diff --git a/.flake8 b/.flake8 index dbeec0b0c..c7c4f1855 100644 --- a/.flake8 +++ b/.flake8 @@ -9,6 +9,7 @@ per-file-ignores = egs/*/ASR/pruned_transducer_stateless*/*.py: E501, egs/*/ASR/*/optim.py: E501, egs/*/ASR/*/scaling.py: E501, + egs/librispeech/ASR/conv_emformer_transducer_stateless/*.py: E501, E203 # invalid escape sequence (cause by tex formular), W605 icefall/utils.py: E501, W605 diff --git a/egs/librispeech/ASR/README.md b/egs/librispeech/ASR/README.md index e2aaa9d7e..318d908d1 100644 --- a/egs/librispeech/ASR/README.md +++ b/egs/librispeech/ASR/README.md @@ -23,6 +23,7 @@ The following table lists the differences among them. | `pruned_transducer_stateless5` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner| | `pruned_transducer_stateless6` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + distillation with hubert| | `pruned_stateless_emformer_rnnt2` | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR| +| `conv_emformer_transducer_stateless` | Emformer | Embedding + Conv1d | Using Emformer augmented with convolution for streaming ASR + mechanisms in reworked model | The decoder in `transducer_stateless` is modified from the paper diff --git a/egs/librispeech/ASR/RESULTS.md b/egs/librispeech/ASR/RESULTS.md index 66410ef40..5eb07fae5 100644 --- a/egs/librispeech/ASR/RESULTS.md +++ b/egs/librispeech/ASR/RESULTS.md @@ -1,5 +1,165 @@ ## Results +### LibriSpeech BPE training results (Pruned Stateless Conv-Emformer RNN-T) + +[conv_emformer_transducer_stateless](./conv_emformer_transducer_stateless) + +It implements [Emformer](https://arxiv.org/abs/2010.10759) augmented with convolution module for streaming ASR. +It is modified from [torchaudio](https://github.com/pytorch/audio). + +See for more details. + +#### Training on full librispeech + +In this model, the lengths of chunk and right context are 32 frames (i.e., 0.32s) and 8 frames (i.e., 0.08s), respectively. + +The WERs are: + +| | test-clean | test-other | comment | decoding mode | +|-------------------------------------|------------|------------|----------------------|----------------------| +| greedy search (max sym per frame 1) | 3.63 | 9.61 | --epoch 30 --avg 10 | simulated streaming | +| greedy search (max sym per frame 1) | 3.64 | 9.65 | --epoch 30 --avg 10 | streaming | +| fast beam search | 3.61 | 9.4 | --epoch 30 --avg 10 | simulated streaming | +| fast beam search | 3.58 | 9.5 | --epoch 30 --avg 10 | streaming | +| modified beam search | 3.56 | 9.41 | --epoch 30 --avg 10 | simulated streaming | +| modified beam search | 3.54 | 9.46 | --epoch 30 --avg 10 | streaming | + +The training command is: + +```bash +./conv_emformer_transducer_stateless/train.py \ + --world-size 6 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --full-libri 1 \ + --max-duration 300 \ + --master-port 12321 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 +``` + +The tensorboard log can be found at + + +The simulated streaming decoding command using greedy search is: +```bash +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True +``` + +The simulated streaming decoding command using fast beam search is: +```bash +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +``` + +The simulated streaming decoding command using modified beam search is: +```bash +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 +``` + +The streaming decoding command using greedy search is: +```bash +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True +``` + +The streaming decoding command using fast beam search is: +```bash +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +``` + +The streaming decoding command using modified beam search is: +```bash +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 +``` + +Pretrained models, training logs, decoding logs, and decoding results +are available at + + ### LibriSpeech BPE training results (Pruned Stateless Emformer RNN-T) [pruned_stateless_emformer_rnnt2](./pruned_stateless_emformer_rnnt2) @@ -280,12 +440,12 @@ The WERs are: | | test-clean | test-other | comment | |-------------------------------------|------------|------------|-------------------------------------------------------------------------------| -| greedy search (max sym per frame 1) | 2.75 | 6.74 | --epoch 30 --avg 6 --use_averaged_model False | -| greedy search (max sym per frame 1) | 2.69 | 6.64 | --epoch 30 --avg 6 --use_averaged_model True | -| fast beam search | 2.72 | 6.67 | --epoch 30 --avg 6 --use_averaged_model False | -| fast beam search | 2.66 | 6.6 | --epoch 30 --avg 6 --use_averaged_model True | -| modified beam search | 2.67 | 6.68 | --epoch 30 --avg 6 --use_averaged_model False | -| modified beam search | 2.62 | 6.57 | --epoch 30 --avg 6 --use_averaged_model True | +| greedy search (max sym per frame 1) | 2.75 | 6.74 | --epoch 30 --avg 6 --use-averaged-model False | +| greedy search (max sym per frame 1) | 2.69 | 6.64 | --epoch 30 --avg 6 --use-averaged-model True | +| fast beam search | 2.72 | 6.67 | --epoch 30 --avg 6 --use-averaged-model False | +| fast beam search | 2.66 | 6.6 | --epoch 30 --avg 6 --use-averaged-model True | +| modified beam search | 2.67 | 6.68 | --epoch 30 --avg 6 --use-averaged-model False | +| modified beam search | 2.62 | 6.57 | --epoch 30 --avg 6 --use-averaged-model True | The training command is: diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py new file mode 120000 index 000000000..a074d6085 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py new file mode 100755 index 000000000..aadac2ae4 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py @@ -0,0 +1,657 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True + +(2) modified beam search +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 + +(3) fast beam search +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=10, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless4/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + feature_lens += params.right_context_length + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.right_context_length), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): hyps + } + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + 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, + sp=sp, + decoding_graph=decoding_graph, + 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) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{params.decoding_method}-beam-size-{params.beam_size}" + ) + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0 + start = params.epoch - params.avg + assert start >= 1 + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py new file mode 120000 index 000000000..0793c5709 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py new file mode 100644 index 000000000..46993da48 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py @@ -0,0 +1,1898 @@ +# Copyright 2022 Xiaomi Corporation (Author: Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# +# It is modified based on +# 1) https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py # noqa +# 2) https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa + +import math +from typing import List, Optional, Tuple + +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv1d, + ScaledConv2d, + ScaledLinear, +) + +from icefall.utils import make_pad_mask + + +LOG_EPSILON = math.log(1e-10) + + +def unstack_states( + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] +) -> List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]: + """Unstack the emformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Args: + states: + A tuple of 2 elements. + ``states[0]`` is the attention caches of a batch of utterance. + ``states[1]`` is the convolution caches of a batch of utterance. + ``len(states[0])`` and ``len(states[1])`` both eqaul to number of layers. # noqa + + Returns: + A list of states. + ``states[i]`` is a tuple of 2 elements of i-th utterance. + ``states[i][0]`` is the attention caches of i-th utterance. + ``states[i][1]`` is the convolution caches of i-th utterance. + ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers. # noqa + """ + + attn_caches, conv_caches = states + batch_size = conv_caches[0].size(0) + num_layers = len(attn_caches) + + list_attn_caches = [None] * batch_size + for i in range(batch_size): + list_attn_caches[i] = [[] for _ in range(num_layers)] + for li, layer in enumerate(attn_caches): + for s in layer: + s_list = s.unbind(dim=1) + for bi, b in enumerate(list_attn_caches): + b[li].append(s_list[bi]) + + list_conv_caches = [None] * batch_size + for i in range(batch_size): + list_conv_caches[i] = [None] * num_layers + for li, layer in enumerate(conv_caches): + c_list = layer.unbind(dim=0) + for bi, b in enumerate(list_conv_caches): + b[li] = c_list[bi] + + ans = [None] * batch_size + for i in range(batch_size): + ans[i] = [list_attn_caches[i], list_conv_caches[i]] + + return ans + + +def stack_states( + state_list: List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]] +) -> Tuple[List[List[torch.Tensor]], List[torch.Tensor]]: + """Stack list of emformer states that correspond to separate utterances + into a single emformer state so that it can be used as an input for + emformer when those utterances are formed into a batch. + + Note: + It is the inverse of :func:`unstack_states`. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the emformer model for a single utterance. + ``states[i]`` is a tuple of 2 elements of i-th utterance. + ``states[i][0]`` is the attention caches of i-th utterance. + ``states[i][1]`` is the convolution caches of i-th utterance. + ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers. # noqa + + Returns: + A new state corresponding to a batch of utterances. + See the input argument of :func:`unstack_states` for the meaning + of the returned tensor. + """ + batch_size = len(state_list) + + attn_caches = [] + for layer in state_list[0][0]: + if batch_size > 1: + # Note: We will stack attn_caches[layer][s][] later to get attn_caches[layer][s] # noqa + attn_caches.append([[s] for s in layer]) + else: + attn_caches.append([s.unsqueeze(1) for s in layer]) + for b, states in enumerate(state_list[1:], 1): + for li, layer in enumerate(states[0]): + for si, s in enumerate(layer): + attn_caches[li][si].append(s) + if b == batch_size - 1: + attn_caches[li][si] = torch.stack( + attn_caches[li][si], dim=1 + ) + + conv_caches = [] + for layer in state_list[0][1]: + if batch_size > 1: + # Note: We will stack conv_caches[layer][] later to get conv_caches[layer] # noqa + conv_caches.append([layer]) + else: + conv_caches.append(layer.unsqueeze(0)) + for b, states in enumerate(state_list[1:], 1): + for li, layer in enumerate(states[1]): + conv_caches[li].append(layer) + if b == batch_size - 1: + conv_caches[li] = torch.stack(conv_caches[li], dim=0) + + return [attn_caches, conv_caches] + + +class ConvolutionModule(nn.Module): + """ConvolutionModule. + + Modified from https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa + + Args: + chunk_length (int): + Length of each chunk. + right_context_length (int): + Length of right context. + channels (int): + The number of input channels and output 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, + chunk_length: int, + right_context_length: int, + channels: int, + kernel_size: int, + bias: bool = True, + ) -> None: + """Construct an ConvolutionModule object.""" + super().__init__() + # kernerl_size should be an odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0, kernel_size + + self.chunk_length = chunk_length + self.right_context_length = right_context_length + self.channels = channels + + self.pointwise_conv1 = ScaledConv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + # After pointwise_conv1 we put x through a gated linear unit + # (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in + # the range 1 to 4, but sometimes, for some reason, for layer 0 the rms + # ends up being very large, between 50 and 100 for different channels. + # This will cause very peaky and sparse derivatives for the sigmoid + # gating function, which will tend to make the loss function not learn + # effectively. (for most layers the average absolute values are in the + # range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for + # different layers, which likely breaks down as 0.5 for the "linear" + # half and 0.2 to 0.3 for the part that goes into the sigmoid. + # The idea is that if we constrain the rms values to a reasonable range + # via a constraint of max_abs=10.0, it will be in a better position to + # start learning something, i.e. to latch onto the correct range. + self.deriv_balancer1 = ActivationBalancer( + channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0 + ) + + # make it causal by padding cached (kernel_size - 1) frames on the left + self.cache_size = kernel_size - 1 + self.depthwise_conv = ScaledConv1d( + channels, + channels, + kernel_size, + stride=1, + padding=0, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channel_dim=1, min_positive=0.05, max_positive=1.0 + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.25, + ) + + def _split_right_context( + self, + pad_utterance: torch.Tensor, + right_context: torch.Tensor, + ) -> torch.Tensor: + """ + Args: + pad_utterance: + Its shape is (cache_size + U, B, D). + right_context: + Its shape is (R, B, D). + + Returns: + Right context segments padding with corresponding context. + Its shape is (num_segs * B, D, cache_size + right_context_length). + """ + U_, B, D = pad_utterance.size() + R = right_context.size(0) + assert self.right_context_length != 0 + assert R % self.right_context_length == 0 + num_chunks = R // self.right_context_length + right_context = right_context.reshape( + num_chunks, self.right_context_length, B, D + ) + right_context = right_context.permute(0, 2, 1, 3).reshape( + num_chunks * B, self.right_context_length, D + ) + + intervals = torch.arange( + 0, self.chunk_length * (num_chunks - 1), self.chunk_length + ) + first = torch.arange( + self.chunk_length, self.chunk_length + self.cache_size + ) + indexes = intervals.unsqueeze(1) + first.unsqueeze(0) + indexes = torch.cat( + [indexes, torch.arange(U_ - self.cache_size, U_).unsqueeze(0)] + ) + padding = pad_utterance[indexes] # (num_chunks, cache_size, B, D) + padding = padding.permute(0, 2, 1, 3).reshape( + num_chunks * B, self.cache_size, D + ) + + pad_right_context = torch.cat([padding, right_context], dim=1) + # (num_chunks * B, cache_size + right_context_length, D) + return pad_right_context.permute(0, 2, 1) + + def _merge_right_context( + self, right_context: torch.Tensor, B: int + ) -> torch.Tensor: + """ + Args: + right_context: + Right context segments. + It shape is (num_segs * B, D, right_context_length). + B: + Batch size. + + Returns: + A tensor of shape (B, D, R), where + R = num_segs * right_context_length. + """ + right_context = right_context.reshape( + -1, B, self.channels, self.right_context_length + ) + right_context = right_context.permute(1, 2, 0, 3) + right_context = right_context.reshape(B, self.channels, -1) + return right_context + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Causal convolution module. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + + Returns: + A tuple of 2 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv and GLU mechanism + x = torch.cat([right_context, utterance], dim=0) # (R + U, B, D) + x = x.permute(1, 2, 0) # (B, D, R + U) + x = self.pointwise_conv1(x) # (B, 2 * D, R + U) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, R + U) + utterance = x[:, :, R:] # (B, D, U) + right_context = x[:, :, :R] # (B, D, R) + + # make causal convolution + cache = torch.zeros( + B, D, self.cache_size, device=x.device, dtype=x.dtype + ) + pad_utterance = torch.cat( + [cache, utterance], dim=2 + ) # (B, D, cache + U) + + # depth-wise conv on utterance + utterance = self.depthwise_conv(pad_utterance) # (B, D, U) + + if self.right_context_length > 0: + # depth-wise conv on right_context + pad_right_context = self._split_right_context( + pad_utterance.permute(2, 0, 1), right_context.permute(2, 0, 1) + ) # (num_segs * B, D, cache_size + right_context_length) + right_context = self.depthwise_conv( + pad_right_context + ) # (num_segs * B, D, right_context_length) + right_context = self._merge_right_context( + right_context, B + ) # (B, D, R) + + x = torch.cat([right_context, utterance], dim=2) # (B, D, R + U) + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, R + U) + + right_context = x[:, :, :R] # (B, D, R) + utterance = x[:, :, R:] # (B, D, U) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + ) + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + cache: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Causal convolution module applied on both utterance and right_context. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + cache (torch.Tensor, optional): + Cached tensor for left padding of shape (B, D, cache_size). + + Returns: + A tuple of 3 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + - updated cache tensor of shape (B, D, cache_size). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv + x = torch.cat([utterance, right_context], dim=0) # (U + R, B, D) + x = x.permute(1, 2, 0) # (B, D, U + R) + x = self.pointwise_conv1(x) # (B, 2 * D, U + R) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, U + R) + + # make causal convolution + assert cache.shape == (B, D, self.cache_size), cache.shape + x = torch.cat([cache, x], dim=2) # (B, D, cache_size + U + R) + # update cache + new_cache = x[:, :, -R - self.cache_size : -R] + + # 1-D depth-wise conv + x = self.depthwise_conv(x) # (B, D, U + R) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, U + R) + + utterance = x[:, :, :U] # (B, D, U) + right_context = x[:, :, U:] # (B, D, R) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + new_cache, + ) + + +class EmformerAttention(nn.Module): + r"""Emformer layer attention module. + + Args: + embed_dim (int): + Embedding dimension. + nhead (int): + Number of attention heads in each Emformer layer. + dropout (float, optional): + Dropout probability. (Default: 0.0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + embed_dim: int, + nhead: int, + dropout: float = 0.0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + if embed_dim % nhead != 0: + raise ValueError( + f"embed_dim ({embed_dim}) is not a multiple of" + f"nhead ({nhead})." + ) + + self.embed_dim = embed_dim + self.nhead = nhead + self.tanh_on_mem = tanh_on_mem + self.negative_inf = negative_inf + self.head_dim = embed_dim // nhead + self.dropout = dropout + + self.emb_to_key_value = ScaledLinear( + embed_dim, 2 * embed_dim, bias=True + ) + self.emb_to_query = ScaledLinear(embed_dim, embed_dim, bias=True) + self.out_proj = ScaledLinear( + embed_dim, embed_dim, bias=True, initial_scale=0.25 + ) + + def _gen_attention_probs( + self, + attention_weights: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Given the entire attention weights, mask out unecessary connections + and optionally with padding positions, to obtain underlying chunk-wise + attention probabilities. + + B: batch size; + Q: length of query; + KV: length of key and value. + + Args: + attention_weights (torch.Tensor): + Attention weights computed on the entire concatenated tensor + with shape (B * nhead, Q, KV). + attention_mask (torch.Tensor): + Mask tensor where chunk-wise connections are filled with `False`, + and other unnecessary connections are filled with `True`, + with shape (Q, KV). + padding_mask (torch.Tensor, optional): + Mask tensor where the padding positions are fill with `True`, + and other positions are filled with `False`, with shapa `(B, KV)`. + + Returns: + A tensor of shape (B * nhead, Q, KV). + """ + attention_weights_float = attention_weights.float() + attention_weights_float = attention_weights_float.masked_fill( + attention_mask.unsqueeze(0), self.negative_inf + ) + if padding_mask is not None: + Q = attention_weights.size(1) + B = attention_weights.size(0) // self.nhead + attention_weights_float = attention_weights_float.view( + B, self.nhead, Q, -1 + ) + attention_weights_float = attention_weights_float.masked_fill( + padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + self.negative_inf, + ) + attention_weights_float = attention_weights_float.view( + B * self.nhead, Q, -1 + ) + + attention_probs = nn.functional.softmax( + attention_weights_float, dim=-1 + ).type_as(attention_weights) + + attention_probs = nn.functional.dropout( + attention_probs, p=self.dropout, training=self.training + ) + return attention_probs + + def _forward_impl( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + left_context_key: Optional[torch.Tensor] = None, + left_context_val: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Underlying chunk-wise attention implementation.""" + U, B, _ = utterance.size() + R = right_context.size(0) + M = memory.size(0) + scaling = float(self.head_dim) ** -0.5 + + # compute query with [right_context, utterance, summary]. + query = self.emb_to_query( + torch.cat([right_context, utterance, summary]) + ) + # compute key and value with [memory, right_context, utterance]. + key, value = self.emb_to_key_value( + torch.cat([memory, right_context, utterance]) + ).chunk(chunks=2, dim=2) + + if left_context_key is not None and left_context_val is not None: + # now compute key and value with + # [memory, right context, left context, uttrance] + # this is used in inference mode + key = torch.cat([key[: M + R], left_context_key, key[M + R :]]) + value = torch.cat( + [value[: M + R], left_context_val, value[M + R :]] + ) + Q = query.size(0) + # KV = key.size(0) + + reshaped_query, reshaped_key, reshaped_value = [ + tensor.contiguous() + .view(-1, B * self.nhead, self.head_dim) + .transpose(0, 1) + for tensor in [query, key, value] + ] # (B * nhead, Q or KV, head_dim) + attention_weights = torch.bmm( + reshaped_query * scaling, reshaped_key.transpose(1, 2) + ) # (B * nhead, Q, KV) + + # compute attention probabilities + attention_probs = self._gen_attention_probs( + attention_weights, attention_mask, padding_mask + ) + + # compute attention outputs + attention = torch.bmm(attention_probs, reshaped_value) + assert attention.shape == (B * self.nhead, Q, self.head_dim) + attention = ( + attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim) + ) + + # apply output projection + outputs = self.out_proj(attention) + + output_right_context_utterance = outputs[: R + U] + output_memory = outputs[R + U :] + if self.tanh_on_mem: + output_memory = torch.tanh(output_memory) + else: + output_memory = torch.clamp(output_memory, min=-10, max=10) + + return output_right_context_utterance, output_memory, key, value + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + # TODO: Modify docs. + """Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of the hard-copied right contexts; + U: length of full utterance; + S: length of summary vectors; + M: length of memory vectors. + + It computes a `big` attention matrix on full utterance and + then utilizes a pre-computed mask to simulate chunk-wise attention. + + It concatenates three blocks: hard-copied right contexts, + full utterance, and summary vectors, as a `big` block, + to compute the query tensor: + query = [right_context, utterance, summary], + with length Q = R + U + S. + It concatenates the three blocks: memory vectors, + hard-copied right contexts, and full utterance as another `big` block, + to compute the key and value tensors: + key & value = [memory, right_context, utterance], + with length KV = M + R + U. + Attention scores is computed with above `big` query and key. + + Then the underlying chunk-wise attention is obtained by applying + the attention mask. Suppose + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + s_i: summary vector of c_i; + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); + s_i (in query) -> l_i, c_i, r_i (in key). + + Args: + utterance (torch.Tensor): + Full utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Hard-copied right context frames, with shape (R, B, D), + where R = num_chunks * right_context_length + summary (torch.Tensor): + Summary elements with shape (S, B, D), where S = num_chunks. + It is an empty tensor without using memory. + memory (torch.Tensor): + Memory elements, with shape (M, B, D), where M = num_chunks - 1. + It is an empty tensor without using memory. + attention_mask (torch.Tensor): + Pre-computed attention mask to simulate underlying chunk-wise + attention, with shape (Q, KV). + padding_mask (torch.Tensor): + Padding mask of key tensor, with shape (B, KV). + + Returns: + A tuple containing 2 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (M, B, D), where M = S - 1 or M = 0. + """ + ( + output_right_context_utterance, + output_memory, + _, + _, + ) = self._forward_impl( + utterance, + right_context, + summary, + memory, + attention_mask, + padding_mask=padding_mask, + ) + return output_right_context_utterance, output_memory[:-1] + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + left_context_key: torch.Tensor, + left_context_val: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right context; + U: length of utterance, i.e., current chunk; + L: length of cached left context; + S: length of summary vectors, S = 1; + M: length of cached memory vectors. + + It concatenates the right context, utterance (i.e., current chunk) + and summary vector of current chunk, to compute the query tensor: + query = [right_context, utterance, summary], + with length Q = R + U + S. + It concatenates the memory vectors, right context, left context, and + current chunk, to compute the key and value tensors: + key & value = [memory, right_context, left_context, utterance], + with length KV = M + R + L + U. + + The chunk-wise attention is: + chunk, right context (in query) -> + left context, chunk, right context, memory vectors (in key); + summary (in query) -> left context, chunk, right context (in key). + + Args: + utterance (torch.Tensor): + Current chunk frames, with shape (U, B, D), where U = chunk_length. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D), + where R = right_context_length. + summary (torch.Tensor): + Summary vector with shape (1, B, D), or empty tensor. + memory (torch.Tensor): + Memory vectors, with shape (M, B, D), or empty tensor. + left_context_key (torch,Tensor): + Cached attention key of left context from preceding computation, + with shape (L, B, D). + left_context_val (torch.Tensor): + Cached attention value of left context from preceding computation, + with shape (L, B, D). + padding_mask (torch.Tensor): + Padding mask of key tensor, with shape (B, KV). + + Returns: + A tuple containing 4 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (1, B, D) or (0, B, D). + - attention key of left context and utterance, which would be cached + for next computation, with shape (L + U, B, D). + - attention value of left context and utterance, which would be + cached for next computation, with shape (L + U, B, D). + """ + U = utterance.size(0) + R = right_context.size(0) + L = left_context_key.size(0) + S = summary.size(0) + M = memory.size(0) + + # TODO: move it outside + # query = [right context, utterance, summary] + Q = R + U + S + # key, value = [memory, right context, left context, uttrance] + KV = M + R + L + U + attention_mask = torch.zeros(Q, KV).to( + dtype=torch.bool, device=utterance.device + ) + # disallow attention bettween the summary vector with the memory bank + attention_mask[-1, :M] = True + ( + output_right_context_utterance, + output_memory, + key, + value, + ) = self._forward_impl( + utterance, + right_context, + summary, + memory, + attention_mask, + padding_mask=padding_mask, + left_context_key=left_context_key, + left_context_val=left_context_val, + ) + return ( + output_right_context_utterance, + output_memory, + key[M + R :], + value[M + R :], + ) + + +class EmformerEncoderLayer(nn.Module): + """Emformer layer that constitutes Emformer. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads. + dim_feedforward (int): + Hidden layer dimension of feedforward network. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (Default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (Default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (Default: 0) + right_context_length (int, optional): + Length of right context. (Default: 0) + memory_size (int, optional): + Number of memory elements to use. (Default: 0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int, + chunk_length: int, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.attention = EmformerAttention( + embed_dim=d_model, + nhead=nhead, + dropout=dropout, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + self.summary_op = nn.AvgPool1d( + kernel_size=chunk_length, stride=chunk_length, ceil_mode=True + ) + + self.feed_forward_macaron = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.feed_forward = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.conv_module = ConvolutionModule( + chunk_length, + right_context_length, + d_model, + cnn_module_kernel, + ) + + self.norm_final = BasicNorm(d_model) + + # try to ensure the output is close to zero-mean + # (or at least, zero-median). + self.balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0 + ) + + self.dropout = nn.Dropout(dropout) + + self.layer_dropout = layer_dropout + self.left_context_length = left_context_length + self.chunk_length = chunk_length + self.memory_size = memory_size + self.d_model = d_model + self.use_memory = memory_size > 0 + + def _update_attn_cache( + self, + next_key: torch.Tensor, + next_val: torch.Tensor, + memory: torch.Tensor, + attn_cache: List[torch.Tensor], + ) -> List[torch.Tensor]: + """Update cached attention state: + 1) output memory of current chunk in the lower layer; + 2) attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + """ + new_memory = torch.cat([attn_cache[0], memory]) + new_key = torch.cat([attn_cache[1], next_key]) + new_val = torch.cat([attn_cache[2], next_val]) + attn_cache[0] = new_memory[new_memory.size(0) - self.memory_size :] + attn_cache[1] = new_key[new_key.size(0) - self.left_context_length :] + attn_cache[2] = new_val[new_val.size(0) - self.left_context_length :] + return attn_cache + + def _apply_conv_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + ) -> torch.Tensor: + """Apply convolution module in training and validation mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context = self.conv_module(utterance, right_context) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance + + def _apply_conv_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + conv_cache: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply convolution module on utterance in inference mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context, conv_cache = self.conv_module.infer( + utterance, right_context, conv_cache + ) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance, conv_cache + + def _apply_attention_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply attention module in training and validation mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + if self.use_memory: + summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + ) + else: + summary = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + output_right_context_utterance, output_memory = self.attention( + utterance=utterance, + right_context=right_context, + summary=summary, + memory=memory, + attention_mask=attention_mask, + padding_mask=padding_mask, + ) + + return output_right_context_utterance, output_memory + + def _apply_attention_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + memory: torch.Tensor, + attn_cache: List[torch.Tensor], + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]: + """Apply attention module in inference mode. + 1) Unpack cached states including: + - memory from previous chunks in the lower layer; + - attention key and value of left context from preceding + chunk's compuation; + 2) Apply attention computation; + 3) Update cached attention states including: + - output memory of current chunk in the lower layer; + - attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + """ + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + pre_memory = attn_cache[0] + left_context_key = attn_cache[1] + left_context_val = attn_cache[2] + + if self.use_memory: + summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + ) + summary = summary[:1] + else: + summary = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + ( + output_right_context_utterance, + output_memory, + next_key, + next_val, + ) = self.attention.infer( + utterance=utterance, + right_context=right_context, + summary=summary, + memory=pre_memory, + left_context_key=left_context_key, + left_context_val=left_context_val, + padding_mask=padding_mask, + ) + attn_cache = self._update_attn_cache( + next_key, next_val, memory, attn_cache + ) + return output_right_context_utterance, output_memory, attn_cache + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + r"""Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of hard-copied right contexts; + U: length of full utterance; + M: length of memory vectors. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + It is an empty tensor without using memory. + attention_mask (torch.Tensor): + Attention mask for underlying attention module, + with shape (Q, KV), where Q = R + U + S, KV = M + R + U. + padding_mask (torch.Tensor): + Padding mask of ker tensor, with shape (B, KV). + + Returns: + A tuple containing 3 tensors: + - output utterance, with shape (U, B, D). + - output right context, with shape (R, B, D). + - output memory, with shape (M, B, D). + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + src_orig = src + + warmup_scale = min(0.1 + warmup, 1.0) + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + if self.training: + alpha = ( + warmup_scale + if torch.rand(()).item() <= (1.0 - self.layer_dropout) + else 0.1 + ) + else: + alpha = 1.0 + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + src_att, output_memory = self._apply_attention_module_forward( + src, R, memory, attention_mask, padding_mask=padding_mask + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv = self._apply_conv_module_forward(src, R) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + output_utterance = src[R:] + output_right_context = src[:R] + return output_utterance, output_right_context, output_memory + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + attn_cache: List[torch.Tensor], + conv_cache: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + torch.Tensor, + List[torch.Tensor], + torch.Tensor, + ]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + M: length of memory. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + attn_cache (List[torch.Tensor]): + Cached attention tensors generated in preceding computation, + including memory, key and value of left context. + conv_cache (torch.Tensor, optional): + Cache tensor of left context for causal convolution. + padding_mask (torch.Tensor): + Padding mask of ker tensor. + + Returns: + (Tensor, Tensor, List[torch.Tensor], Tensor): + - output utterance, with shape (U, B, D); + - output right_context, with shape (R, B, D); + - output memory, with shape (1, B, D) or (0, B, D). + - output state. + - updated conv_cache. + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + ( + src_att, + output_memory, + attn_cache, + ) = self._apply_attention_module_infer( + src, R, memory, attn_cache, padding_mask=padding_mask + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv, conv_cache = self._apply_conv_module_infer(src, R, conv_cache) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + output_utterance = src[R:] + output_right_context = src[:R] + return ( + output_utterance, + output_right_context, + output_memory, + attn_cache, + conv_cache, + ) + + +def _gen_attention_mask_block( + col_widths: List[int], + col_mask: List[bool], + num_rows: int, + device: torch.device, +) -> torch.Tensor: + assert len(col_widths) == len( + col_mask + ), "Length of col_widths must match that of col_mask" + + mask_block = [ + torch.ones(num_rows, col_width, device=device) + if is_ones_col + else torch.zeros(num_rows, col_width, device=device) + for col_width, is_ones_col in zip(col_widths, col_mask) + ] + return torch.cat(mask_block, dim=1) + + +class EmformerEncoder(nn.Module): + """Implements the Emformer architecture introduced in + *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency + Streaming Speech Recognition* + [:footcite:`shi2021emformer`]. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads in each emformer layer. + dim_feedforward (int): + Hidden layer dimension of each emformer layer's feedforward network. + num_encoder_layers (int): + Number of emformer layers to instantiate. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (default: 0) + right_context_length (int, optional): + Length of right context. (default: 0) + memory_size (int, optional): + Number of memory elements to use. (default: 0) + tanh_on_mem (bool, optional): + If ``true``, applies tanh to memory elements. (default: ``false``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (default: -1e8) + """ + + def __init__( + self, + chunk_length: int, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.use_memory = memory_size > 0 + self.init_memory_op = nn.AvgPool1d( + kernel_size=chunk_length, + stride=chunk_length, + ceil_mode=True, + ) + + self.emformer_layers = nn.ModuleList( + [ + EmformerEncoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + chunk_length=chunk_length, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + for layer_idx in range(num_encoder_layers) + ] + ) + + self.num_encoder_layers = num_encoder_layers + self.d_model = d_model + self.left_context_length = left_context_length + self.right_context_length = right_context_length + self.chunk_length = chunk_length + self.memory_size = memory_size + self.cnn_module_kernel = cnn_module_kernel + + def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor: + """Hard copy each chunk's right context and concat them.""" + T = x.shape[0] + num_chunks = math.ceil( + (T - self.right_context_length) / self.chunk_length + ) + # first (num_chunks - 1) right context block + intervals = torch.arange( + 0, self.chunk_length * (num_chunks - 1), self.chunk_length + ) + first = torch.arange( + self.chunk_length, self.chunk_length + self.right_context_length + ) + indexes = intervals.unsqueeze(1) + first.unsqueeze(0) + # cat last right context block + indexes = torch.cat( + [ + indexes, + torch.arange(T - self.right_context_length, T).unsqueeze(0), + ] + ) + right_context_blocks = x[indexes.reshape(-1)] + return right_context_blocks + + def _gen_attention_mask_col_widths( + self, chunk_idx: int, U: int + ) -> List[int]: + """Calculate column widths (key, value) in attention mask for the + chunk_idx chunk.""" + num_chunks = math.ceil(U / self.chunk_length) + rc = self.right_context_length + lc = self.left_context_length + rc_start = chunk_idx * rc + rc_end = rc_start + rc + chunk_start = max(chunk_idx * self.chunk_length - lc, 0) + chunk_end = min((chunk_idx + 1) * self.chunk_length, U) + R = rc * num_chunks + + if self.use_memory: + m_start = max(chunk_idx - self.memory_size, 0) + M = num_chunks - 1 + col_widths = [ + m_start, # before memory + chunk_idx - m_start, # memory + M - chunk_idx, # after memory + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + else: + col_widths = [ + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + + return col_widths + + def _gen_attention_mask(self, utterance: torch.Tensor) -> torch.Tensor: + """Generate attention mask to simulate underlying chunk-wise attention + computation, where chunk-wise connections are filled with `False`, + and other unnecessary connections beyond chunk are filled with `True`. + + R: length of hard-copied right contexts; + U: length of full utterance; + S: length of summary vectors; + M: length of memory vectors; + Q: length of attention query; + KV: length of attention key and value. + + The shape of attention mask is (Q, KV). + If self.use_memory is `True`: + query = [right_context, utterance, summary]; + key, value = [memory, right_context, utterance]; + Q = R + U + S, KV = M + R + U. + Otherwise: + query = [right_context, utterance] + key, value = [right_context, utterance] + Q = R + U, KV = R + U. + + Suppose: + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + s_i: summary vector of c_i. + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); + s_i (in query) -> l_i, c_i, r_i (in key). + """ + U = utterance.size(0) + num_chunks = math.ceil(U / self.chunk_length) + + right_context_mask = [] + utterance_mask = [] + summary_mask = [] + + if self.use_memory: + num_cols = 9 + # right context and utterance both attend to memory, right context, + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4, 7] for idx in range(num_cols) + ] + # summary attends to right context, utterance + summary_cols_mask = [idx in [4, 7] for idx in range(num_cols)] + masks_to_concat = [right_context_mask, utterance_mask, summary_mask] + else: + num_cols = 6 + # right context and utterance both attend to right context and + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4] for idx in range(num_cols) + ] + summary_cols_mask = None + masks_to_concat = [right_context_mask, utterance_mask] + + for chunk_idx in range(num_chunks): + col_widths = self._gen_attention_mask_col_widths(chunk_idx, U) + + right_context_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + self.right_context_length, + utterance.device, + ) + right_context_mask.append(right_context_mask_block) + + utterance_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + min( + self.chunk_length, + U - chunk_idx * self.chunk_length, + ), + utterance.device, + ) + utterance_mask.append(utterance_mask_block) + + if summary_cols_mask is not None: + summary_mask_block = _gen_attention_mask_block( + col_widths, summary_cols_mask, 1, utterance.device + ) + summary_mask.append(summary_mask_block) + + attention_mask = ( + 1 - torch.cat([torch.cat(mask) for mask in masks_to_concat]) + ).to(torch.bool) + return attention_mask + + def forward( + self, x: torch.Tensor, lengths: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and validation mode. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + + Returns: + A tuple of 2 tensors: + - output utterance frames, with shape (U, B, D). + - output_lengths, with shape (B,), without containing the + right_context at the end. + """ + U = x.size(0) - self.right_context_length + + right_context = self._gen_right_context(x) + utterance = x[:U] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + attention_mask = self._gen_attention_mask(utterance) + memory = ( + self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[ + :-1 + ] + if self.use_memory + else torch.empty(0).to(dtype=x.dtype, device=x.device) + ) + padding_mask = make_pad_mask( + memory.size(0) + right_context.size(0) + output_lengths + ) + + output = utterance + for layer in self.emformer_layers: + output, right_context, memory = layer( + output, + right_context, + memory, + attention_mask, + padding_mask=padding_mask, + warmup=warmup, + ) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + lengths: torch.Tensor, + num_processed_frames: torch.Tensor, + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ]: + """Forward pass for streaming inference. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa + Cached states containing: + - past_lens: number of past frames for each sample in batch + - attn_caches: attention states from preceding chunk's computation, + where each element corresponds to each emformer layer + - conv_caches: left context for causal convolution, where each + element corresponds to each layer. + + Returns: + (Tensor, Tensor, List[List[torch.Tensor]], List[torch.Tensor]): + - output utterance frames, with shape (U, B, D). + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + """ + assert num_processed_frames.shape == (x.size(1),) + + attn_caches = states[0] + assert len(attn_caches) == self.num_encoder_layers, len(attn_caches) + for i in range(len(attn_caches)): + assert attn_caches[i][0].shape == ( + self.memory_size, + x.size(1), + self.d_model, + ), attn_caches[i][0].shape + assert attn_caches[i][1].shape == ( + self.left_context_length, + x.size(1), + self.d_model, + ), attn_caches[i][1].shape + assert attn_caches[i][2].shape == ( + self.left_context_length, + x.size(1), + self.d_model, + ), attn_caches[i][2].shape + + conv_caches = states[1] + assert len(conv_caches) == self.num_encoder_layers, len(conv_caches) + for i in range(len(conv_caches)): + assert conv_caches[i].shape == ( + x.size(1), + self.d_model, + self.cnn_module_kernel - 1, + ), conv_caches[i].shape + + right_context = x[-self.right_context_length :] + utterance = x[: -self.right_context_length] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + memory = ( + self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1) + if self.use_memory + else torch.empty(0).to(dtype=x.dtype, device=x.device) + ) + + # calcualte padding mask to mask out initial zero caches + chunk_mask = make_pad_mask(output_lengths).to(x.device) + memory_mask = ( + torch.div( + num_processed_frames, self.chunk_length, rounding_mode="floor" + ).view(x.size(1), 1) + <= torch.arange(self.memory_size, device=x.device).expand( + x.size(1), self.memory_size + ) + ).flip(1) + left_context_mask = ( + num_processed_frames.view(x.size(1), 1) + <= torch.arange(self.left_context_length, device=x.device).expand( + x.size(1), self.left_context_length + ) + ).flip(1) + right_context_mask = torch.zeros( + x.size(1), + self.right_context_length, + dtype=torch.bool, + device=x.device, + ) + padding_mask = torch.cat( + [memory_mask, right_context_mask, left_context_mask, chunk_mask], + dim=1, + ) + + output = utterance + output_attn_caches: List[List[torch.Tensor]] = [] + output_conv_caches: List[torch.Tensor] = [] + for layer_idx, layer in enumerate(self.emformer_layers): + ( + output, + right_context, + memory, + output_attn_cache, + output_conv_cache, + ) = layer.infer( + output, + right_context, + memory, + padding_mask=padding_mask, + attn_cache=attn_caches[layer_idx], + conv_cache=conv_caches[layer_idx], + ) + output_attn_caches.append(output_attn_cache) + output_conv_caches.append(output_conv_cache) + + output_states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = ( + output_attn_caches, + output_conv_caches, + ) + return output, output_lengths, output_states + + +class Emformer(EncoderInterface): + def __init__( + self, + num_features: int, + chunk_length: 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, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 3, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.subsampling_factor = subsampling_factor + self.right_context_length = right_context_length + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + if chunk_length % subsampling_factor != 0: + raise NotImplementedError( + "chunk_length must be a mutiple of subsampling_factor." + ) + if ( + left_context_length != 0 + and left_context_length % subsampling_factor != 0 + ): + raise NotImplementedError( + "left_context_length must be 0 or a mutiple of subsampling_factor." # noqa + ) + if ( + right_context_length != 0 + and right_context_length % subsampling_factor != 0 + ): + raise NotImplementedError( + "right_context_length must be 0 or a mutiple of subsampling_factor." # noqa + ) + + # 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 + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder = EmformerEncoder( + chunk_length=chunk_length // subsampling_factor, + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + num_encoder_layers=num_encoder_layers, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length // subsampling_factor, + right_context_length=right_context_length // subsampling_factor, + memory_size=memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and non-streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + x_lens (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + """ + x = self.encoder_embed(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + x_lens = (((x_lens - 1) >> 1) - 1) >> 1 + assert x.size(0) == x_lens.max().item() + + output, output_lengths = self.encoder( + x, x_lens, warmup=warmup + ) # (T, N, C) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + num_processed_frames: torch.Tensor, + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ]: + """Forward pass for streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa + Cached states containing: + - past_lens: number of past frames for each sample in batch + - attn_caches: attention states from preceding chunk's computation, + where each element corresponds to each emformer layer + - conv_caches: left context for causal convolution, where each + element corresponds to each layer. + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + """ + x = self.encoder_embed(x) + # drop the first and last frames + x = x[:, 1:-1, :] + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + x_lens = (((x_lens - 1) >> 1) - 1) >> 1 + x_lens -= 2 + assert x.size(0) == x_lens.max().item() + + num_processed_frames = num_processed_frames >> 2 + + output, output_lengths, output_states = self.encoder.infer( + x, x_lens, num_processed_frames, states + ) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths, output_states + + +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 + T' = ((T-1)//2 - 1)//2, which approximates T' == T//4 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + layer1_channels: int = 8, + layer2_channels: int = 32, + layer3_channels: int = 128, + ) -> None: + """ + Args: + in_channels: + Number of channels in. The input shape is (N, T, in_channels). + Caution: It requires: T >=7, in_channels >=7 + out_channels + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + """ + assert in_channels >= 7 + super().__init__() + + self.conv = nn.Sequential( + ScaledConv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=1, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ) + self.out = ScaledLinear( + layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels + ) + # set learn_eps=False because out_norm is preceded by `out`, and `out` + # itself has learned scale, so the extra degree of freedom is not + # needed. + self.out_norm = BasicNorm(out_channels, learn_eps=False) + # constrain median of output to be close to zero. + self.out_balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55 + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + 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) + x = self.conv(x) + # 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) + x = self.out_norm(x) + x = self.out_balancer(x) + return x diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py new file mode 120000 index 000000000..b9aa0ae08 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/export.py new file mode 100755 index 000000000..4930881ea --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/export.py @@ -0,0 +1,287 @@ +#!/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: +./conv_emformer_transducer_stateless/export.py \ + --exp-dir ./conv_emformer_transducer_stateless/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 30 \ + --avg 10 \ + --use-averaged-model=True \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --jit False + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `conv_emformer_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 + ./conv_emformer_transducer_stateless/decode.py \ + --exp-dir ./conv_emformer_transducer_stateless/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 100 \ + --bpe-model data/lang_bpe_500/bpe.model \ + --use-averaged-model=False \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for averaging. + Note: Epoch counts from 0. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/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", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + add_model_arguments(parser) + + return parser + + +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + + 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) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.eval() + + if params.jit: + # We won't use the forward() method of the model in C++, so just ignore + # it here. + # Otherwise, one of its arguments is a ragged tensor and is not + # torch scriptabe. + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torch.jit.script") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py new file mode 120000 index 000000000..815fd4bb6 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py new file mode 120000 index 000000000..ebb6d774d --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py new file mode 120000 index 000000000..e2deb4492 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py new file mode 120000 index 000000000..09d802cc4 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py new file mode 100644 index 000000000..31ad3f50a --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py @@ -0,0 +1,176 @@ +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from typing import List, Optional, Tuple + +import k2 +import torch +from beam_search import Hypothesis, HypothesisList + +from icefall.utils import AttributeDict + + +class Stream(object): + def __init__( + self, + params: AttributeDict, + decoding_graph: Optional[k2.Fsa] = None, + device: torch.device = torch.device("cpu"), + LOG_EPS: float = math.log(1e-10), + ) -> None: + """ + Args: + params: + It's the return value of :func:`get_params`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + device: + The device to run this stream. + """ + self.device = device + self.LOG_EPS = LOG_EPS + + # Containing attention caches and convolution caches + self.states: Optional[ + Tuple[List[List[torch.Tensor]], List[torch.Tensor]] + ] = None + # Initailize zero states. + self.init_states(params) + + # It uses different attributes for different decoding methods. + self.context_size = params.context_size + self.decoding_method = params.decoding_method + if params.decoding_method == "greedy_search": + self.hyp = [params.blank_id] * params.context_size + elif params.decoding_method == "modified_beam_search": + self.hyps = HypothesisList() + self.hyps.add( + Hypothesis( + ys=[params.blank_id] * params.context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + ) + ) + elif params.decoding_method == "fast_beam_search": + # feature_len is needed to get partial results. + # The rnnt_decoding_stream for fast_beam_search. + self.rnnt_decoding_stream: k2.RnntDecodingStream = ( + k2.RnntDecodingStream(decoding_graph) + ) + self.hyp: Optional[List[int]] = None + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + + self.ground_truth: str = "" + + self.feature: Optional[torch.Tensor] = None + # Make sure all feature frames can be used. + # Add 2 here since we will drop the first and last after subsampling. + self.chunk_length = params.chunk_length + self.pad_length = ( + params.right_context_length + 2 * params.subsampling_factor + 3 + ) + self.num_frames = 0 + self.num_processed_frames = 0 + + # After all feature frames are processed, we set this flag to True + self._done = False + + def set_feature(self, feature: torch.Tensor) -> None: + assert feature.dim() == 2, feature.dim() + self.num_frames = feature.size(0) + # tail padding + self.feature = torch.nn.functional.pad( + feature, + (0, 0, 0, self.pad_length), + mode="constant", + value=self.LOG_EPS, + ) + + def set_ground_truth(self, ground_truth: str) -> None: + self.ground_truth = ground_truth + + def init_states(self, params: AttributeDict) -> None: + attn_caches = [ + [ + torch.zeros( + params.memory_size, params.encoder_dim, device=self.device + ), + torch.zeros( + params.left_context_length // params.subsampling_factor, + params.encoder_dim, + device=self.device, + ), + torch.zeros( + params.left_context_length // params.subsampling_factor, + params.encoder_dim, + device=self.device, + ), + ] + for _ in range(params.num_encoder_layers) + ] + conv_caches = [ + torch.zeros( + params.encoder_dim, + params.cnn_module_kernel - 1, + device=self.device, + ) + for _ in range(params.num_encoder_layers) + ] + self.states = (attn_caches, conv_caches) + + def get_feature_chunk(self) -> torch.Tensor: + """Get a chunk of feature frames. + + Returns: + A tensor of shape (ret_length, feature_dim). + """ + update_length = min( + self.num_frames - self.num_processed_frames, self.chunk_length + ) + ret_length = update_length + self.pad_length + + ret_feature = self.feature[ + self.num_processed_frames : self.num_processed_frames + ret_length + ] + # Cut off used frames. + # self.feature = self.feature[update_length:] + + self.num_processed_frames += update_length + if self.num_processed_frames >= self.num_frames: + self._done = True + + return ret_feature + + @property + def done(self) -> bool: + """Return True if all feature frames are processed.""" + return self._done + + def decoding_result(self) -> List[int]: + """Obtain current decoding result.""" + if self.decoding_method == "greedy_search": + return self.hyp[self.context_size :] + elif self.decoding_method == "modified_beam_search": + best_hyp = self.hyps.get_most_probable(length_norm=True) + return best_hyp.ys[self.context_size :] + else: + assert self.decoding_method == "fast_beam_search" + return self.hyp diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py new file mode 100755 index 000000000..4fac405b0 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py @@ -0,0 +1,978 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True + +(2) modified beam search +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 + +(3) fast beam search +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" +import argparse +import logging +import warnings +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +from lhotse import CutSet +import numpy as np +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import Hypothesis, HypothesisList, get_hyps_shape +from emformer import LOG_EPSILON, stack_states, unstack_states +from kaldifeat import Fbank, FbankOptions +from stream import Stream +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import one_best_decoding +from icefall.utils import ( + AttributeDict, + get_texts, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_emformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An interger indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--sampling-rate", + type=float, + default=16000, + help="Sample rate of the audio", + ) + + parser.add_argument( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel", + ) + + add_model_arguments(parser) + + return parser + + +def greedy_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], +) -> None: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + + Args: + model: + The transducer model. + encoder_out: + Output from the encoder. Its shape is (N, T, C), where N >= 1. + streams: + A list of Stream objects. + """ + assert len(streams) == encoder_out.size(0) + assert encoder_out.ndim == 3 + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + T = encoder_out.size(1) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + # decoder_out is of shape (batch_size, 1, decoder_out_dim) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + for t in range(T): + # current_encoder_out's shape: (batch_size, 1, encoder_out_dim) + current_encoder_out = encoder_out[:, t : t + 1, :] # noqa + + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + # logits'shape (batch_size, vocab_size) + logits = logits.squeeze(1).squeeze(1) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + streams[i].hyp.append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder( + decoder_input, + need_pad=False, + ) + decoder_out = model.joiner.decoder_proj(decoder_out) + + +def modified_beam_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], + beam: int = 4, +): + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The RNN-T model. + encoder_out: + A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of + the encoder model. + streams: + A list of stream objects. + beam: + Number of active paths during the beam search. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert len(streams) == encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + batch_size = len(streams) + T = encoder_out.size(1) + + B = [stream.hyps for stream in streams] + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim) + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.stack( + [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0 + ) # (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim) + + # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor + # as index, so we use `to(torch.int64)` below. + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, decoder_out, project_input=False + ) + # logits is of shape (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) + + log_probs = logits.log_softmax(dim=-1) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + + vocab_size = log_probs.size(-1) + + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor( + shape=log_probs_shape, value=log_probs + ) + + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_ys = hyp.ys[:] + new_token = topk_token_indexes[k] + if new_token != blank_id: + new_ys.append(new_token) + + new_log_prob = topk_log_probs[k] + new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob) + B[i].add(new_hyp) + + for i in range(batch_size): + streams[i].hyps = B[i] + + +def fast_beam_search_one_best( + model: nn.Module, + streams: List[Stream], + encoder_out: torch.Tensor, + processed_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, +) -> None: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using modified beam search, and then + the shortest path within the lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + streams: + A list of stream objects. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + processed_lens: + A tensor of shape (N,) containing the number of processed frames + in `encoder_out` before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + """ + assert encoder_out.ndim == 3 + + context_size = model.decoder.context_size + vocab_size = model.decoder.vocab_size + + B, T, C = encoder_out.shape + assert B == len(streams) + + config = k2.RnntDecodingConfig( + vocab_size=vocab_size, + decoder_history_len=context_size, + beam=beam, + max_contexts=max_contexts, + max_states=max_states, + ) + individual_streams = [] + for i in range(B): + individual_streams.append(streams[i].rnnt_decoding_stream) + decoding_streams = k2.RnntDecodingStreams(individual_streams, config) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + # shape is a RaggedShape of shape (B, context) + # contexts is a Tensor of shape (shape.NumElements(), context_size) + shape, contexts = decoding_streams.get_contexts() + # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64 + contexts = contexts.to(torch.int64) + # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim) + decoder_out = model.decoder(contexts, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # current_encoder_out is of shape + # (shape.NumElements(), 1, joiner_dim) + # fmt: off + current_encoder_out = torch.index_select( + encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64) + ) + # fmt: on + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + logits = logits.squeeze(1).squeeze(1) + log_probs = logits.log_softmax(dim=-1) + decoding_streams.advance(log_probs) + + decoding_streams.terminate_and_flush_to_streams() + + lattice = decoding_streams.format_output(processed_lens.tolist()) + + best_path = one_best_decoding(lattice) + hyps = get_texts(best_path) + + for i in range(B): + streams[i].hyp = hyps[i] + + +def decode_one_chunk( + model: nn.Module, + streams: List[Stream], + params: AttributeDict, + decoding_graph: Optional[k2.Fsa] = None, +) -> List[int]: + """ + Args: + model: + The Transducer model. + streams: + A list of Stream objects. + params: + It is returned by :func:`get_params`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + + Returns: + A list of indexes indicating the finished streams. + """ + device = next(model.parameters()).device + + feature_list = [] + feature_len_list = [] + state_list = [] + num_processed_frames_list = [] + + for stream in streams: + # We should first get `stream.num_processed_frames` + # before calling `stream.get_feature_chunk()` + # since `stream.num_processed_frames` would be updated + num_processed_frames_list.append(stream.num_processed_frames) + feature = stream.get_feature_chunk() + feature_len = feature.size(0) + feature_list.append(feature) + feature_len_list.append(feature_len) + state_list.append(stream.states) + + features = pad_sequence( + feature_list, batch_first=True, padding_value=LOG_EPSILON + ).to(device) + feature_lens = torch.tensor(feature_len_list, device=device) + num_processed_frames = torch.tensor( + num_processed_frames_list, device=device + ) + + # Make sure it has at least 1 frame after subsampling, first-and-last-frame cutting, and right context cutting # noqa + tail_length = ( + 3 * params.subsampling_factor + params.right_context_length + 3 + ) + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPSILON, + ) + + # Stack states of all streams + states = stack_states(state_list) + + encoder_out, encoder_out_lens, states = model.encoder.infer( + x=features, + x_lens=feature_lens, + states=states, + num_processed_frames=num_processed_frames, + ) + + if params.decoding_method == "greedy_search": + greedy_search( + model=model, + streams=streams, + encoder_out=encoder_out, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=streams, + encoder_out=encoder_out, + beam=params.beam_size, + ) + elif params.decoding_method == "fast_beam_search": + # feature_len is needed to get partial results. + # The rnnt_decoding_stream for fast_beam_search. + fast_beam_search_one_best( + model=model, + streams=streams, + encoder_out=encoder_out, + processed_lens=(num_processed_frames >> 2) + encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + + # Update cached states of each stream + state_list = unstack_states(states) + for i, s in enumerate(state_list): + streams[i].states = s + + finished_streams = [i for i, stream in enumerate(streams) if stream.done] + return finished_streams + + +def create_streaming_feature_extractor() -> Fbank: + """Create a CPU streaming feature extractor. + + At present, we assume it returns a fbank feature extractor with + fixed options. In the future, we will support passing in the options + from outside. + + Returns: + Return a CPU streaming feature extractor. + """ + opts = FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + return Fbank(opts) + + +def decode_dataset( + cuts: CutSet, + model: nn.Module, + params: AttributeDict, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +): + """Decode dataset. + + Args: + cuts: + Lhotse Cutset containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The Transducer model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + device = next(model.parameters()).device + + log_interval = 300 + + fbank = create_streaming_feature_extractor() + + decode_results = [] + streams = [] + for num, cut in enumerate(cuts): + # Each utterance has a Stream. + stream = Stream( + params=params, + decoding_graph=decoding_graph, + device=device, + LOG_EPS=LOG_EPSILON, + ) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + # The trained model is using normalized samples + assert audio.max() <= 1, "Should be normalized to [-1, 1])" + + samples = torch.from_numpy(audio).squeeze(0) + feature = fbank(samples) + stream.set_feature(feature) + stream.set_ground_truth(cut.supervisions[0].text) + + streams.append(stream) + + while len(streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + while len(streams) > 0: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if params.decoding_method == "greedy_search": + key = "greedy_search" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ) + else: + key = f"beam_size_{params.beam_size}" + + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=sorted(results)) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / "streaming" / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + # for streaming + params.suffix += f"-streaming-chunk-length-{params.chunk_length}" + params.suffix += f"-left-context-length-{params.left_context_length}" + params.suffix += f"-right-context-length-{params.right_context_length}" + params.suffix += f"-memory-size-{params.memory_size}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{params.decoding_method}-beam-size-{params.beam_size}" + ) + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-streaming-decode") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + params.device = device + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.eval() + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_sets = ["test-clean", "test-other"] + test_cuts = [test_clean_cuts, test_other_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): + results_dict = decode_dataset( + cuts=test_cut, + model=model, + params=params, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + torch.manual_seed(20220410) + main() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py new file mode 100644 index 000000000..8cde6205b --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py @@ -0,0 +1,194 @@ +#!/usr/bin/env python3 +# +# Copyright 2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import torch +from emformer import ConvolutionModule, Emformer, stack_states, unstack_states + + +def test_convolution_module_forward(): + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 3 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + + utterance, right_context = conv_module(utterance, right_context) + assert utterance.shape == (U, B, D), utterance.shape + assert right_context.shape == (R, B, D), right_context.shape + + +def test_convolution_module_infer(): + from emformer import ConvolutionModule + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 1 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + cache = torch.randn(B, D, kernel_size - 1) + + utterance, right_context, new_cache = conv_module.infer( + utterance, right_context, cache + ) + assert utterance.shape == (U, B, D), utterance.shape + assert right_context.shape == (R, B, D), right_context.shape + assert new_cache.shape == (B, D, kernel_size - 1), new_cache.shape + + +def test_state_stack_unstack(): + num_features = 80 + chunk_length = 32 + encoder_dim = 512 + num_encoder_layers = 2 + kernel_size = 31 + left_context_length = 32 + right_context_length = 8 + memory_size = 32 + + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=encoder_dim, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + ) + + for batch_size in [1, 2]: + attn_caches = [ + [ + torch.zeros(memory_size, batch_size, encoder_dim), + torch.zeros(left_context_length // 4, batch_size, encoder_dim), + torch.zeros( + left_context_length // 4, + batch_size, + encoder_dim, + ), + ] + for _ in range(num_encoder_layers) + ] + conv_caches = [ + torch.zeros(batch_size, encoder_dim, kernel_size - 1) + for _ in range(num_encoder_layers) + ] + states = [attn_caches, conv_caches] + x = torch.randn(batch_size, 23, num_features) + x_lens = torch.full((batch_size,), 23) + num_processed_frames = torch.full((batch_size,), 0) + y, y_lens, states = model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + state_list = unstack_states(states) + states2 = stack_states(state_list) + + for ss, ss2 in zip(states[0], states2[0]): + for s, s2 in zip(ss, ss2): + assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}" + + for s, s2 in zip(states[1], states2[1]): + assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}" + + +def test_torchscript_consistency_infer(): + r"""Verify that scripting Emformer does not change the behavior of method `infer`.""" # noqa + num_features = 80 + chunk_length = 32 + encoder_dim = 512 + num_encoder_layers = 2 + kernel_size = 31 + left_context_length = 32 + right_context_length = 8 + memory_size = 32 + batch_size = 2 + + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=encoder_dim, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + ).eval() + attn_caches = [ + [ + torch.zeros(memory_size, batch_size, encoder_dim), + torch.zeros(left_context_length // 4, batch_size, encoder_dim), + torch.zeros( + left_context_length // 4, + batch_size, + encoder_dim, + ), + ] + for _ in range(num_encoder_layers) + ] + conv_caches = [ + torch.zeros(batch_size, encoder_dim, kernel_size - 1) + for _ in range(num_encoder_layers) + ] + states = [attn_caches, conv_caches] + x = torch.randn(batch_size, 23, num_features) + x_lens = torch.full((batch_size,), 23) + num_processed_frames = torch.full((batch_size,), 0) + y, y_lens, out_states = model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + sc_model = torch.jit.script(model).eval() + sc_y, sc_y_lens, sc_out_states = sc_model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + assert torch.allclose(y, sc_y) + + +if __name__ == "__main__": + test_convolution_module_forward() + test_convolution_module_infer() + test_state_stack_unstack() + test_torchscript_consistency_infer() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py new file mode 100755 index 000000000..106f3e511 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py @@ -0,0 +1,1136 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./conv_emformer_transducer_stateless/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --full-libri 1 \ + --max-duration 300 \ + --master-port 12321 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 + +# For mix precision training: +./conv_emformer_transducer_stateless/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --full-libri 1 \ + --max-duration 300 \ + --master-port 12321 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from emformer import Emformer +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, Eve +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--encoder-dim", + type=int, + default=512, + help="Attention dim for the Emformer", + ) + + parser.add_argument( + "--nhead", + type=int, + default=8, + help="Number of attention heads for the Emformer", + ) + + parser.add_argument( + "--dim-feedforward", + type=int, + default=2048, + help="Feed-forward dimension for the Emformer", + ) + + parser.add_argument( + "--num-encoder-layers", + type=int, + default=12, + help="Number of encoder layers for the Emformer", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=int, + default=31, + help="Kernel size for the convolution module.", + ) + + parser.add_argument( + "--left-context-length", + type=int, + default=32, + help="""Number of frames before subsampling for left context + in the Emformer.""", + ) + + parser.add_argument( + "--chunk-length", + type=int, + default=32, + help="""Number of frames before subsampling for each chunk + in the Emformer.""", + ) + + parser.add_argument( + "--right-context-length", + type=int, + default=8, + help="""Number of frames before subsampling for right context + in the Emformer.""", + ) + + parser.add_argument( + "--memory-size", + type=int, + default=0, + help="Number of entries in the memory for the Emformer", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--initial-lr", + type=float, + default=0.003, + help="""The initial learning rate. This value should not need to be + changed.""", + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate decreases. + We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" + "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=8000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=20, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for Emformer + "feature_dim": 80, + "subsampling_factor": 4, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Emformer( + num_features=params.feature_dim, + chunk_length=params.chunk_length, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + cnn_module_kernel=params.cnn_module_kernel, + left_context_length=params.left_context_length, + right_context_length=params.right_context_length, + memory_size=params.memory_size, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = ( + model.device + if isinstance(model, DDP) + else next(model.parameters()).device + ) + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = ( + (feature_lens // params.subsampling_factor).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}" + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + + if batch_idx > 0 and batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + 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)) + if params.full_libri is False: + params.valid_interval = 1600 + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank]) + + optimizer = Eve(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2 ** 22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + # warmup = 0.0 is so that the derivs for the pruned loss stay zero + # (i.e. are not remembered by the decaying-average in adam), because + # we want to avoid these params being subject to shrinkage in adam. + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=0.0, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + except RuntimeError as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py index ce8b04afd..7c936b257 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py @@ -557,7 +557,7 @@ class HypothesisList(object): return ", ".join(s) -def _get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape: +def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape: """Return a ragged shape with axes [utt][num_hyps]. Args: @@ -648,7 +648,7 @@ def modified_beam_search( finalized_B = B[batch_size:] + finalized_B B = B[:batch_size] - hyps_shape = _get_hyps_shape(B).to(device) + hyps_shape = get_hyps_shape(B).to(device) A = [list(b) for b in B] B = [HypothesisList() for _ in range(batch_size)]