diff --git a/README.md b/README.md index ff93e8fad..f7aed9dc3 100644 --- a/README.md +++ b/README.md @@ -107,6 +107,17 @@ The best CER we currently have is: We provide a Colab notebook to run a pre-trained conformer CTC model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1WnG17io5HEZ0Gn_cnh_VzK5QYOoiiklC?usp=sharing) +#### Transducer Stateless Model + +The best CER we currently have is: + +| | test | +|-----|------| +| CER | 5.7 | + + +We provide a Colab notebook to run a pre-trained TransducerStateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/14XaT2MhnBkK-3_RqqWq3K90Xlbin-GZC?usp=sharing) + #### TDNN LSTM CTC Model The CER for this model is: diff --git a/egs/aishell/ASR/RESULTS.md b/egs/aishell/ASR/RESULTS.md index 465f8ce85..dd27e1f35 100644 --- a/egs/aishell/ASR/RESULTS.md +++ b/egs/aishell/ASR/RESULTS.md @@ -1,5 +1,37 @@ ## Results +### Aishell training results (Transducer-stateless) +#### 2021-12-29 +(Pingfeng Luo) : The tensorboard log for training is available at + +||test| +|--|--| +|CER| 5.7% | + +You can use the following commands to reproduce our results: + +```bash +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8" +./transducer_stateless/train.py \ + --bucketing-sampler True \ + --world-size 8 \ + --lang-dir data/lang_char \ + --num-epochs 40 \ + --start-epoch 0 \ + --exp-dir transducer_stateless/exp_char \ + --max-duration 160 \ + --lr-factor 3 + +./transducer_stateless/decode.py \ + --epoch 39 \ + --avg 10 \ + --lang-dir data/lang_char \ + --exp-dir transducer_stateless/exp_char \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 +``` + ### Aishell training results (Conformer-MMI) #### 2021-12-04 (Pingfeng Luo): Result of diff --git a/egs/aishell/ASR/conformer_ctc/decode.py b/egs/aishell/ASR/conformer_ctc/decode.py index dc593eeb9..c38c4c65f 100755 --- a/egs/aishell/ASR/conformer_ctc/decode.py +++ b/egs/aishell/ASR/conformer_ctc/decode.py @@ -538,9 +538,13 @@ def main(): logging.info(f"Number of model parameters: {num_param}") aishell = AishellAsrDataModule(args) + test_cuts = aishell.test_cuts() + test_dl = aishell.test_dataloaders(test_cuts) test_sets = ["test"] - for test_set, test_dl in zip(test_sets, aishell.test_dataloaders()): + test_dls = [test_dl] + + for test_set, test_dl in zip(test_sets, test_dls): results_dict = decode_dataset( dl=test_dl, params=params, diff --git a/egs/aishell/ASR/conformer_ctc/subsampling.py b/egs/aishell/ASR/conformer_ctc/subsampling.py index 720ed6c22..542fb0364 100644 --- a/egs/aishell/ASR/conformer_ctc/subsampling.py +++ b/egs/aishell/ASR/conformer_ctc/subsampling.py @@ -22,8 +22,8 @@ import torch.nn as nn class Conv2dSubsampling(nn.Module): """Convolutional 2D subsampling (to 1/4 length). - Convert an input of shape [N, T, idim] to an output - with shape [N, T', odim], where + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where T' = ((T-1)//2 - 1)//2, which approximates T' == T//4 It is based on @@ -34,10 +34,10 @@ class Conv2dSubsampling(nn.Module): """ Args: idim: - Input dim. The input shape is [N, T, idim]. + Input dim. The input shape is (N, T, idim). Caution: It requires: T >=7, idim >=7 odim: - Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim] + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim) """ assert idim >= 7 super().__init__() @@ -58,18 +58,18 @@ class Conv2dSubsampling(nn.Module): Args: x: - Its shape is [N, T, idim]. + Its shape is (N, T, idim). Returns: - Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim] + Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim) """ - # On entry, x is [N, T, idim] - x = x.unsqueeze(1) # [N, T, idim] -> [N, 1, T, idim] i.e., [N, C, H, W] + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) x = self.conv(x) - # Now x is of shape [N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2] + # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2) b, c, t, f = x.size() x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) - # Now x is of shape [N, ((T-1)//2 - 1))//2, odim] + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) return x @@ -80,8 +80,8 @@ class VggSubsampling(nn.Module): This paper is not 100% explicit so I am guessing to some extent, and trying to compare with other VGG implementations. - Convert an input of shape [N, T, idim] to an output - with shape [N, T', odim], where + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where T' = ((T-1)//2 - 1)//2, which approximates T' = T//4 """ @@ -93,10 +93,10 @@ class VggSubsampling(nn.Module): Args: idim: - Input dim. The input shape is [N, T, idim]. + Input dim. The input shape is (N, T, idim). Caution: It requires: T >=7, idim >=7 odim: - Output dim. The output shape is [N, ((T-1)//2 - 1)//2, odim] + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim) """ super().__init__() @@ -149,10 +149,10 @@ class VggSubsampling(nn.Module): Args: x: - Its shape is [N, T, idim]. + Its shape is (N, T, idim). Returns: - Return a tensor of shape [N, ((T-1)//2 - 1)//2, odim] + Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim) """ x = x.unsqueeze(1) x = self.layers(x) diff --git a/egs/aishell/ASR/conformer_ctc/train.py b/egs/aishell/ASR/conformer_ctc/train.py index 629d7a373..a4bc8e3bb 100755 --- a/egs/aishell/ASR/conformer_ctc/train.py +++ b/egs/aishell/ASR/conformer_ctc/train.py @@ -614,8 +614,8 @@ def run(rank, world_size, args): optimizer.load_state_dict(checkpoints["optimizer"]) aishell = AishellAsrDataModule(args) - train_dl = aishell.train_dataloaders() - valid_dl = aishell.valid_dataloaders() + train_dl = aishell.train_dataloaders(aishell.train_cuts()) + valid_dl = aishell.valid_dataloaders(aishell.valid_cuts()) for epoch in range(params.start_epoch, params.num_epochs): train_dl.sampler.set_epoch(epoch) diff --git a/egs/aishell/ASR/conformer_mmi/decode.py b/egs/aishell/ASR/conformer_mmi/decode.py index 1d0b3daad..35a7d98fc 100755 --- a/egs/aishell/ASR/conformer_mmi/decode.py +++ b/egs/aishell/ASR/conformer_mmi/decode.py @@ -557,9 +557,13 @@ def main(): logging.info(f"Number of model parameters: {num_param}") aishell = AishellAsrDataModule(args) + test_cuts = aishell.test_cuts() + test_dl = aishell.test_dataloaders(test_cuts) test_sets = ["test"] - for test_set, test_dl in zip(test_sets, aishell.test_dataloaders()): + test_dls = [test_dl] + + for test_set, test_dl in zip(test_sets, test_dls): results_dict = decode_dataset( dl=test_dl, params=params, diff --git a/egs/aishell/ASR/conformer_mmi/train.py b/egs/aishell/ASR/conformer_mmi/train.py index 14ddaf5fd..79c16d1cc 100755 --- a/egs/aishell/ASR/conformer_mmi/train.py +++ b/egs/aishell/ASR/conformer_mmi/train.py @@ -608,8 +608,9 @@ def run(rank, world_size, args): optimizer.load_state_dict(checkpoints["optimizer"]) aishell = AishellAsrDataModule(args) - train_dl = aishell.train_dataloaders() - valid_dl = aishell.valid_dataloaders() + train_cuts = aishell.train_cuts() + train_dl = aishell.train_dataloaders(train_cuts) + valid_dl = aishell.valid_dataloaders(aishell.valid_cuts()) for epoch in range(params.start_epoch, params.num_epochs): train_dl.sampler.set_epoch(epoch) diff --git a/egs/aishell/ASR/local/prepare_lang.py b/egs/aishell/ASR/local/prepare_lang.py index 0880019b3..e5ae89ec4 100755 --- a/egs/aishell/ASR/local/prepare_lang.py +++ b/egs/aishell/ASR/local/prepare_lang.py @@ -33,6 +33,7 @@ consisting of words and tokens (i.e., phones) and does the following: 5. Generate L_disambig.pt, in k2 format. """ +import argparse import math from collections import defaultdict from pathlib import Path @@ -314,8 +315,16 @@ def lexicon_to_fst( return fsa +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", type=str, help="The lang dir, data/lang_phone" + ) + return parser.parse_args() + + def main(): - out_dir = Path("data/lang_phone") + out_dir = Path(get_args().lang_dir) lexicon_filename = out_dir / "lexicon.txt" sil_token = "SIL" sil_prob = 0.5 diff --git a/egs/aishell/ASR/prepare.sh b/egs/aishell/ASR/prepare.sh index 1e78d79d9..fe8a747dc 100755 --- a/egs/aishell/ASR/prepare.sh +++ b/egs/aishell/ASR/prepare.sh @@ -124,7 +124,7 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then ./local/generate_unique_lexicon.py --lang-dir $lang_phone_dir if [ ! -f $lang_phone_dir/L_disambig.pt ]; then - ./local/prepare_lang.py + ./local/prepare_lang.py --lang-dir $lang_phone_dir fi # Train a bigram P for MMI training @@ -133,7 +133,8 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then aishell_text=$dl_dir/aishell/data_aishell/transcript/aishell_transcript_v0.8.txt aishell_train_uid=$dl_dir/aishell/data_aishell/transcript/aishell_train_uid find data/aishell/data_aishell/wav/train -name "*.wav" | sed 's/\.wav//g' | awk -F '/' '{print $NF}' > $aishell_train_uid - awk 'NR==FNR{uid[$1]=$1} NR!=FNR{if($1 in uid) print $0}' $aishell_train_uid $aishell_text | cut -d " " -f 2- > $lang_phone_dir/transcript_words.txt + awk 'NR==FNR{uid[$1]=$1} NR!=FNR{if($1 in uid) print $0}' $aishell_train_uid $aishell_text | + cut -d " " -f 2- > $lang_phone_dir/transcript_words.txt fi if [ ! -f $lang_phone_dir/transcript_tokens.txt ]; then diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py index 9075ecb7e..65caa656e 100644 --- a/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py +++ b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py @@ -19,7 +19,7 @@ import argparse import logging from functools import lru_cache from pathlib import Path -from typing import List, Union +from typing import List from lhotse import CutSet, Fbank, FbankConfig, load_manifest from lhotse.dataset import ( @@ -34,11 +34,10 @@ from lhotse.dataset import ( from lhotse.dataset.input_strategies import OnTheFlyFeatures from torch.utils.data import DataLoader -from icefall.dataset.datamodule import DataModule from icefall.utils import str2bool -class AishellAsrDataModule(DataModule): +class AishellAsrDataModule: """ DataModule for k2 ASR experiments. It assumes there is always one train and valid dataloader, @@ -52,11 +51,15 @@ class AishellAsrDataModule(DataModule): - cut concatenation, - augmentation, - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. """ + def __init__(self, args: argparse.Namespace): + self.args = args + @classmethod def add_arguments(cls, parser: argparse.ArgumentParser): - super().add_arguments(parser) group = parser.add_argument_group( title="ASR data related options", description="These options are used for the preparation of " @@ -65,7 +68,7 @@ class AishellAsrDataModule(DataModule): "augmentations, etc.", ) group.add_argument( - "--feature-dir", + "--manifest-dir", type=Path, default=Path("data/fbank"), help="Path to directory with train/valid/test cuts.", @@ -145,15 +148,48 @@ class AishellAsrDataModule(DataModule): "collect the batches.", ) - def train_dataloaders(self) -> DataLoader: - logging.info("About to get train cuts") - cuts_train = self.train_cuts() + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + def train_dataloaders(self, cuts_train: CutSet) -> DataLoader: logging.info("About to get Musan cuts") - cuts_musan = load_manifest(self.args.feature_dir / "cuts_musan.json.gz") + cuts_musan = load_manifest( + self.args.manifest_dir / "cuts_musan.json.gz" + ) + + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + transforms.append( + CutMix( + cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True + ) + ) + else: + logging.info("Disable MUSAN") - logging.info("About to create train dataset") - transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))] if self.args.concatenate_cuts: logging.info( f"Using cut concatenation with duration factor " @@ -168,15 +204,25 @@ class AishellAsrDataModule(DataModule): ) ] + transforms - input_transforms = [ - SpecAugment( - num_frame_masks=2, - features_mask_size=27, - num_feature_masks=2, - frames_mask_size=100, + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info( + f"Time warp factor: {self.args.spec_aug_time_warp_factor}" ) - ] + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=2, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + logging.info("About to create train dataset") train = K2SpeechRecognitionDataset( cut_transforms=transforms, input_transforms=input_transforms, @@ -232,10 +278,7 @@ class AishellAsrDataModule(DataModule): return train_dl - def valid_dataloaders(self) -> DataLoader: - logging.info("About to get dev cuts") - cuts_valid = self.valid_cuts() - + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: transforms = [] if self.args.concatenate_cuts: transforms = [ @@ -258,7 +301,7 @@ class AishellAsrDataModule(DataModule): cut_transforms=transforms, return_cuts=self.args.return_cuts, ) - valid_sampler = SingleCutSampler( + valid_sampler = BucketingSampler( cuts_valid, max_duration=self.args.max_duration, shuffle=False, @@ -274,58 +317,39 @@ class AishellAsrDataModule(DataModule): return valid_dl - def test_dataloaders(self) -> Union[DataLoader, List[DataLoader]]: - cuts = self.test_cuts() - is_list = isinstance(cuts, list) - test_loaders = [] - if not is_list: - cuts = [cuts] - - for cuts_test in cuts: - logging.debug("About to create test dataset") - test = K2SpeechRecognitionDataset( - input_strategy=OnTheFlyFeatures( - Fbank(FbankConfig(num_mel_bins=80)) - ) - if self.args.on_the_fly_feats - else PrecomputedFeatures(), - return_cuts=self.args.return_cuts, - ) - sampler = SingleCutSampler( - cuts_test, max_duration=self.args.max_duration - ) - logging.debug("About to create test dataloader") - test_dl = DataLoader( - test, batch_size=None, sampler=sampler, num_workers=1 - ) - test_loaders.append(test_dl) - - if is_list: - return test_loaders - else: - return test_loaders[0] + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + return_cuts=self.args.return_cuts, + ) + sampler = BucketingSampler( + cuts, max_duration=self.args.max_duration, shuffle=False + ) + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl @lru_cache() def train_cuts(self) -> CutSet: logging.info("About to get train cuts") - cuts_train = load_manifest(self.args.feature_dir / "cuts_train.json.gz") + cuts_train = load_manifest( + self.args.manifest_dir / "cuts_train.json.gz" + ) return cuts_train @lru_cache() def valid_cuts(self) -> CutSet: logging.info("About to get dev cuts") - cuts_valid = load_manifest(self.args.feature_dir / "cuts_dev.json.gz") - return cuts_valid + return load_manifest(self.args.manifest_dir / "cuts_dev.json.gz") @lru_cache() def test_cuts(self) -> List[CutSet]: - test_sets = ["test"] - cuts = [] - for test_set in test_sets: - logging.debug("About to get test cuts") - cuts.append( - load_manifest( - self.args.feature_dir / f"cuts_{test_set}.json.gz" - ) - ) - return cuts + logging.info("About to get test cuts") + return load_manifest(self.args.manifest_dir / "cuts_test.json.gz") diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/decode.py b/egs/aishell/ASR/tdnn_lstm_ctc/decode.py index c41d7da17..f3c8e8f44 100755 --- a/egs/aishell/ASR/tdnn_lstm_ctc/decode.py +++ b/egs/aishell/ASR/tdnn_lstm_ctc/decode.py @@ -366,6 +366,8 @@ def main(): model.eval() aishell = AishellAsrDataModule(args) + test_cuts = aishell.test_cuts() + test_dl = aishell.test_dataloaders(test_cuts) # CAUTION: `test_sets` is for displaying only. # If you want to skip test-clean, you have to skip # it inside the for loop. That is, use @@ -373,7 +375,9 @@ def main(): # if test_set == 'test-clean': continue # test_sets = ["test"] - for test_set, test_dl in zip(test_sets, aishell.test_dataloaders()): + test_dls = [test_dl] + + for test_set, test_dl in zip(test_sets, test_dls): results_dict = decode_dataset( dl=test_dl, params=params, diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/train.py b/egs/aishell/ASR/tdnn_lstm_ctc/train.py index 410f07c53..a0045115d 100755 --- a/egs/aishell/ASR/tdnn_lstm_ctc/train.py +++ b/egs/aishell/ASR/tdnn_lstm_ctc/train.py @@ -553,8 +553,8 @@ def run(rank, world_size, args): scheduler.load_state_dict(checkpoints["scheduler"]) aishell = AishellAsrDataModule(args) - train_dl = aishell.train_dataloaders() - valid_dl = aishell.valid_dataloaders() + train_dl = aishell.train_dataloaders(aishell.train_cuts()) + valid_dl = aishell.valid_dataloaders(aishell.valid_cuts()) for epoch in range(params.start_epoch, params.num_epochs): train_dl.sampler.set_epoch(epoch) diff --git a/egs/aishell/ASR/transducer_stateless/README.md b/egs/aishell/ASR/transducer_stateless/README.md new file mode 100644 index 000000000..622cb837c --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/README.md @@ -0,0 +1,21 @@ +## Introduction + +The decoder, i.e., the prediction network, is from +https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 +(Rnn-Transducer with Stateless Prediction Network) + +You can use the following command to start the training: + +```bash +cd egs/aishell/ASR + +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./transducer_stateless/train.py \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 0 \ + --exp-dir transducer_stateless/exp \ + --max-duration 250 \ + --lr-factor 2.5 +``` diff --git a/egs/aishell/ASR/transducer_stateless/__init__.py b/egs/aishell/ASR/transducer_stateless/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/aishell/ASR/transducer_stateless/asr_datamodule.py b/egs/aishell/ASR/transducer_stateless/asr_datamodule.py new file mode 120000 index 000000000..a73848de9 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/asr_datamodule.py @@ -0,0 +1 @@ +../conformer_ctc/asr_datamodule.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless/beam_search.py b/egs/aishell/ASR/transducer_stateless/beam_search.py new file mode 100644 index 000000000..9ed9b2ad1 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/beam_search.py @@ -0,0 +1,339 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import Dict, List, Optional + +import numpy as np +import torch +from model import Transducer + + +def greedy_search( + model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int +) -> List[int]: + """ + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + max_sym_per_frame: + Maximum number of symbols per frame. If it is set to 0, the WER + would be 100%. + Returns: + Return the decoded result. + """ + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + + device = model.device + + decoder_input = torch.tensor( + [blank_id] * context_size, device=device + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + + T = encoder_out.size(1) + t = 0 + hyp = [blank_id] * context_size + + # Maximum symbols per utterance. + max_sym_per_utt = 1000 + + # symbols per frame + sym_per_frame = 0 + + # symbols per utterance decoded so far + sym_per_utt = 0 + + while t < T and sym_per_utt < max_sym_per_utt: + if sym_per_frame >= max_sym_per_frame: + sym_per_frame = 0 + t += 1 + continue + + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :] + # fmt: on + logits = model.joiner(current_encoder_out, decoder_out) + # logits is (1, 1, 1, vocab_size) + + y = logits.argmax().item() + if y != blank_id: + hyp.append(y) + decoder_input = torch.tensor( + [hyp[-context_size:]], device=device + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + + sym_per_utt += 1 + sym_per_frame += 1 + else: + sym_per_frame = 0 + t += 1 + hyp = hyp[context_size:] # remove blanks + + return hyp + + +@dataclass +class Hypothesis: + # The predicted tokens so far. + # Newly predicted tokens are appended to `ys`. + ys: List[int] + + # The log prob of ys + log_prob: float + + @property + def key(self) -> str: + """Return a string representation of self.ys""" + return "_".join(map(str, self.ys)) + + +class HypothesisList(object): + def __init__(self, data: Optional[Dict[str, Hypothesis]] = None): + """ + Args: + data: + A dict of Hypotheses. Its key is its `value.key`. + """ + if data is None: + self._data = {} + else: + self._data = data + + @property + def data(self): + return self._data + + # def add(self, ys: List[int], log_prob: float): + def add(self, hyp: Hypothesis): + """Add a Hypothesis to `self`. + + If `hyp` already exists in `self`, its probability is updated using + `log-sum-exp` with the existed one. + + Args: + hyp: + The hypothesis to be added. + """ + key = hyp.key + if key in self: + old_hyp = self._data[key] + old_hyp.log_prob = np.logaddexp(old_hyp.log_prob, hyp.log_prob) + else: + self._data[key] = hyp + + def get_most_probable(self, length_norm: bool = False) -> Hypothesis: + """Get the most probable hypothesis, i.e., the one with + the largest `log_prob`. + + Args: + length_norm: + If True, the `log_prob` of a hypothesis is normalized by the + number of tokens in it. + + """ + if length_norm: + return max( + self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys) + ) + else: + return max(self._data.values(), key=lambda hyp: hyp.log_prob) + + def remove(self, hyp: Hypothesis) -> None: + """Remove a given hypothesis. + + Args: + hyp: + The hypothesis to be removed from `self`. + Note: It must be contained in `self`. Otherwise, + an exception is raised. + """ + key = hyp.key + assert key in self, f"{key} does not exist" + del self._data[key] + + def filter(self, threshold: float) -> "HypothesisList": + """Remove all Hypotheses whose log_prob is less than threshold. + + Caution: + `self` is not modified. Instead, a new HypothesisList is returned. + + Returns: + Return a new HypothesisList containing all hypotheses from `self` + that have `log_prob` being greater than the given `threshold`. + """ + ans = HypothesisList() + for key, hyp in self._data.items(): + if hyp.log_prob > threshold: + ans.add(hyp) # shallow copy + return ans + + def topk(self, k: int) -> "HypothesisList": + """Return the top-k hypothesis.""" + hyps = list(self._data.items()) + + hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k] + + ans = HypothesisList(dict(hyps)) + return ans + + def __contains__(self, key: str): + return key in self._data + + def __iter__(self): + return iter(self._data.values()) + + def __len__(self) -> int: + return len(self._data) + + def __str__(self) -> str: + s = [] + for key in self: + s.append(key) + return ", ".join(s) + + +def beam_search( + model: Transducer, + encoder_out: torch.Tensor, + beam: int = 4, +) -> List[int]: + """ + It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf + + espnet/nets/beam_search_transducer.py#L247 is used as a reference. + + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + beam: + Beam size. + Returns: + Return the decoded result. + """ + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + + device = model.device + + decoder_input = torch.tensor( + [blank_id] * context_size, device=device + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + + T = encoder_out.size(1) + t = 0 + + B = HypothesisList() + B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0)) + + max_sym_per_utt = 20000 + + sym_per_utt = 0 + + decoder_cache: Dict[str, torch.Tensor] = {} + + while t < T and sym_per_utt < max_sym_per_utt: + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :] + # fmt: on + A = B + B = HypothesisList() + + joint_cache: Dict[str, torch.Tensor] = {} + + # TODO(fangjun): Implement prefix search to update the `log_prob` + # of hypotheses in A + + while True: + y_star = A.get_most_probable() + A.remove(y_star) + + cached_key = y_star.key + + if cached_key not in decoder_cache: + decoder_input = torch.tensor( + [y_star.ys[-context_size:]], device=device + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_cache[cached_key] = decoder_out + else: + decoder_out = decoder_cache[cached_key] + + cached_key += f"-t-{t}" + if cached_key not in joint_cache: + logits = model.joiner(current_encoder_out, decoder_out) + + # TODO(fangjun): Ccale the blank posterior + + log_prob = logits.log_softmax(dim=-1) + # log_prob is (1, 1, 1, vocab_size) + log_prob = log_prob.squeeze() + # Now log_prob is (vocab_size,) + joint_cache[cached_key] = log_prob + else: + log_prob = joint_cache[cached_key] + + # First, process the blank symbol + skip_log_prob = log_prob[blank_id] + new_y_star_log_prob = y_star.log_prob + skip_log_prob.item() + + # ys[:] returns a copy of ys + B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob)) + + # Second, process other non-blank labels + values, indices = log_prob.topk(beam + 1) + for i, v in zip(indices.tolist(), values.tolist()): + if i == blank_id: + continue + new_ys = y_star.ys + [i] + new_log_prob = y_star.log_prob + v + A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob)) + + # Check whether B contains more than "beam" elements more probable + # than the most probable in A + A_most_probable = A.get_most_probable() + + kept_B = B.filter(A_most_probable.log_prob) + + if len(kept_B) >= beam: + B = kept_B.topk(beam) + break + + t += 1 + + best_hyp = B.get_most_probable(length_norm=True) + ys = best_hyp.ys[context_size:] # [context_size:] to remove blanks + return ys diff --git a/egs/aishell/ASR/transducer_stateless/conformer.py b/egs/aishell/ASR/transducer_stateless/conformer.py new file mode 100644 index 000000000..81d7708f9 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/conformer.py @@ -0,0 +1,920 @@ +#!/usr/bin/env python3 +# Copyright (c) 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import math +import warnings +from typing import Optional, Tuple + +import torch +from torch import Tensor, nn +from transformer import Transformer + +from icefall.utils import make_pad_mask + + +class Conformer(Transformer): + """ + Args: + num_features (int): Number of input features + output_dim (int): Number of output dimension + subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers) + d_model (int): attention dimension + nhead (int): number of head + dim_feedforward (int): feedforward dimention + num_encoder_layers (int): number of encoder layers + dropout (float): dropout rate + cnn_module_kernel (int): Kernel size of convolution module + normalize_before (bool): whether to use layer_norm before the first block. + vgg_frontend (bool): whether to use vgg frontend. + """ + + def __init__( + self, + num_features: int, + output_dim: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + cnn_module_kernel: int = 31, + normalize_before: bool = True, + vgg_frontend: bool = False, + ) -> None: + super(Conformer, self).__init__( + num_features=num_features, + output_dim=output_dim, + subsampling_factor=subsampling_factor, + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + num_encoder_layers=num_encoder_layers, + dropout=dropout, + normalize_before=normalize_before, + vgg_frontend=vgg_frontend, + ) + + self.encoder_pos = RelPositionalEncoding(d_model, dropout) + + encoder_layer = ConformerEncoderLayer( + d_model, + nhead, + dim_feedforward, + dropout, + cnn_module_kernel, + normalize_before, + ) + self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers) + self.normalize_before = normalize_before + if self.normalize_before: + self.after_norm = nn.LayerNorm(d_model) + else: + # Note: TorchScript detects that self.after_norm could be used inside forward() + # and throws an error without this change. + self.after_norm = identity + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + Returns: + Return a tuple containing 2 tensors: + - logits, its shape is (batch_size, output_seq_len, output_dim) + - logit_lens, a tensor of shape (batch_size,) containing the number + of frames in `logits` before padding. + """ + x = self.encoder_embed(x) + x, pos_emb = self.encoder_pos(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + lengths = ((x_lens - 1) // 2 - 1) // 2 + assert x.size(0) == lengths.max().item() + mask = make_pad_mask(lengths) + + x = self.encoder(x, pos_emb, src_key_padding_mask=mask) # (T, N, C) + + if self.normalize_before: + x = self.after_norm(x) + + logits = self.encoder_output_layer(x) + logits = logits.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return logits, lengths + + +class ConformerEncoderLayer(nn.Module): + """ + ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks. + See: "Conformer: Convolution-augmented Transformer for Speech Recognition" + + Args: + d_model: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + dim_feedforward: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + normalize_before: whether to use layer_norm before the first block. + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int = 2048, + dropout: float = 0.1, + cnn_module_kernel: int = 31, + normalize_before: bool = True, + ) -> None: + super(ConformerEncoderLayer, self).__init__() + self.self_attn = RelPositionMultiheadAttention( + d_model, nhead, dropout=0.0 + ) + + self.feed_forward = nn.Sequential( + nn.Linear(d_model, dim_feedforward), + Swish(), + nn.Dropout(dropout), + nn.Linear(dim_feedforward, d_model), + ) + + self.feed_forward_macaron = nn.Sequential( + nn.Linear(d_model, dim_feedforward), + Swish(), + nn.Dropout(dropout), + nn.Linear(dim_feedforward, d_model), + ) + + self.conv_module = ConvolutionModule(d_model, cnn_module_kernel) + + self.norm_ff_macaron = nn.LayerNorm( + d_model + ) # for the macaron style FNN module + self.norm_ff = nn.LayerNorm(d_model) # for the FNN module + self.norm_mha = nn.LayerNorm(d_model) # for the MHA module + + self.ff_scale = 0.5 + + self.norm_conv = nn.LayerNorm(d_model) # for the CNN module + self.norm_final = nn.LayerNorm( + d_model + ) # for the final output of the block + + self.dropout = nn.Dropout(dropout) + + self.normalize_before = normalize_before + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + src_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + pos_emb: Positional embedding tensor (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, N is the batch size, E is the feature number + """ + + # macaron style feed forward module + residual = src + if self.normalize_before: + src = self.norm_ff_macaron(src) + src = residual + self.ff_scale * self.dropout( + self.feed_forward_macaron(src) + ) + if not self.normalize_before: + src = self.norm_ff_macaron(src) + + # multi-headed self-attention module + residual = src + if self.normalize_before: + src = self.norm_mha(src) + src_att = self.self_attn( + src, + src, + src, + pos_emb=pos_emb, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + )[0] + src = residual + self.dropout(src_att) + if not self.normalize_before: + src = self.norm_mha(src) + + # convolution module + residual = src + if self.normalize_before: + src = self.norm_conv(src) + src = residual + self.dropout(self.conv_module(src)) + if not self.normalize_before: + src = self.norm_conv(src) + + # feed forward module + residual = src + if self.normalize_before: + src = self.norm_ff(src) + src = residual + self.ff_scale * self.dropout(self.feed_forward(src)) + if not self.normalize_before: + src = self.norm_ff(src) + + if self.normalize_before: + src = self.norm_final(src) + + return src + + +class ConformerEncoder(nn.TransformerEncoder): + r"""ConformerEncoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the ConformerEncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + norm: the layer normalization component (optional). + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = conformer_encoder(src, pos_emb) + """ + + def __init__( + self, encoder_layer: nn.Module, num_layers: int, norm: nn.Module = None + ) -> None: + super(ConformerEncoder, self).__init__( + encoder_layer=encoder_layer, num_layers=num_layers, norm=norm + ) + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required). + pos_emb: Positional embedding tensor (required). + mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + """ + output = src + + for mod in self.layers: + output = mod( + output, + pos_emb, + src_mask=mask, + src_key_padding_mask=src_key_padding_mask, + ) + + if self.norm is not None: + output = self.norm(output) + + return output + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + + """ + + def __init__( + self, d_model: int, dropout_rate: float, max_len: int = 5000 + ) -> None: + """Construct an PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.xscale = math.sqrt(self.d_model) + self.dropout = torch.nn.Dropout(p=dropout_rate) + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x: Tensor) -> None: + """Reset the positional encodings.""" + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(1) >= x.size(1) * 2 - 1: + # Note: TorchScript doesn't implement operator== for torch.Device + if self.pe.dtype != x.dtype or str(self.pe.device) != str( + x.device + ): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i Tuple[Tensor, Tensor]: + """Add positional encoding. + + Args: + x (torch.Tensor): Input tensor (batch, time, `*`). + + Returns: + torch.Tensor: Encoded tensor (batch, time, `*`). + torch.Tensor: Encoded tensor (batch, 2*time-1, `*`). + + """ + self.extend_pe(x) + x = x * self.xscale + pos_emb = self.pe[ + :, + self.pe.size(1) // 2 + - x.size(1) + + 1 : self.pe.size(1) // 2 # noqa E203 + + x.size(1), + ] + return self.dropout(x), self.dropout(pos_emb) + + +class RelPositionMultiheadAttention(nn.Module): + r"""Multi-Head Attention layer with relative position encoding + + See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + + Args: + embed_dim: total dimension of the model. + num_heads: parallel attention heads. + dropout: a Dropout layer on attn_output_weights. Default: 0.0. + + Examples:: + + >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads) + >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + num_heads: int, + dropout: float = 0.0, + ) -> None: + super(RelPositionMultiheadAttention, self).__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + + self.in_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=True) + self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True) + + # linear transformation for positional encoding. + self.linear_pos = nn.Linear(embed_dim, embed_dim, bias=False) + # these two learnable bias are used in matrix c and matrix d + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + + self._reset_parameters() + + def _reset_parameters(self) -> None: + nn.init.xavier_uniform_(self.in_proj.weight) + nn.init.constant_(self.in_proj.bias, 0.0) + nn.init.constant_(self.out_proj.bias, 0.0) + + nn.init.xavier_uniform_(self.pos_bias_u) + nn.init.xavier_uniform_(self.pos_bias_v) + + def forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. When given a binary mask and a value is True, + the corresponding value on the attention layer will be ignored. When given + a byte mask and a value is non-zero, the corresponding value on the attention + layer will be ignored + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + - Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the position + with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + - Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + return self.multi_head_attention_forward( + query, + key, + value, + pos_emb, + self.embed_dim, + self.num_heads, + self.in_proj.weight, + self.in_proj.bias, + self.dropout, + self.out_proj.weight, + self.out_proj.bias, + training=self.training, + key_padding_mask=key_padding_mask, + need_weights=need_weights, + attn_mask=attn_mask, + ) + + def rel_shift(self, x: Tensor) -> Tensor: + """Compute relative positional encoding. + + Args: + x: Input tensor (batch, head, time1, 2*time1-1). + time1 means the length of query vector. + + Returns: + Tensor: tensor of shape (batch, head, time1, time2) + (note: time2 has the same value as time1, but it is for + the key, while time1 is for the query). + """ + (batch_size, num_heads, time1, n) = x.shape + assert n == 2 * time1 - 1 + # Note: TorchScript requires explicit arg for stride() + batch_stride = x.stride(0) + head_stride = x.stride(1) + time1_stride = x.stride(2) + n_stride = x.stride(3) + return x.as_strided( + (batch_size, num_heads, time1, time1), + (batch_stride, head_stride, time1_stride - n_stride, n_stride), + storage_offset=n_stride * (time1 - 1), + ) + + def multi_head_attention_forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + embed_dim_to_check: int, + num_heads: int, + in_proj_weight: Tensor, + in_proj_bias: Tensor, + dropout_p: float, + out_proj_weight: Tensor, + out_proj_bias: Tensor, + training: bool = True, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + embed_dim_to_check: total dimension of the model. + num_heads: parallel attention heads. + in_proj_weight, in_proj_bias: input projection weight and bias. + dropout_p: probability of an element to be zeroed. + out_proj_weight, out_proj_bias: the output projection weight and bias. + training: apply dropout if is ``True``. + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. This is an binary mask. When the value is True, + the corresponding value on the attention layer will be filled with -inf. + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence + length, N is the batch size, E is the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions + will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == embed_dim_to_check + assert key.size(0) == value.size(0) and key.size(1) == value.size(1) + + head_dim = embed_dim // num_heads + assert ( + head_dim * num_heads == embed_dim + ), "embed_dim must be divisible by num_heads" + scaling = float(head_dim) ** -0.5 + + if torch.equal(query, key) and torch.equal(key, value): + # self-attention + q, k, v = nn.functional.linear( + query, in_proj_weight, in_proj_bias + ).chunk(3, dim=-1) + + elif torch.equal(key, value): + # encoder-decoder attention + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1) + + else: + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = embed_dim * 2 + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + k = nn.functional.linear(key, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim * 2 + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + v = nn.functional.linear(value, _w, _b) + + if attn_mask is not None: + assert ( + attn_mask.dtype == torch.float32 + or attn_mask.dtype == torch.float64 + or attn_mask.dtype == torch.float16 + or attn_mask.dtype == torch.uint8 + or attn_mask.dtype == torch.bool + ), "Only float, byte, and bool types are supported for attn_mask, not {}".format( + attn_mask.dtype + ) + if attn_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for attn_mask is deprecated. Use bool tensor instead." + ) + attn_mask = attn_mask.to(torch.bool) + + if attn_mask.dim() == 2: + attn_mask = attn_mask.unsqueeze(0) + if list(attn_mask.size()) != [1, query.size(0), key.size(0)]: + raise RuntimeError( + "The size of the 2D attn_mask is not correct." + ) + elif attn_mask.dim() == 3: + if list(attn_mask.size()) != [ + bsz * num_heads, + query.size(0), + key.size(0), + ]: + raise RuntimeError( + "The size of the 3D attn_mask is not correct." + ) + else: + raise RuntimeError( + "attn_mask's dimension {} is not supported".format( + attn_mask.dim() + ) + ) + # attn_mask's dim is 3 now. + + # convert ByteTensor key_padding_mask to bool + if ( + key_padding_mask is not None + and key_padding_mask.dtype == torch.uint8 + ): + warnings.warn( + "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead." + ) + key_padding_mask = key_padding_mask.to(torch.bool) + + q = q.contiguous().view(tgt_len, bsz, num_heads, head_dim) + k = k.contiguous().view(-1, bsz, num_heads, head_dim) + v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1) + + src_len = k.size(0) + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz, "{} == {}".format( + key_padding_mask.size(0), bsz + ) + assert key_padding_mask.size(1) == src_len, "{} == {}".format( + key_padding_mask.size(1), src_len + ) + + q = q.transpose(0, 1) # (batch, time1, head, d_k) + + pos_emb_bsz = pos_emb.size(0) + assert pos_emb_bsz in (1, bsz) # actually it is 1 + p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim) + p = p.transpose(1, 2) # (batch, head, 2*time1-1, d_k) + + q_with_bias_u = (q + self.pos_bias_u).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + q_with_bias_v = (q + self.pos_bias_v).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + # compute attention score + # first compute matrix a and matrix c + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + k = k.permute(1, 2, 3, 0) # (batch, head, d_k, time2) + matrix_ac = torch.matmul( + q_with_bias_u, k + ) # (batch, head, time1, time2) + + # compute matrix b and matrix d + matrix_bd = torch.matmul( + q_with_bias_v, p.transpose(-2, -1) + ) # (batch, head, time1, 2*time1-1) + matrix_bd = self.rel_shift(matrix_bd) + + attn_output_weights = ( + matrix_ac + matrix_bd + ) * scaling # (batch, head, time1, time2) + + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, -1 + ) + + assert list(attn_output_weights.size()) == [ + bsz * num_heads, + tgt_len, + src_len, + ] + + if attn_mask is not None: + if attn_mask.dtype == torch.bool: + attn_output_weights.masked_fill_(attn_mask, float("-inf")) + else: + attn_output_weights += attn_mask + + if key_padding_mask is not None: + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + attn_output_weights = attn_output_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + float("-inf"), + ) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, src_len + ) + + attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1) + attn_output_weights = nn.functional.dropout( + attn_output_weights, p=dropout_p, training=training + ) + + attn_output = torch.bmm(attn_output_weights, v) + assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim] + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(tgt_len, bsz, embed_dim) + ) + attn_output = nn.functional.linear( + attn_output, out_proj_weight, out_proj_bias + ) + + if need_weights: + # average attention weights over heads + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + return attn_output, attn_output_weights.sum(dim=1) / num_heads + else: + return attn_output, None + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Conformer model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__( + self, channels: int, kernel_size: int, bias: bool = True + ) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + self.pointwise_conv1 = nn.Conv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.depthwise_conv = nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + padding=(kernel_size - 1) // 2, + groups=channels, + bias=bias, + ) + self.norm = nn.LayerNorm(channels) + self.pointwise_conv2 = nn.Conv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.activation = Swish() + + def forward(self, x: Tensor) -> Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + # x is (batch, channels, time) + x = x.permute(0, 2, 1) + x = self.norm(x) + x = x.permute(0, 2, 1) + + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + return x.permute(2, 0, 1) + + +class Swish(torch.nn.Module): + """Construct an Swish object.""" + + def forward(self, x: Tensor) -> Tensor: + """Return Swich activation function.""" + return x * torch.sigmoid(x) + + +def identity(x): + return x diff --git a/egs/aishell/ASR/transducer_stateless/decode.py b/egs/aishell/ASR/transducer_stateless/decode.py new file mode 100755 index 000000000..f27e4cdcf --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/decode.py @@ -0,0 +1,473 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Tuple + +import torch +import torch.nn as nn +from asr_datamodule import AishellAsrDataModule +from beam_search import beam_search, greedy_search +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer + +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import average_checkpoints, load_checkpoint +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + write_error_stats, + str2bool, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=10, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_stateless/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="The lang dir", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="Used only when --decoding-method is beam_search", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=3, + help="Maximum number of symbols per frame", + ) + parser.add_argument( + "--export", + type=str2bool, + default=False, + help="""When enabled, the averaged model is saved to + transducer_stateless/exp/pretrained.pt. Note: only model.state_dict() + is saved. pretrained.pt contains a dict {"model": model.state_dict()}, + which can be loaded by `icefall.checkpoint.load_checkpoint()`. + """, + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + } + ) + return params + + +def get_encoder_model(params: AttributeDict): + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict): + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict): + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict): + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + batch: dict, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + lexicon: + It contains the token symbol table and the word symbol table. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = model.device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, encoder_out=encoder_out_i, beam=params.beam_size + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append([lexicon.token_table[i] for i in hyp]) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + else: + return {f"beam_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + lexicon=lexicon, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((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[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + + # 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" + ) + # we compute CER for aishell dataset. + results_char = [] + for res in results: + results_char.append((list("".join(res[0])), list("".join(res[1])))) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results_char, 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\tCER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER 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() + AishellAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ("greedy_search", "beam_search") + params.res_dir = params.exp_dir / params.decoding_method + + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + if params.decoding_method == "beam_search": + params.suffix += f"-beam-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + lexicon = Lexicon(params.lang_dir) + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + params.blank_id = graph_compiler.texts_to_ids("")[0][0] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + if params.export: + logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt") + torch.save( + {"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt" + ) + return + + model.to(device) + model.eval() + model.device = device + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + aishell = AishellAsrDataModule(args) + test_cuts = aishell.test_cuts() + test_dl = aishell.test_dataloaders(test_cuts) + + test_sets = ["test"] + test_dls = [test_dl] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + lexicon=lexicon, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/aishell/ASR/transducer_stateless/decoder.py b/egs/aishell/ASR/transducer_stateless/decoder.py new file mode 100644 index 000000000..dca084477 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/decoder.py @@ -0,0 +1,98 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class Decoder(nn.Module): + """This class modifies the stateless decoder from the following paper: + + RNN-transducer with stateless prediction network + https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 + + It removes the recurrent connection from the decoder, i.e., the prediction + network. Different from the above paper, it adds an extra Conv1d + right after the embedding layer. + + TODO: Implement https://arxiv.org/pdf/2109.07513.pdf + """ + + def __init__( + self, + vocab_size: int, + embedding_dim: int, + blank_id: int, + context_size: int, + ): + """ + Args: + vocab_size: + Number of tokens of the modeling unit including blank. + embedding_dim: + Dimension of the input embedding. + blank_id: + The ID of the blank symbol. + context_size: + Number of previous words to use to predict the next word. + 1 means bigram; 2 means trigram. n means (n+1)-gram. + """ + super().__init__() + self.embedding = nn.Embedding( + num_embeddings=vocab_size, + embedding_dim=embedding_dim, + padding_idx=blank_id, + ) + self.blank_id = blank_id + + assert context_size >= 1, context_size + self.context_size = context_size + if context_size > 1: + self.conv = nn.Conv1d( + in_channels=embedding_dim, + out_channels=embedding_dim, + kernel_size=context_size, + padding=0, + groups=embedding_dim, + bias=False, + ) + + def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, U) with blank prepended. + need_pad: + True to left pad the input. Should be True during training. + False to not pad the input. Should be False during inference. + Returns: + Return a tensor of shape (N, U, embedding_dim). + """ + embeding_out = self.embedding(y) + if self.context_size > 1: + embeding_out = embeding_out.permute(0, 2, 1) + if need_pad is True: + embeding_out = F.pad( + embeding_out, pad=(self.context_size - 1, 0) + ) + else: + # During inference time, there is no need to do extra padding + # as we only need one output + assert embeding_out.size(-1) == self.context_size + embeding_out = self.conv(embeding_out) + embeding_out = embeding_out.permute(0, 2, 1) + return embeding_out diff --git a/egs/aishell/ASR/transducer_stateless/encoder_interface.py b/egs/aishell/ASR/transducer_stateless/encoder_interface.py new file mode 100644 index 000000000..257facce4 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/encoder_interface.py @@ -0,0 +1,43 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Tuple + +import torch +import torch.nn as nn + + +class EncoderInterface(nn.Module): + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + A tensor of shape (batch_size, input_seq_len, num_features) + containing the input features. + x_lens: + A tensor of shape (batch_size,) containing the number of frames + in `x` before padding. + Returns: + Return a tuple containing two tensors: + - encoder_out, a tensor of (batch_size, out_seq_len, output_dim) + containing unnormalized probabilities, i.e., the output of a + linear layer. + - encoder_out_lens, a tensor of shape (batch_size,) containing + the number of frames in `encoder_out` before padding. + """ + raise NotImplementedError("Please implement it in a subclass") diff --git a/egs/aishell/ASR/transducer_stateless/export.py b/egs/aishell/ASR/transducer_stateless/export.py new file mode 100755 index 000000000..641555bdb --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/export.py @@ -0,0 +1,248 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" +Usage: +./transducer_stateless/export.py \ + --exp-dir ./transducer_stateless/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `transducer_stateless/decode.py`, you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./transducer_stateless/decode.py \ + --exp-dir ./transducer_stateless/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 1 \ + --bpe-model data/lang_bpe_500/bpe.model +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer + +from icefall.checkpoint import average_checkpoints, load_checkpoint +from icefall.env import get_env_info +from icefall.utils import AttributeDict, str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=20, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--avg", + type=int, + default=10, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_stateless/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + } + ) + return params + + +def get_encoder_model(params: AttributeDict): + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict): + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict): + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict): + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + assert args.jit is False, "Support torchscript will be added later" + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + model.to(device) + + if params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.eval() + + model.to("cpu") + model.eval() + + if params.jit: + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torch.jit.script") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/aishell/ASR/transducer_stateless/joiner.py b/egs/aishell/ASR/transducer_stateless/joiner.py new file mode 100644 index 000000000..2ef3f1de6 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/joiner.py @@ -0,0 +1,54 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn + + +class Joiner(nn.Module): + def __init__(self, input_dim: int, output_dim: int): + super().__init__() + + self.output_linear = nn.Linear(input_dim, output_dim) + + def forward( + self, encoder_out: torch.Tensor, decoder_out: torch.Tensor + ) -> torch.Tensor: + """ + Args: + encoder_out: + Output from the encoder. Its shape is (N, T, C). + decoder_out: + Output from the decoder. Its shape is (N, U, C). + Returns: + Return a tensor of shape (N, T, U, C). + """ + assert encoder_out.ndim == decoder_out.ndim == 3 + assert encoder_out.size(0) == decoder_out.size(0) + assert encoder_out.size(2) == decoder_out.size(2) + + encoder_out = encoder_out.unsqueeze(2) + # Now encoder_out is (N, T, 1, C) + + decoder_out = decoder_out.unsqueeze(1) + # Now decoder_out is (N, 1, U, C) + + logit = encoder_out + decoder_out + logit = torch.tanh(logit) + + output = self.output_linear(logit) + + return output diff --git a/egs/aishell/ASR/transducer_stateless/model.py b/egs/aishell/ASR/transducer_stateless/model.py new file mode 100644 index 000000000..2f0f9a183 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/model.py @@ -0,0 +1,125 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Note we use `rnnt_loss` from torchaudio, which exists only in +torchaudio >= v0.10.0. It also means you have to use torch >= v1.10.0 +""" +import k2 +import torch +import torch.nn as nn +import torchaudio +import torchaudio.functional +from encoder_interface import EncoderInterface + +from icefall.utils import add_sos + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + ): + """ + Args: + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, C) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, C) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, C). It should contain + one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, C) and (N, U, C). Its + output shape is (N, T, U, C). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + ) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + Returns: + Return the transducer loss. + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + encoder_out, x_lens = self.encoder(x, x_lens) + assert torch.all(x_lens > 0) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + decoder_out = self.decoder(sos_y_padded) + + logits = self.joiner(encoder_out, decoder_out) + + # rnnt_loss requires 0 padded targets + # Note: y does not start with SOS + y_padded = y.pad(mode="constant", padding_value=0) + + assert hasattr(torchaudio.functional, "rnnt_loss"), ( + f"Current torchaudio version: {torchaudio.__version__}\n" + "Please install a version >= 0.10.0" + ) + + loss = torchaudio.functional.rnnt_loss( + logits=logits, + targets=y_padded, + logit_lengths=x_lens, + target_lengths=y_lens, + blank=blank_id, + reduction="sum", + ) + + return loss diff --git a/egs/aishell/ASR/transducer_stateless/pretrained.py b/egs/aishell/ASR/transducer_stateless/pretrained.py new file mode 100755 index 000000000..65ac5f3ff --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/pretrained.py @@ -0,0 +1,326 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +(1) greedy search +./transducer_stateless/pretrained.py \ + --checkpoint ./transducer_stateless/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav \ + +(1) beam search +./transducer_stateless/pretrained.py \ + --checkpoint ./transducer_stateless/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav \ + +You can also use `./transducer_stateless/exp/epoch-xx.pt`. + +Note: ./transducer_stateless/exp/pretrained.pt is generated by +./transducer_stateless/export.py +""" + + +import argparse +import logging +import math +from typing import List +from pathlib import Path + +import kaldifeat +import torch +import torchaudio +from beam_search import beam_search, greedy_search +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer +from torch.nn.utils.rnn import pad_sequence + +from icefall.env import get_env_info +from icefall.utils import AttributeDict +from icefall.lexicon import Lexicon +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--lang-dir", + type=str, + help="""Path to lang. + Used only when method is ctc-decoding. + """, + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="Used only when --method is beam_search", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=3, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + "sample_rate": 16000, + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + } + ) + return params + + +def get_encoder_model(params: AttributeDict): + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict): + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict): + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict): + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +def main(): + parser = get_parser() + args = parser.parse_args() + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + lexicon = Lexicon(params.lang_dir) + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + params.blank_id = graph_compiler.texts_to_ids("")[0][0] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info("Creating model") + model = get_transducer_model(params) + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + with torch.no_grad(): + encoder_out, encoder_out_lens = model.encoder( + x=features, x_lens=feature_lengths + ) + + num_waves = encoder_out.size(0) + hyps = [] + msg = f"Using {params.method}" + if params.method == "beam_search": + msg += f" with beam size {params.beam_size}" + logging.info(msg) + for i in range(num_waves): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, encoder_out=encoder_out_i, beam=params.beam_size + ) + else: + raise ValueError(f"Unsupported method: {params.method}") + + hyps.append([lexicon.token_table[i] for i in hyp]) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/aishell/ASR/transducer_stateless/subsampling.py b/egs/aishell/ASR/transducer_stateless/subsampling.py new file mode 120000 index 000000000..6fee09e58 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/subsampling.py @@ -0,0 +1 @@ +../conformer_ctc/subsampling.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless/test_decoder.py b/egs/aishell/ASR/transducer_stateless/test_decoder.py new file mode 100755 index 000000000..fe0bdee70 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/test_decoder.py @@ -0,0 +1,58 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +To run this file, do: + + cd icefall/egs/aishell/ASR + python ./transducer_stateless/test_decoder.py +""" + +import torch +from decoder import Decoder + + +def test_decoder(): + vocab_size = 3 + blank_id = 0 + embedding_dim = 128 + context_size = 4 + + decoder = Decoder( + vocab_size=vocab_size, + embedding_dim=embedding_dim, + blank_id=blank_id, + context_size=context_size, + ) + N = 100 + U = 20 + x = torch.randint(low=0, high=vocab_size, size=(N, U)) + y = decoder(x) + assert y.shape == (N, U, embedding_dim) + + # for inference + x = torch.randint(low=0, high=vocab_size, size=(N, context_size)) + y = decoder(x, need_pad=False) + assert y.shape == (N, 1, embedding_dim) + + +def main(): + test_decoder() + + +if __name__ == "__main__": + main() diff --git a/egs/aishell/ASR/transducer_stateless/train.py b/egs/aishell/ASR/transducer_stateless/train.py new file mode 100755 index 000000000..7da8e28a1 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/train.py @@ -0,0 +1,672 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang +# Mingshuang Luo) +# Copyright 2021 (Pingfeng Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import logging +from pathlib import Path +from shutil import copyfile +from typing import Optional, Tuple + +import k2 +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AishellAsrDataModule +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.utils import fix_random_seed +from model import Transducer +from torch import Tensor +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.nn.utils import clip_grad_norm_ +from torch.utils.tensorboard import SummaryWriter +from transformer import Noam + +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=0, + help="""Resume training from from this epoch. + If it is positive, it will load checkpoint from + transducer_stateless/exp/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_stateless/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--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( + "--lr-factor", + type=float, + default=5.0, + help="The lr_factor for Noam optimizer", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - attention_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + # parameters for Noam + "warm_step": 80000, # For the 100h subset, use 8k + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict): + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict): + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict): + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict): + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None, +) -> None: + """Load checkpoint from file. + + If params.start_epoch is positive, it will load the checkpoint from + `params.start_epoch - 1`. Otherwise, this function does nothing. + + Apart from loading state dict for `model`, `optimizer` and `scheduler`, + it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + optimizer: + The optimizer that we are using. + scheduler: + The learning rate scheduler we are using. + Returns: + Return None. + """ + if params.start_epoch <= 0: + return + + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + saved_params = load_checkpoint( + filename, + model=model, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: nn.Module, + graph_compiler: CharCtcTrainingGraphCompiler, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = graph_compiler.texts_to_ids(texts) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + loss = model(x=feature, x_lens=feature_lens, y=y) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: nn.Module, + 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: nn.Module, + optimizer: torch.optim.Optimizer, + graph_compiler: CharCtcTrainingGraphCompiler, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + """ + model.train() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + loss, loss_info = compute_loss( + params=params, + model=model, + 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. + + optimizer.zero_grad() + loss.backward() + clip_grad_norm_(model.parameters(), 5.0, 2.0) + optimizer.step() + + if batch_idx % params.log_interval == 0: + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}" + ) + + if batch_idx % params.log_interval == 0: + + if tb_writer is not None: + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + + if batch_idx > 0 and batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(42) + 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, + oov="", + ) + + params.blank_id = graph_compiler.texts_to_ids("")[0][0] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoints = load_checkpoint_if_available(params=params, model=model) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank]) + model.device = device + + optimizer = Noam( + model.parameters(), + model_size=params.attention_dim, + factor=params.lr_factor, + warm_step=params.warm_step, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + aishell = AishellAsrDataModule(args) + train_cuts = aishell.train_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + return 1.0 <= c.duration <= 20.0 + + num_in_total = len(train_cuts) + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + num_left = len(train_cuts) + num_removed = num_in_total - num_left + removed_percent = num_removed / num_in_total * 100 + + logging.info(f"Before removing short and long utterances: {num_in_total}") + logging.info(f"After removing short and long utterances: {num_left}") + logging.info(f"Removed {num_removed} utterances ({removed_percent:.5f}%)") + + train_dl = aishell.train_dataloaders(train_cuts) + valid_dl = aishell.valid_dataloaders(aishell.valid_cuts()) + + for epoch in range(params.start_epoch, params.num_epochs): + train_dl.sampler.set_epoch(epoch) + + cur_lr = optimizer._rate + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + if rank == 0: + logging.info("epoch {}, learning rate {}".format(epoch, cur_lr)) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + optimizer=optimizer, + graph_compiler=graph_compiler, + train_dl=train_dl, + valid_dl=valid_dl, + tb_writer=tb_writer, + world_size=world_size, + ) + + save_checkpoint( + params=params, + model=model, + optimizer=optimizer, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def main(): + parser = get_parser() + AishellAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_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/aishell/ASR/transducer_stateless/transformer.py b/egs/aishell/ASR/transducer_stateless/transformer.py new file mode 100644 index 000000000..e851dcc32 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless/transformer.py @@ -0,0 +1,418 @@ +# Copyright 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import math +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from subsampling import Conv2dSubsampling, VggSubsampling + +from icefall.utils import make_pad_mask + + +class Transformer(EncoderInterface): + def __init__( + self, + num_features: int, + output_dim: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + normalize_before: bool = True, + vgg_frontend: bool = False, + ) -> None: + """ + Args: + num_features: + The input dimension of the model. + output_dim: + The output dimension of the model. + subsampling_factor: + Number of output frames is num_in_frames // subsampling_factor. + Currently, subsampling_factor MUST be 4. + d_model: + Attention dimension. + nhead: + Number of heads in multi-head attention. + Must satisfy d_model // nhead == 0. + dim_feedforward: + The output dimension of the feedforward layers in encoder. + num_encoder_layers: + Number of encoder layers. + dropout: + Dropout in encoder. + normalize_before: + If True, use pre-layer norm; False to use post-layer norm. + vgg_frontend: + True to use vgg style frontend for subsampling. + """ + super().__init__() + + self.num_features = num_features + self.output_dim = output_dim + self.subsampling_factor = subsampling_factor + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + if vgg_frontend: + self.encoder_embed = VggSubsampling(num_features, d_model) + else: + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder_pos = PositionalEncoding(d_model, dropout) + + encoder_layer = TransformerEncoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + dropout=dropout, + normalize_before=normalize_before, + ) + + if normalize_before: + encoder_norm = nn.LayerNorm(d_model) + else: + encoder_norm = None + + self.encoder = nn.TransformerEncoder( + encoder_layer=encoder_layer, + num_layers=num_encoder_layers, + norm=encoder_norm, + ) + + # TODO(fangjun): remove dropout + self.encoder_output_layer = nn.Sequential( + nn.Dropout(p=dropout), nn.Linear(d_model, output_dim) + ) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + Returns: + Return a tuple containing 2 tensors: + - logits, its shape is (batch_size, output_seq_len, output_dim) + - logit_lens, a tensor of shape (batch_size,) containing the number + of frames in `logits` before padding. + """ + x = self.encoder_embed(x) + x = self.encoder_pos(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + lengths = ((x_lens - 1) // 2 - 1) // 2 + assert x.size(0) == lengths.max().item() + + mask = make_pad_mask(lengths) + x = self.encoder(x, src_key_padding_mask=mask) # (T, N, C) + + logits = self.encoder_output_layer(x) + logits = logits.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return logits, lengths + + +class TransformerEncoderLayer(nn.Module): + """ + Modified from torch.nn.TransformerEncoderLayer. + Add support of normalize_before, + i.e., use layer_norm before the first block. + + Args: + d_model: + the number of expected features in the input (required). + nhead: + the number of heads in the multiheadattention models (required). + dim_feedforward: + the dimension of the feedforward network model (default=2048). + dropout: + the dropout value (default=0.1). + activation: + the activation function of intermediate layer, relu or + gelu (default=relu). + normalize_before: + whether to use layer_norm before the first block. + + Examples:: + >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> out = encoder_layer(src) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int = 2048, + dropout: float = 0.1, + activation: str = "relu", + normalize_before: bool = True, + ) -> None: + super(TransformerEncoderLayer, self).__init__() + self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0) + # Implementation of Feedforward model + self.linear1 = nn.Linear(d_model, dim_feedforward) + self.dropout = nn.Dropout(dropout) + self.linear2 = nn.Linear(dim_feedforward, d_model) + + self.norm1 = nn.LayerNorm(d_model) + self.norm2 = nn.LayerNorm(d_model) + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + + self.normalize_before = normalize_before + + def __setstate__(self, state): + if "activation" not in state: + state["activation"] = nn.functional.relu + super(TransformerEncoderLayer, self).__setstate__(state) + + def forward( + self, + src: torch.Tensor, + src_mask: Optional[torch.Tensor] = None, + src_key_padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional) + + Shape: + src: (S, N, E). + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, + N is the batch size, E is the feature number + """ + residual = src + if self.normalize_before: + src = self.norm1(src) + src2 = self.self_attn( + src, + src, + src, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + )[0] + src = residual + self.dropout1(src2) + if not self.normalize_before: + src = self.norm1(src) + + residual = src + if self.normalize_before: + src = self.norm2(src) + src2 = self.linear2(self.dropout(self.activation(self.linear1(src)))) + src = residual + self.dropout2(src2) + if not self.normalize_before: + src = self.norm2(src) + return src + + +def _get_activation_fn(activation: str): + if activation == "relu": + return nn.functional.relu + elif activation == "gelu": + return nn.functional.gelu + + raise RuntimeError( + "activation should be relu/gelu, not {}".format(activation) + ) + + +class PositionalEncoding(nn.Module): + """This class implements the positional encoding + proposed in the following paper: + + - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf + + PE(pos, 2i) = sin(pos / (10000^(2i/d_modle)) + PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle)) + + Note:: + + 1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model))) + = exp(-1* 2i / d_model * log(100000)) + = exp(2i * -(log(10000) / d_model)) + """ + + def __init__(self, d_model: int, dropout: float = 0.1) -> None: + """ + Args: + d_model: + Embedding dimension. + dropout: + Dropout probability to be applied to the output of this module. + """ + super().__init__() + self.d_model = d_model + self.xscale = math.sqrt(self.d_model) + self.dropout = nn.Dropout(p=dropout) + # not doing: self.pe = None because of errors thrown by torchscript + self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32) + + def extend_pe(self, x: torch.Tensor) -> None: + """Extend the time t in the positional encoding if required. + + The shape of `self.pe` is (1, T1, d_model). The shape of the input x + is (N, T, d_model). If T > T1, then we change the shape of self.pe + to (N, T, d_model). Otherwise, nothing is done. + + Args: + x: + It is a tensor of shape (N, T, C). + Returns: + Return None. + """ + if self.pe is not None: + if self.pe.size(1) >= x.size(1): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32) + position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) + div_term = torch.exp( + torch.arange(0, self.d_model, 2, dtype=torch.float32) + * -(math.log(10000.0) / self.d_model) + ) + pe[:, 0::2] = torch.sin(position * div_term) + pe[:, 1::2] = torch.cos(position * div_term) + pe = pe.unsqueeze(0) + # Now pe is of shape (1, T, d_model), where T is x.size(1) + self.pe = pe.to(device=x.device, dtype=x.dtype) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """ + Add positional encoding. + + Args: + x: + Its shape is (N, T, C) + + Returns: + Return a tensor of shape (N, T, C) + """ + self.extend_pe(x) + x = x * self.xscale + self.pe[:, : x.size(1), :] + return self.dropout(x) + + +class Noam(object): + """ + Implements Noam optimizer. + + Proposed in + "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf + + Modified from + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py # noqa + + Args: + params: + iterable of parameters to optimize or dicts defining parameter groups + model_size: + attention dimension of the transformer model + factor: + learning rate factor + warm_step: + warmup steps + """ + + def __init__( + self, + params, + model_size: int = 256, + factor: float = 10.0, + warm_step: int = 25000, + weight_decay=0, + ) -> None: + """Construct an Noam object.""" + self.optimizer = torch.optim.Adam( + params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay + ) + self._step = 0 + self.warmup = warm_step + self.factor = factor + self.model_size = model_size + self._rate = 0 + + @property + def param_groups(self): + """Return param_groups.""" + return self.optimizer.param_groups + + def step(self): + """Update parameters and rate.""" + self._step += 1 + rate = self.rate() + for p in self.optimizer.param_groups: + p["lr"] = rate + self._rate = rate + self.optimizer.step() + + def rate(self, step=None): + """Implement `lrate` above.""" + if step is None: + step = self._step + return ( + self.factor + * self.model_size ** (-0.5) + * min(step ** (-0.5), step * self.warmup ** (-1.5)) + ) + + def zero_grad(self): + """Reset gradient.""" + self.optimizer.zero_grad() + + def state_dict(self): + """Return state_dict.""" + return { + "_step": self._step, + "warmup": self.warmup, + "factor": self.factor, + "model_size": self.model_size, + "_rate": self._rate, + "optimizer": self.optimizer.state_dict(), + } + + def load_state_dict(self, state_dict): + """Load state_dict.""" + for key, value in state_dict.items(): + if key == "optimizer": + self.optimizer.load_state_dict(state_dict["optimizer"]) + else: + setattr(self, key, value)