,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + 100000 + ) def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: diff --git a/egs/librispeech/SSL/zipformer/finetune.py b/egs/librispeech/SSL/zipformer/finetune.py index 336c358136..2e521f1772 100755 --- a/egs/librispeech/SSL/zipformer/finetune.py +++ b/egs/librispeech/SSL/zipformer/finetune.py @@ -99,7 +99,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + 100000 + ) def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: diff --git a/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py index 572e040bd9..d5dd8d71f4 100755 --- a/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py +++ b/egs/librispeech/SSL/zipformer_ctc/finetune_ctc.py @@ -93,7 +93,7 @@ def get_adjusted_batch_count(params: AttributeDict) -> float: * params.accum_grad * (params.max_duration * params.world_size) / params.ref_duration - ) + 100000 + ) def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: