From a380556b887e63716c70da96cf77d1647541ad92 Mon Sep 17 00:00:00 2001 From: Fangjun Kuang Date: Thu, 5 May 2022 21:04:57 +0800 Subject: [PATCH 1/4] Copy files for editing. --- .../asr_datamodule.py | 1 + .../beam_search.py | 1 + .../decode-giga.py | 644 ++++++++++ .../pruned_transducer_stateless5/decode.py | 626 +++++++++ .../pruned_transducer_stateless5/decoder.py | 1 + .../encoder_interface.py | 1 + .../gigaspeech.py | 1 + .../gigaspeech_scoring.py | 1 + .../pruned_transducer_stateless5/joiner.py | 1 + .../librispeech.py | 1 + .../ASR/pruned_transducer_stateless5/model.py | 1 + .../ASR/pruned_transducer_stateless5/optim.py | 1 + .../pruned_transducer_stateless5/scaling.py | 1 + .../ASR/pruned_transducer_stateless5/train.py | 1125 +++++++++++++++++ 14 files changed, 2406 insertions(+) create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless5/decode.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/decoder.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech_scoring.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/librispeech.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/model.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/optim.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless5/train.py diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py new file mode 120000 index 000000000..3ba9ada4f --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py new file mode 120000 index 000000000..e9bbcf2a9 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py new file mode 100755 index 000000000..a715a2a5c --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py @@ -0,0 +1,644 @@ +#!/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. +""" +Usage: +(1) greedy search +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 1500 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +import argparse +import logging +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 AsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from gigaspeech import GigaSpeech +from gigaspeech_scoring import asr_text_post_processing +from train import get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + 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. + 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_stateless3/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 + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=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 or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for computed nbest oracle WER + when the decoding method is fast_beam_search_nbest_oracle. + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding_method is fast_beam_search_nbest_oracle. + """, + ) + return parser + + +def post_processing( + results: List[Tuple[List[List[str]], List[List[str]]]], +) -> List[Tuple[List[List[str]], List[List[str]]]]: + new_results = [] + for ref, hyp in results: + new_ref = asr_text_post_processing(" ".join(ref)).split() + new_hyp = asr_text_post_processing(" ".join(hyp)).split() + new_results.append((new_ref, new_hyp)) + return new_results + + +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 or fast_beam_search_nbest_oracle. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = model.device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + ) + 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, + 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 + } + elif params.decoding_method == "fast_beam_search_nbest_oracle": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}_" + f"num_paths_{params.num_paths}_" + f"nbest_scale_{params.nbest_scale}" + ): 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[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = post_processing(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / "giga" / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.decoding_method == "fast_beam_search": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif params.decoding_method == "fast_beam_search_nbest_oracle": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + params.suffix += f"-num-paths-{params.num_paths}" + params.suffix += f"-nbest-scale-{params.nbest_scale}" + 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}" + + 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.unk_id() + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + 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 start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.to(device) + model.eval() + model.device = device + model.unk_id = params.unk_id + + # In beam_search.py, we are using model.decoder() and model.joiner(), + # so we have to switch to the branch for the GigaSpeech dataset. + model.decoder = model.decoder_giga + model.joiner = model.joiner_giga + + if params.decoding_method in ( + "fast_beam_search", + "fast_beam_search_nbest_oracle", + ): + 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}") + + asr_datamodule = AsrDataModule(args) + gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir) + + test_cuts = gigaspeech.test_cuts() + dev_cuts = gigaspeech.dev_cuts() + + test_dl = asr_datamodule.test_dataloaders(test_cuts) + dev_dl = asr_datamodule.test_dataloaders(dev_cuts) + + test_sets = ["test", "dev"] + test_sets_dl = [test_dl, dev_dl] + + for test_set, dl in zip(test_sets, test_sets_dl): + results_dict = decode_dataset( + dl=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/pruned_transducer_stateless5/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py new file mode 100755 index 000000000..9a6b5a117 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py @@ -0,0 +1,626 @@ +#!/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. +""" +Usage: +(1) greedy search +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 1500 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +import argparse +import logging +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 AsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from librispeech import LibriSpeech +from train import get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + 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. + 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_stateless3/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 + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=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 or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for computed nbest oracle WER + when the decoding method is fast_beam_search_nbest_oracle. + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding_method is fast_beam_search_nbest_oracle. + """, + ) + 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 or fast_beam_search_nbest_oracle. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = model.device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + ) + 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, + 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 + } + elif params.decoding_method == "fast_beam_search_nbest_oracle": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}_" + f"num_paths_{params.num_paths}_" + f"nbest_scale_{params.nbest_scale}" + ): 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() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.decoding_method == "fast_beam_search": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif params.decoding_method == "fast_beam_search_nbest_oracle": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + params.suffix += f"-num-paths-{params.num_paths}" + params.suffix += f"-nbest-scale-{params.nbest_scale}" + 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}" + + 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.unk_id() + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + 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 start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.to(device) + model.eval() + model.device = device + model.unk_id = params.unk_id + + if params.decoding_method in ( + "fast_beam_search", + "fast_beam_search_nbest_oracle", + ): + 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}") + + asr_datamodule = AsrDataModule(args) + librispeech = LibriSpeech(manifest_dir=args.manifest_dir) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts) + test_other_dl = asr_datamodule.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/pruned_transducer_stateless5/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decoder.py new file mode 120000 index 000000000..d8c760036 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decoder.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py new file mode 120000 index 000000000..10505621c --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/encoder_interface.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech.py new file mode 120000 index 000000000..5242c652a --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/gigaspeech.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech_scoring.py b/egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech_scoring.py new file mode 120000 index 000000000..bbe2e7769 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/gigaspeech_scoring.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/gigaspeech_scoring.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py new file mode 120000 index 000000000..e023db4a9 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/joiner.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/librispeech.py b/egs/librispeech/ASR/pruned_transducer_stateless5/librispeech.py new file mode 120000 index 000000000..b76723bf5 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/librispeech.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/librispeech.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/model.py b/egs/librispeech/ASR/pruned_transducer_stateless5/model.py new file mode 120000 index 000000000..78ed28971 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/model.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless5/optim.py new file mode 120000 index 000000000..6d41654f4 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/optim.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py new file mode 120000 index 000000000..5c79c7a94 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/scaling.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py new file mode 100755 index 000000000..4966ea57f --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py @@ -0,0 +1,1125 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang +# Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +cd egs/librispeech/ASR/ +./prepare.sh +./prepare_giga_speech.sh + +./pruned_transducer_stateless3/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --exp-dir pruned_transducer_stateless3/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless3/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --use_fp16 1 \ + --exp-dir pruned_transducer_stateless3/exp \ + --full-libri 1 \ + --max-duration 550 + +""" + + +import argparse +import logging +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AsrDataModule +from conformer import Conformer +from decoder import Decoder +from gigaspeech import GigaSpeech +from joiner import Joiner +from lhotse import CutSet, load_manifest +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from librispeech import LibriSpeech +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 +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 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( + "--full-libri", + type=str2bool, + default=True, + help="When enabled, use 960h LibriSpeech. " + "Otherwise, use 100h subset.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=0, + help="""Resume training from from this epoch. + If it is positive, it will load checkpoint from + transducer_stateless3/exp/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_stateless3/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=4, + 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( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--giga-prob", + type=float, + default=0.5, + help="The probability to select a batch from the GigaSpeech dataset", + ) + + 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 conformer + "feature_dim": 80, + "subsampling_factor": 4, + "encoder_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + # 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 = Conformer( + num_features=params.feature_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + decoder_giga = get_decoder_model(params) + joiner_giga = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + decoder_giga=decoder_giga, + joiner_giga=joiner_giga, + 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, + 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 positive, 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. + 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 > 0: + 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, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: nn.Module, + 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. + 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, + 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 is_libri(c: Cut) -> bool: + """Return True if this cut is from the LibriSpeech dataset. + + Note: + During data preparation, we set the custom field in + the supervision segment of GigaSpeech to dict(origin='giga') + See ../local/preprocess_gigaspeech.py. + """ + return c.supervisions[0].custom is None + + +def compute_loss( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device + 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) + + libri = is_libri(supervisions["cut"][0]) + + 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, + libri=libri, + 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: nn.Module, + 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: nn.Module, + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + giga_train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + rng: random.Random, + scaler: GradScaler, + 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. + giga_train_dl: + Dataloader for the GigaSpeech training dataset. + valid_dl: + Dataloader for the validation dataset. + rng: + For selecting which dataset to use. + scaler: + The scaler used for mix precision 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() + + libri_tot_loss = MetricsTracker() + giga_tot_loss = MetricsTracker() + tot_loss = MetricsTracker() + + # index 0: for LibriSpeech + # index 1: for GigaSpeech + # This sets the probabilities for choosing which datasets + dl_weights = [1 - params.giga_prob, params.giga_prob] + + iter_libri = iter(train_dl) + iter_giga = iter(giga_train_dl) + + batch_idx = 0 + + while True: + idx = rng.choices((0, 1), weights=dl_weights, k=1)[0] + dl = iter_libri if idx == 0 else iter_giga + + try: + batch = next(dl) + except StopIteration: + name = "libri" if idx == 0 else "giga" + logging.info(f"{name} reaches end of dataloader") + break + + batch_idx += 1 + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + libri = is_libri(batch["supervisions"]["cut"][0]) + + 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 + + if libri: + libri_tot_loss = ( + libri_tot_loss * (1 - 1 / params.reset_interval) + ) + loss_info + prefix = "libri" # for logging only + else: + giga_tot_loss = ( + giga_tot_loss * (1 - 1 / params.reset_interval) + ) + loss_info + prefix = "giga" + + # 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 ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, {prefix}_loss[{loss_info}], " + f"tot_loss[{tot_loss}], " + f"libri_tot_loss[{libri_tot_loss}], " + f"giga_tot_loss[{giga_tot_loss}], " + f"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, + f"train/current_{prefix}_", + params.batch_idx_train, + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + libri_tot_loss.write_summary( + tb_writer, "train/libri_tot_", params.batch_idx_train + ) + giga_tot_loss.write_summary( + tb_writer, "train/giga_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 filter_short_and_long_utterances(cuts: CutSet) -> CutSet: + 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 + + cuts = cuts.filter(remove_short_and_long_utt) + + return cuts + + +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) + rng = random.Random(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}") + + checkpoints = load_checkpoint_if_available(params=params, model=model) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + model.device = device + + 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 = LibriSpeech(manifest_dir=args.manifest_dir) + + 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() + + train_cuts = filter_short_and_long_utterances(train_cuts) + + gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir) + # XL 10k hours + # L 2.5k hours + # M 1k hours + # S 250 hours + # XS 10 hours + # DEV 12 hours + # Test 40 hours + if params.full_libri: + logging.info("Using the XL subset of GigaSpeech (10k hours)") + train_giga_cuts = gigaspeech.train_XL_cuts() + else: + logging.info("Using the S subset of GigaSpeech (250 hours)") + train_giga_cuts = gigaspeech.train_S_cuts() + + train_giga_cuts = filter_short_and_long_utterances(train_giga_cuts) + + if args.enable_musan: + cuts_musan = load_manifest( + Path(args.manifest_dir) / "cuts_musan.json.gz" + ) + else: + cuts_musan = None + + asr_datamodule = AsrDataModule(args) + + train_dl = asr_datamodule.train_dataloaders( + train_cuts, + dynamic_bucketing=False, + on_the_fly_feats=False, + cuts_musan=cuts_musan, + ) + + giga_train_dl = asr_datamodule.train_dataloaders( + train_giga_cuts, + dynamic_bucketing=True, + on_the_fly_feats=False, + cuts_musan=cuts_musan, + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = asr_datamodule.valid_dataloaders(valid_cuts) + + # It's time consuming to include `giga_train_dl` here + # for dl in [train_dl, giga_train_dl]: + for dl in [train_dl]: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=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): + scheduler.step_epoch(epoch) + fix_random_seed(params.seed + epoch) + train_dl.sampler.set_epoch(epoch) + + 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, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + giga_train_dl=giga_train_dl, + valid_dl=valid_dl, + rng=rng, + 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, + 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: nn.Module, + 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 0 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() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + assert 0 <= args.giga_prob < 1, args.giga_prob + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() From eac839478bfe091c2d41a3008900ea1e47cff542 Mon Sep 17 00:00:00 2001 From: Fangjun Kuang Date: Thu, 5 May 2022 21:05:35 +0800 Subject: [PATCH 2/4] Copy files. --- .../pruned_transducer_stateless5/conformer.py | 1073 +++++++++++++++++ .../pruned_transducer_stateless5/sampling.py | 293 +++++ 2 files changed, 1366 insertions(+) create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py new file mode 100644 index 000000000..548d8e275 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py @@ -0,0 +1,1073 @@ +#!/usr/bin/env python3 +# Copyright (c) 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +import warnings +from typing import Optional, Tuple, List +from sampling import create_knowledge_base, KnowledgeBaseLookup + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv1d, + ScaledConv2d, + ScaledLinear, +) +from torch import Tensor, nn + +from icefall.utils import make_pad_mask + + +class Conformer(EncoderInterface): + """ + Args: + num_features (int): Number of input features + subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers) + d_model (int): attention dimension, also the output dimension + nhead (int): number of head + dim_feedforward (int): feedforward dimention + num_encoder_layers (int): number of encoder layers + dropout (float): dropout rate + layer_dropout (float): layer-dropout rate. + cnn_module_kernel (int): Kernel size of convolution module + knowledge_M: softmax size in knowledge base + knowledge_N: number of softmaxes in knowledge base + knowledge_D: feature dimension in knowledge base + knowledge_K: number of samples to use each time in knowledge base + knowledge_share: number of successive layers that share the same knowledge + base + """ + def __init__( + self, + num_features: 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 = 31, + knowledge_M: int = 256, + knowledge_N: int = 2, + knowledge_D: int = 256, + knowledge_K: int = 16, + knowledge_share: int = 4, + ) -> None: + super(Conformer, self).__init__() + + self.num_features = num_features + self.subsampling_factor = subsampling_factor + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + + + num_knowledge_bases = (num_encoder_layers + knowledge_share - 1) // knowledge_share # round up + self.knowledge_base = nn.ParameterList([ create_knowledge_base(knowledge_M, knowledge_N, + knowledge_D) + for _ in range(num_knowledge_bases) ]) + + # 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_pos = RelPositionalEncoding(d_model, dropout) + + + encoder_layers = [ ConformerEncoderLayer( + self.knowledge_base[n // knowledge_share], + d_model, + nhead, + dim_feedforward, + dropout, + layer_dropout, + cnn_module_kernel, + knowledge_M, + knowledge_N, + knowledge_D, + knowledge_K + ) for n in range(num_encoder_layers) ] + self.encoder = ConformerEncoder(encoder_layers) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + 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: + Return a tuple containing 2 tensors: + - embeddings: its shape is (batch_size, output_seq_len, d_model) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + x = self.encoder_embed(x) + x, pos_emb = self.encoder_pos(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # Caution: We assume the subsampling factor is 4! + lengths = ((x_lens - 1) // 2 - 1) // 2 + assert x.size(0) == lengths.max().item() + mask = make_pad_mask(lengths) + + x = self.encoder( + x, pos_emb, src_key_padding_mask=mask, warmup=warmup + ) # (T, N, C) + + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return x, lengths + + +class ConformerEncoderLayer(nn.Module): + """ + ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks. + See: "Conformer: Convolution-augmented Transformer for Speech Recognition" + + Args: + knowledge_base: shared knowledge base parameter matrix, to be passed to constructors + of lookup modules + d_model: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + dim_feedforward: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + knowledge_M, knowledge_N, knowledge_D, knowledge_K: parameters for knowledge-base, + see docs for KnowlegeBaseLookup. + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + knowledge_base: nn.Parameter, + d_model: int, + nhead: int, + dim_feedforward: int = 2048, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + knowledge_M: int = 256, + knowledge_N: int = 2, + knowledge_D: int = 256, + knowledge_K: int = 16, + ) -> None: + super(ConformerEncoderLayer, self).__init__() + + self.layer_dropout = layer_dropout + + self.d_model = d_model + + self.self_attn = RelPositionMultiheadAttention( + d_model, nhead, dropout=0.0 + ) + + 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.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.conv_module = ConvolutionModule(d_model, cnn_module_kernel) + + self.lookup = KnowledgeBaseLookup(knowledge_M, knowledge_N, + knowledge_D, knowledge_K, + d_model, + knowledge_base) + + 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) + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + src_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + ) -> Tensor: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + pos_emb: Positional embedding tensor (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + warmup: controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, N is the batch size, E is the feature number + """ + 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)) + + # multi-headed self-attention module + src_att = self.self_attn( + src, + src, + src, + pos_emb=pos_emb, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + )[0] + src = src + self.dropout(src_att) + + # convolution module + src = src + self.dropout(self.conv_module(src)) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + # knowledge-base lookup + src = src + self.dropout(self.lookup(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + return src + + +class ConformerEncoder(nn.Module): + r"""ConformerEncoder is a stack of N encoder layers + + Args: + encoder_layers: the list of ConformerEncoderLayer modules. + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = conformer_encoder(src, pos_emb) + """ + + def __init__(self, encoder_layers: List[nn.Module]) -> None: + super().__init__() + self.layers = nn.ModuleList(encoder_layers) + self.num_layers = len(encoder_layers) + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required). + pos_emb: Positional embedding tensor (required). + mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + """ + output = src + + for i, mod in enumerate(self.layers): + output = mod( + output, + pos_emb, + src_mask=mask, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + ) + + return output + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + + """ + + def __init__( + self, d_model: int, dropout_rate: float, max_len: int = 5000 + ) -> None: + """Construct an PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.dropout = torch.nn.Dropout(p=dropout_rate) + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x: Tensor) -> None: + """Reset the positional encodings.""" + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(1) >= x.size(1) * 2 - 1: + # Note: TorchScript doesn't implement operator== for torch.Device + if self.pe.dtype != x.dtype or str(self.pe.device) != str( + x.device + ): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i Tuple[Tensor, Tensor]: + """Add positional encoding. + + Args: + x (torch.Tensor): Input tensor (batch, time, `*`). + + Returns: + torch.Tensor: Encoded tensor (batch, time, `*`). + torch.Tensor: Encoded tensor (batch, 2*time-1, `*`). + + """ + self.extend_pe(x) + pos_emb = self.pe[ + :, + self.pe.size(1) // 2 + - x.size(1) + + 1 : self.pe.size(1) // 2 # noqa E203 + + x.size(1), + ] + return self.dropout(x), self.dropout(pos_emb) + + +class RelPositionMultiheadAttention(nn.Module): + r"""Multi-Head Attention layer with relative position encoding + + See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + + Args: + embed_dim: total dimension of the model. + num_heads: parallel attention heads. + dropout: a Dropout layer on attn_output_weights. Default: 0.0. + + Examples:: + + >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads) + >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + num_heads: int, + dropout: float = 0.0, + ) -> None: + super(RelPositionMultiheadAttention, self).__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + + self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True) + self.out_proj = ScaledLinear( + embed_dim, embed_dim, bias=True, initial_scale=0.25 + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False) + # these two learnable bias are used in matrix c and matrix d + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach()) + self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach()) + self._reset_parameters() + + def _pos_bias_u(self): + return self.pos_bias_u * self.pos_bias_u_scale.exp() + + def _pos_bias_v(self): + return self.pos_bias_v * self.pos_bias_v_scale.exp() + + def _reset_parameters(self) -> None: + nn.init.normal_(self.pos_bias_u, std=0.01) + nn.init.normal_(self.pos_bias_v, std=0.01) + + def forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. When given a binary mask and a value is True, + the corresponding value on the attention layer will be ignored. When given + a byte mask and a value is non-zero, the corresponding value on the attention + layer will be ignored + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + - Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the position + with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + - Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + return self.multi_head_attention_forward( + query, + key, + value, + pos_emb, + self.embed_dim, + self.num_heads, + self.in_proj.get_weight(), + self.in_proj.get_bias(), + self.dropout, + self.out_proj.get_weight(), + self.out_proj.get_bias(), + training=self.training, + key_padding_mask=key_padding_mask, + need_weights=need_weights, + attn_mask=attn_mask, + ) + + def rel_shift(self, x: Tensor) -> Tensor: + """Compute relative positional encoding. + + Args: + x: Input tensor (batch, head, time1, 2*time1-1). + time1 means the length of query vector. + + Returns: + Tensor: tensor of shape (batch, head, time1, time2) + (note: time2 has the same value as time1, but it is for + the key, while time1 is for the query). + """ + (batch_size, num_heads, time1, n) = x.shape + assert n == 2 * time1 - 1 + # Note: TorchScript requires explicit arg for stride() + batch_stride = x.stride(0) + head_stride = x.stride(1) + time1_stride = x.stride(2) + n_stride = x.stride(3) + return x.as_strided( + (batch_size, num_heads, time1, time1), + (batch_stride, head_stride, time1_stride - n_stride, n_stride), + storage_offset=n_stride * (time1 - 1), + ) + + def multi_head_attention_forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + embed_dim_to_check: int, + num_heads: int, + in_proj_weight: Tensor, + in_proj_bias: Tensor, + dropout_p: float, + out_proj_weight: Tensor, + out_proj_bias: Tensor, + training: bool = True, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + embed_dim_to_check: total dimension of the model. + num_heads: parallel attention heads. + in_proj_weight, in_proj_bias: input projection weight and bias. + dropout_p: probability of an element to be zeroed. + out_proj_weight, out_proj_bias: the output projection weight and bias. + training: apply dropout if is ``True``. + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. This is an binary mask. When the value is True, + the corresponding value on the attention layer will be filled with -inf. + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence + length, N is the batch size, E is the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions + will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == embed_dim_to_check + assert key.size(0) == value.size(0) and key.size(1) == value.size(1) + + head_dim = embed_dim // num_heads + assert ( + head_dim * num_heads == embed_dim + ), "embed_dim must be divisible by num_heads" + + scaling = float(head_dim) ** -0.5 + + if torch.equal(query, key) and torch.equal(key, value): + # self-attention + q, k, v = nn.functional.linear( + query, in_proj_weight, in_proj_bias + ).chunk(3, dim=-1) + + elif torch.equal(key, value): + # encoder-decoder attention + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1) + + else: + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = embed_dim * 2 + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + k = nn.functional.linear(key, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim * 2 + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + v = nn.functional.linear(value, _w, _b) + + if attn_mask is not None: + assert ( + attn_mask.dtype == torch.float32 + or attn_mask.dtype == torch.float64 + or attn_mask.dtype == torch.float16 + or attn_mask.dtype == torch.uint8 + or attn_mask.dtype == torch.bool + ), "Only float, byte, and bool types are supported for attn_mask, not {}".format( + attn_mask.dtype + ) + if attn_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for attn_mask is deprecated. Use bool tensor instead." + ) + attn_mask = attn_mask.to(torch.bool) + + if attn_mask.dim() == 2: + attn_mask = attn_mask.unsqueeze(0) + if list(attn_mask.size()) != [1, query.size(0), key.size(0)]: + raise RuntimeError( + "The size of the 2D attn_mask is not correct." + ) + elif attn_mask.dim() == 3: + if list(attn_mask.size()) != [ + bsz * num_heads, + query.size(0), + key.size(0), + ]: + raise RuntimeError( + "The size of the 3D attn_mask is not correct." + ) + else: + raise RuntimeError( + "attn_mask's dimension {} is not supported".format( + attn_mask.dim() + ) + ) + # attn_mask's dim is 3 now. + + # convert ByteTensor key_padding_mask to bool + if ( + key_padding_mask is not None + and key_padding_mask.dtype == torch.uint8 + ): + warnings.warn( + "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead." + ) + key_padding_mask = key_padding_mask.to(torch.bool) + + q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim) + k = k.contiguous().view(-1, bsz, num_heads, head_dim) + v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1) + + src_len = k.size(0) + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz, "{} == {}".format( + key_padding_mask.size(0), bsz + ) + assert key_padding_mask.size(1) == src_len, "{} == {}".format( + key_padding_mask.size(1), src_len + ) + + q = q.transpose(0, 1) # (batch, time1, head, d_k) + + pos_emb_bsz = pos_emb.size(0) + assert pos_emb_bsz in (1, bsz) # actually it is 1 + p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim) + p = p.transpose(1, 2) # (batch, head, 2*time1-1, d_k) + + q_with_bias_u = (q + self._pos_bias_u()).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + q_with_bias_v = (q + self._pos_bias_v()).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + # compute attention score + # first compute matrix a and matrix c + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + k = k.permute(1, 2, 3, 0) # (batch, head, d_k, time2) + matrix_ac = torch.matmul( + q_with_bias_u, k + ) # (batch, head, time1, time2) + + # compute matrix b and matrix d + matrix_bd = torch.matmul( + q_with_bias_v, p.transpose(-2, -1) + ) # (batch, head, time1, 2*time1-1) + matrix_bd = self.rel_shift(matrix_bd) + + attn_output_weights = ( + matrix_ac + matrix_bd + ) # (batch, head, time1, time2) + + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, -1 + ) + + assert list(attn_output_weights.size()) == [ + bsz * num_heads, + tgt_len, + src_len, + ] + + if attn_mask is not None: + if attn_mask.dtype == torch.bool: + attn_output_weights.masked_fill_(attn_mask, float("-inf")) + else: + attn_output_weights += attn_mask + + if key_padding_mask is not None: + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + attn_output_weights = attn_output_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + float("-inf"), + ) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, src_len + ) + + attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1) + attn_output_weights = nn.functional.dropout( + attn_output_weights, p=dropout_p, training=training + ) + + attn_output = torch.bmm(attn_output_weights, v) + assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim] + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(tgt_len, bsz, embed_dim) + ) + attn_output = nn.functional.linear( + attn_output, out_proj_weight, out_proj_bias + ) + + if need_weights: + # average attention weights over heads + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + return attn_output, attn_output_weights.sum(dim=1) / num_heads + else: + return attn_output, None + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Conformer model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__( + self, channels: int, kernel_size: int, bias: bool = True + ) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + self.pointwise_conv1 = 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 + ) + + self.depthwise_conv = ScaledConv1d( + channels, + channels, + kernel_size, + stride=1, + padding=(kernel_size - 1) // 2, + 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 forward(self, x: Tensor) -> Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + return x.permute(2, 0, 1) + + +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 + + +if __name__ == "__main__": + feature_dim = 50 + c = Conformer(num_features=feature_dim, d_model=128, nhead=4) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + warmup=0.5, + ) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py b/egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py new file mode 100644 index 000000000..26f0d26b0 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py @@ -0,0 +1,293 @@ +#!/usr/bin/env python3 + +# This was copied from /ceph-dan/torch-sampling/torch_sampling/sampling_ref.py, +# its git history is there. + +import timeit +import torch +from torch import Tensor +from torch import nn +from torch.cuda.amp import GradScaler, custom_fwd, custom_bwd +from typing import Tuple, Optional +from scaling import ScaledLinear +import random +from torch_scheduled_sampling import sample_combined + +# The main exports of this file are the module KnowledgeBaseLookup and the +# function create_knowledge_base. + + + + + + +def create_knowledge_base(M: int, N: int, D: int) -> nn.Parameter: + std = 0.1 + a = (3 ** 0.5) * std # this sqrt(3) thing is intended to get variance of + # 0.1 from uniform distribution + ans = nn.Parameter(torch.ones(M ** N, D)) + nn.init.uniform_(ans, -a, a) + return ans + +def join_indexes(indexes: Tensor, M: int) -> Tensor: + """ + Combines N-tuples of indexes into single indexes that can be used for + lookup in the knowledge base. Args: + indexes: tensor of torch.int64 of shape (*, K, N), with elements in + {0..M-1} + M: the size of the original softmaxes, is upper bound on elements + in indexes + Returns: + joined_indexes: of shape (*, K), joined_indexes[...,k] equals + joined_indexes[...,0,k] + joined_indexes[...,1,k]*(M**1) ... + joined_indexes[...,1,k]*(M**(N-1))] + """ + N = indexes.shape[-1] + n_powers = M ** torch.arange(N, device=indexes.device) # [ 1, M, ..., M**(N-1) ] + return (indexes * n_powers).sum(dim=-1) + + +# Note, we don't use this, we +def weighted_matrix_lookup(weights: Tensor, + indexes: Tensor, + knowledge_base: Tensor) -> Tensor: + """ + Weighted combination of specified rows of a matrix. + weights: Tensor of shape (*, K), can contain any value but probably in [0..1]. + indexes: Tensor of shape (*, K), with elements in [0..C-1] + knowledge_base: Tensor of shape (C-1, D), whose rows we'll be looking up + Returns: + tensor of shape (*, D), containing weighted sums of rows of + `knowledge_base` + """ + if True: + return WeightedMatrixLookupFunction.apply(weights, indexes, knowledge_base) + else: + # simpler but less memory-efficient implementation + lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten()) + D = knowledge_base.shape[-1] + weights = weights.unsqueeze(-2) # (*, 1, K) + lookup = lookup.reshape(*indexes.shape, D) # (*, K, D) + ans = torch.matmul(weights, lookup) # ans: (*, 1, D) + ans = ans.squeeze(-2) + assert list(ans.shape) == list(weights.shape[:-2]) + [D] + return ans + + +class WeightedMatrixLookupFunction(torch.autograd.Function): + @staticmethod + @custom_fwd + def forward(ctx, weights: Tensor, indexes: Tensor, knowledge_base: Tensor) -> Tensor: + """ + Weighted combination of specified rows of a matrix. + weights: Tensor of shape (*, K), can contain any value but probably in [0..1]. + indexes: Tensor of shape (*, K), with elements in [0..C-1] + knowledge_base: Tensor of shape (C, D), whose rows we'll be looking up + Returns: + tensor of shape (*, D), containing weighted sums of rows of + `knowledge_base` + """ + if random.random() < 0.001: + print("dtype[1] = ", weights.dtype) + ctx.save_for_backward(weights.detach(), indexes.detach(), + knowledge_base.detach()) + with torch.no_grad(): + lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten()) + D = knowledge_base.shape[-1] + weights = weights.unsqueeze(-2) # (*, 1, K) + lookup = lookup.reshape(*indexes.shape, D) # (*, K, D) + ans = torch.matmul(weights, lookup) # ans: (*, 1, D) + ans = ans.squeeze(-2) #(*, D) + return ans + + @staticmethod + @custom_bwd + def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, Tensor]: + # ans_grad: (*, D) + weights, indexes, knowledge_base = ctx.saved_tensors + knowledge_base.requires_grad = True + dtype = ans_grad.dtype + ans_grad = ans_grad.to(weights.dtype) + assert weights.requires_grad == False + D = knowledge_base.shape[-1] + with torch.enable_grad(): + # we'll use torch's autograd to differentiate this operation, which + # is nontrivial [and anyway we need `lookup` to compute weight grad. + # We don't save `lookup` because it's large, that is the reason + # we override Torch autograd. + lookup = torch.index_select(knowledge_base, dim=0, index=indexes.flatten()) + lookup = lookup.reshape(*indexes.shape, D) # (*, K, D) + weights = weights.unsqueeze(-1) # (*, K, 1) + # forward pass: was: + ## ans = torch.matmul(weights, lookup) + ## ans: (*, 1, D) + ## ans = ans.squeeze(-2) # ans, ans_grad: (*, D) + weights_grad = torch.matmul(lookup, # (*, K, D) + ans_grad.unsqueeze(-1)) # (*, D, 1) + weights_grad = weights_grad.squeeze(-1) # (*, K, 1) -> (*, K) + lookup_grad = weights * ans_grad.unsqueeze(-2) # (*, K, 1) * (*, 1, D) = (*, K, D) + lookup.backward(gradient=lookup_grad) + return weights_grad.to(dtype), None, knowledge_base.grad.to(dtype) + + +class KnowledgeBaseLookup(nn.Module): + """ + Create knowledge-base lookup module. (The knowledge-base parameter, which is + large, is shared between these modules). + Args: + M: int, softmax size, e.g. in [32..128] + N: int, number of softmaxes, in [2..3] + D: int, embedding dimension in knowledge base, e.g. 256 + K: number of samples (affects speed/accuracy tradeoff), e.g. 16. + embedding_dim: the dimension to project from and to, e.g. the + d_model of the conformer. + """ + def __init__(self, M: int, N: int, D: int, + K: int, embedding_dim: int, + knowledge_base: nn.Parameter): + super(KnowledgeBaseLookup, self).__init__() + self.knowledge_base = knowledge_base # shared! + self.in_proj = ScaledLinear(embedding_dim, M * N, + initial_scale=1.0) + # initial_scale = 4.0 because the knowlege_base activations are + # quite small -- if we use our optimizer they'll have stddev <= 0.1. + self.out_proj = ScaledLinear(D, embedding_dim, + initial_scale = 4.0) + self.M = M + self.N = N + self.K = K + + def forward(self, x: Tensor) -> Tensor: + """ + Forward function that does knowledge-base lookup. + Args: + x: input, of shape (*, E) where E is embedding_dim + as passed to constructor + y: output of knowledge-base lookup, of shape (*, E) + + # TODO: later we can try multiplying by a projection of x or something like that. + """ + assert torch.all(x - x == 0) + x = self.in_proj(x) # now (*, M*N) + assert torch.all(x - x == 0) + x = x.reshape(*x.shape[:-1], self.N, self.M) # now (*, N, M) + x = x.log_softmax(dim=-1) # now normalized logprobs, dim= (*, N, M) + assert torch.all(x - x == 0) + if random.random() < 0.001: + entropy = (x * x.exp()).sum(dim=-1).mean() + print("Entropy = ", entropy) + # only need 'combined_indexes', call them 'indexes'. + _, indexes, weights = sample_combined(x, self.K, input_is_log=True) + x = weighted_matrix_lookup(weights, indexes, self.knowledge_base) # now (*, D) + x = self.out_proj(x) # now (*, self.embedding_dim) + return x + + +def _test_knowledge_base_lookup(): + K = 16 + N = 2 + M = 128 + D = 256 + E = 255 + + knowledge_base: nn.Parameter = create_knowledge_base(M, N, D) + m = KnowledgeBaseLookup(M, N, D, K, E, knowledge_base) + + B = 30 + T = 40 + x = torch.randn(B, T, E) + x.requires_grad = True + y = m(x) + assert y.shape == x.shape + y.sum().backward() # make sure backward doesn't crash.. + print("y = ", y) + print("x.grad = ", x.grad) + print("knowlege_base.grad norm = ", knowledge_base.grad.norm()) + + dtype = torch.float32 + device = torch.device('cuda') + train_pairs = [ (torch.randn(B, T, E, device=device, dtype=dtype), torch.randn(B, T, E, device=device, dtype=dtype)) for _ in range(10) ] + from optim import Eve + optimizer = Eve(m.parameters(), lr=0.005, eps=1.0e-04) + m = m.to(device).to(dtype) + + + start = timeit.default_timer() + +# Epoch 0, batch 0, loss 1.0109944343566895 +# Epoch 10, batch 0, loss 1.0146660804748535 +# Epoch 20, batch 0, loss 1.0119813680648804 +# Epoch 30, batch 0, loss 1.0105408430099487 +# Epoch 40, batch 0, loss 1.0077732801437378 +# Epoch 50, batch 0, loss 1.0050103664398193 +# Epoch 60, batch 0, loss 1.0033129453659058 +# Epoch 70, batch 0, loss 1.0014232397079468 +# Epoch 80, batch 0, loss 0.9977912306785583 +# Epoch 90, batch 0, loss 0.8274348974227905 +# Epoch 100, batch 0, loss 0.3368612825870514 +# Epoch 110, batch 0, loss 0.11323091387748718 +# Time taken: 17.591704960912466 + for epoch in range(150): + for n, (x,y) in enumerate(train_pairs): + y_out = m(x) + loss = ((y_out - y)**2).mean() * 100.0 + if n % 10 == 0 and epoch % 10 == 0: + print(f"Epoch {epoch}, batch {n}, loss {loss.item()}") + loss.backward() + optimizer.step() + optimizer.zero_grad() + + stop = timeit.default_timer() + print('Time taken: ', stop - start) + +def _test_knowledge_base_lookup_autocast(): + K = 16 + N = 2 + M = 128 + D = 256 + E = 255 + + knowledge_base: nn.Parameter = create_knowledge_base(M, N, D) + m = KnowledgeBaseLookup(M, N, D, K, E, knowledge_base) + + B = 30 + T = 40 + x = torch.randn(B, T, E) + x.requires_grad = True + y = m(x) + assert y.shape == x.shape + y.sum().backward() # make sure backward doesn't crash.. + print("y = ", y) + print("x.grad = ", x.grad) + print("knowlege_base.grad norm = ", knowledge_base.grad.norm()) + + device = torch.device('cuda') + train_pairs = [ (torch.randn(B, T, E, device=device), torch.randn(B, T, E, device=device)) for _ in range(10) ] + from optim import Eve + optimizer = Eve(m.parameters(), lr=0.005, eps=1.0e-04) + m = m.to(device) + + scaler = GradScaler(enabled=True) + + start = timeit.default_timer() + + + for epoch in range(150): + for n, (x,y) in enumerate(train_pairs): + y_out = m(x) + with torch.cuda.amp.autocast(enabled=True): + loss = ((y_out - y)**2).mean() * 100.0 + if n % 10 == 0 and epoch % 10 == 0: + print(f"Epoch {epoch}, batch {n}, loss {loss.item()}") + scaler.scale(loss).backward() + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + stop = timeit.default_timer() + print('Time taken: ', stop - start) + + + +if __name__ == '__main__': + _test_knowledge_base_lookup() + _test_knowledge_base_lookup_autocast() From e38c6aa7fa724aa99574dd28dc15e89e47218c8f Mon Sep 17 00:00:00 2001 From: Fangjun Kuang Date: Thu, 5 May 2022 21:15:13 +0800 Subject: [PATCH 3/4] Update encoder model parameters. --- .../decode-giga.py | 18 ++++---- .../pruned_transducer_stateless5/decode.py | 18 ++++---- .../test_model.py | 44 +++++++++++++++++++ .../ASR/pruned_transducer_stateless5/train.py | 20 +++++---- 4 files changed, 73 insertions(+), 27 deletions(-) create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py index a715a2a5c..9ae17fd11 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py @@ -18,36 +18,36 @@ """ Usage: (1) greedy search -./pruned_transducer_stateless3/decode-giga.py \ +./pruned_transducer_stateless5/decode-giga.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 100 \ --decoding-method greedy_search (2) beam search -./pruned_transducer_stateless3/decode-giga.py \ +./pruned_transducer_stateless5/decode-giga.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 100 \ --decoding-method beam_search \ --beam-size 4 (3) modified beam search -./pruned_transducer_stateless3/decode-giga.py \ +./pruned_transducer_stateless5/decode-giga.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 100 \ --decoding-method modified_beam_search \ --beam-size 4 (4) fast beam search -./pruned_transducer_stateless3/decode-giga.py \ +./pruned_transducer_stateless5/decode-giga.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 1500 \ --decoding-method fast_beam_search \ --beam 4 \ @@ -128,7 +128,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="pruned_transducer_stateless3/exp", + default="pruned_transducer_stateless5/exp", help="The experiment dir", ) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py index 9a6b5a117..865709833 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py @@ -18,36 +18,36 @@ """ Usage: (1) greedy search -./pruned_transducer_stateless3/decode.py \ +./pruned_transducer_stateless5/decode.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 100 \ --decoding-method greedy_search (2) beam search -./pruned_transducer_stateless3/decode.py \ +./pruned_transducer_stateless5/decode.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 100 \ --decoding-method beam_search \ --beam-size 4 (3) modified beam search -./pruned_transducer_stateless3/decode.py \ +./pruned_transducer_stateless5/decode.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 100 \ --decoding-method modified_beam_search \ --beam-size 4 (4) fast beam search -./pruned_transducer_stateless3/decode.py \ +./pruned_transducer_stateless5/decode.py \ --epoch 28 \ --avg 15 \ - --exp-dir ./pruned_transducer_stateless3/exp \ + --exp-dir ./pruned_transducer_stateless5/exp \ --max-duration 1500 \ --decoding-method fast_beam_search \ --beam 4 \ @@ -127,7 +127,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="pruned_transducer_stateless3/exp", + default="pruned_transducer_stateless5/exp", help="The experiment dir", ) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py new file mode 100755 index 000000000..1162eb379 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/test_model.py @@ -0,0 +1,44 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./pruned_transducer_stateless5/test_model.py +""" + +from train import get_params, get_transducer_model + + +def test_model(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + model = get_transducer_model(params) + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + +def main(): + test_model() + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py index 4966ea57f..dcedcfec6 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py @@ -25,22 +25,22 @@ cd egs/librispeech/ASR/ ./prepare.sh ./prepare_giga_speech.sh -./pruned_transducer_stateless3/train.py \ +./pruned_transducer_stateless5/train.py \ --world-size 4 \ --num-epochs 30 \ --start-epoch 0 \ - --exp-dir pruned_transducer_stateless3/exp \ + --exp-dir pruned_transducer_stateless5/exp \ --full-libri 1 \ --max-duration 300 # For mix precision training: -./pruned_transducer_stateless3/train.py \ +./pruned_transducer_stateless5/train.py \ --world-size 4 \ --num-epochs 30 \ --start-epoch 0 \ --use_fp16 1 \ - --exp-dir pruned_transducer_stateless3/exp \ + --exp-dir pruned_transducer_stateless5/exp \ --full-libri 1 \ --max-duration 550 @@ -154,7 +154,7 @@ def get_parser(): parser.add_argument( "--exp-dir", type=str, - default="pruned_transducer_stateless3/exp", + default="pruned_transducer_stateless5/exp", help="""The experiment dir. It specifies the directory where all training related files, e.g., checkpoints, log, etc, are saved @@ -346,10 +346,11 @@ def get_params() -> AttributeDict: # parameters for conformer "feature_dim": 80, "subsampling_factor": 4, - "encoder_dim": 512, - "nhead": 8, - "dim_feedforward": 2048, - "num_encoder_layers": 12, + "encoder_dim": 256, + "nhead": 4, + "dim_feedforward": 1024, + "num_encoder_layers": 18, + "knowledge_D": 512, # parameters for decoder "decoder_dim": 512, # parameters for joiner @@ -372,6 +373,7 @@ def get_encoder_model(params: AttributeDict) -> nn.Module: nhead=params.nhead, dim_feedforward=params.dim_feedforward, num_encoder_layers=params.num_encoder_layers, + knowledge_D=params.knowledge_D, ) return encoder From 6d809bad0bfe6ea1beb36eabf7c71e2c5f7716e9 Mon Sep 17 00:00:00 2001 From: Fangjun Kuang Date: Thu, 5 May 2022 22:06:37 +0800 Subject: [PATCH 4/4] Merge changes from master. --- .flake8 | 2 + egs/librispeech/ASR/README.md | 2 + .../ASR/pruned_transducer_stateless4/train.py | 2 +- .../decode-giga.py | 93 +++++++++++++------ .../pruned_transducer_stateless5/decode.py | 93 +++++++++++++------ .../pruned_transducer_stateless5/sampling.py | 6 +- .../ASR/pruned_transducer_stateless5/train.py | 81 +++++++++++++--- icefall/checkpoint.py | 4 +- 8 files changed, 208 insertions(+), 75 deletions(-) diff --git a/.flake8 b/.flake8 index 190387886..8b165135e 100644 --- a/.flake8 +++ b/.flake8 @@ -20,4 +20,6 @@ exclude = .git, **/data/**, icefall/shared/make_kn_lm.py, + egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py, + egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py, icefall/__init__.py diff --git a/egs/librispeech/ASR/README.md b/egs/librispeech/ASR/README.md index c053076a3..14dbfe95f 100644 --- a/egs/librispeech/ASR/README.md +++ b/egs/librispeech/ASR/README.md @@ -19,6 +19,8 @@ The following table lists the differences among them. | `pruned_transducer_stateless` | Conformer | Embedding + Conv1d | Using k2 pruned RNN-T loss | | `pruned_transducer_stateless2` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss | | `pruned_transducer_stateless3` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss + using GigaSpeech as extra training data | +| `pruned_transducer_stateless4` | Conformer(modified) | Embedding + Conv1d | Same as pruned_transducer_stateless2 but supports saving averaged model periodically.| +| `pruned_transducer_stateless5` | Conformer(modified) | Embedding + Conv1d | Same as pruned_transducer_stateless3 but with knowledge bank| The decoder in `transducer_stateless` is modified from the paper diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/train.py b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py index 147bcf658..de126a8e3 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless4/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py @@ -411,7 +411,7 @@ def get_transducer_model(params: AttributeDict) -> nn.Module: def load_checkpoint_if_available( params: AttributeDict, model: nn.Module, - model_avg: nn.Module = None, + model_avg: Optional[nn.Module] = None, optimizer: Optional[torch.optim.Optimizer] = None, scheduler: Optional[LRSchedulerType] = None, ) -> Optional[Dict[str, Any]]: diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py index 9ae17fd11..5d51ab478 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode-giga.py @@ -1,6 +1,8 @@ #!/usr/bin/env python3 # -# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang +# Zengwei Yao) +# # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -81,6 +83,7 @@ from train import get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, + average_checkpoints_with_averaged_model, find_checkpoints, load_checkpoint, ) @@ -88,6 +91,7 @@ from icefall.utils import ( AttributeDict, setup_logger, store_transcripts, + str2bool, write_error_stats, ) @@ -102,7 +106,7 @@ def get_parser(): type=int, default=28, help="""It specifies the checkpoint to use for decoding. - Note: Epoch counts from 0. + Note: Epoch counts from 1. You can specify --avg to use more checkpoints for model averaging.""", ) @@ -125,6 +129,17 @@ def get_parser(): "'--epoch' and '--iter'", ) + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + parser.add_argument( "--exp-dir", type=str, @@ -538,6 +553,9 @@ def main(): 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") @@ -560,34 +578,53 @@ def main(): logging.info("About to create model") model = get_transducer_model(params) - 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) + 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: - start = params.epoch - params.avg + 1 - filenames = [] - for i in range(start, params.epoch + 1): - if start >= 0: - filenames.append(f"{params.exp_dir}/epoch-{i}.pt") - logging.info(f"averaging {filenames}") + assert params.iter == 0 and 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(filenames, device=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() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py index 865709833..1cfdd57a3 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py @@ -1,6 +1,8 @@ #!/usr/bin/env python3 # -# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang +# Zengwei Yao) +# # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -80,6 +82,7 @@ from train import get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, + average_checkpoints_with_averaged_model, find_checkpoints, load_checkpoint, ) @@ -87,6 +90,7 @@ from icefall.utils import ( AttributeDict, setup_logger, store_transcripts, + str2bool, write_error_stats, ) @@ -101,7 +105,7 @@ def get_parser(): type=int, default=28, help="""It specifies the checkpoint to use for decoding. - Note: Epoch counts from 0. + Note: Epoch counts from 1. You can specify --avg to use more checkpoints for model averaging.""", ) @@ -124,6 +128,17 @@ def get_parser(): "'--epoch' and '--iter'", ) + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + parser.add_argument( "--exp-dir", type=str, @@ -525,6 +540,9 @@ def main(): 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") @@ -547,34 +565,53 @@ def main(): logging.info("About to create model") model = get_transducer_model(params) - 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) + 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: - start = params.epoch - params.avg + 1 - filenames = [] - for i in range(start, params.epoch + 1): - if start >= 0: - filenames.append(f"{params.exp_dir}/epoch-{i}.pt") - logging.info(f"averaging {filenames}") + assert params.iter == 0 and 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(filenames, device=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() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py b/egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py index 26f0d26b0..d53062c84 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/sampling.py @@ -86,8 +86,8 @@ class WeightedMatrixLookupFunction(torch.autograd.Function): tensor of shape (*, D), containing weighted sums of rows of `knowledge_base` """ - if random.random() < 0.001: - print("dtype[1] = ", weights.dtype) + # if random.random() < 0.001: + # print("dtype[1] = ", weights.dtype) ctx.save_for_backward(weights.detach(), indexes.detach(), knowledge_base.detach()) with torch.no_grad(): @@ -174,7 +174,7 @@ class KnowledgeBaseLookup(nn.Module): assert torch.all(x - x == 0) if random.random() < 0.001: entropy = (x * x.exp()).sum(dim=-1).mean() - print("Entropy = ", entropy) + # print("Entropy = ", entropy) # only need 'combined_indexes', call them 'indexes'. _, indexes, weights = sample_combined(x, self.K, input_is_log=True) x = weighted_matrix_lookup(weights, indexes, self.knowledge_base) # now (*, D) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py index dcedcfec6..62eeffb0e 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py @@ -1,7 +1,8 @@ #!/usr/bin/env python3 -# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang # Wei Kang -# Mingshuang Luo) +# Mingshuang Luo +# Zengwei Yao) # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -48,6 +49,7 @@ cd egs/librispeech/ASR/ import argparse +import copy import logging import random import warnings @@ -81,7 +83,10 @@ 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 +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 @@ -135,10 +140,10 @@ def get_parser(): parser.add_argument( "--start-epoch", type=int, - default=0, + default=1, help="""Resume training from from this epoch. If it is positive, it will load checkpoint from - transducer_stateless3/exp/epoch-{start_epoch-1}.pt + exp-dir/epoch-{start_epoch-1}.pt """, ) @@ -272,6 +277,19 @@ def get_parser(): """, ) + parser.add_argument( + "--average-period", + type=int, + default=1000, + 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, @@ -423,6 +441,7 @@ def get_transducer_model(params: AttributeDict) -> nn.Module: def load_checkpoint_if_available( params: AttributeDict, model: nn.Module, + model_avg: Optional[nn.Module] = None, optimizer: Optional[torch.optim.Optimizer] = None, scheduler: Optional[LRSchedulerType] = None, ) -> Optional[Dict[str, Any]]: @@ -430,7 +449,7 @@ def load_checkpoint_if_available( 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 positive, it will load the checkpoint from + 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 @@ -442,6 +461,8 @@ def load_checkpoint_if_available( 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: @@ -451,7 +472,7 @@ def load_checkpoint_if_available( """ if params.start_batch > 0: filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" - elif params.start_epoch > 0: + elif params.start_epoch > 1: filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" else: return None @@ -461,6 +482,7 @@ def load_checkpoint_if_available( saved_params = load_checkpoint( filename, model=model, + model_avg=model_avg, optimizer=optimizer, scheduler=scheduler, ) @@ -485,6 +507,7 @@ def load_checkpoint_if_available( def save_checkpoint( params: AttributeDict, model: nn.Module, + model_avg: Optional[nn.Module] = None, optimizer: Optional[torch.optim.Optimizer] = None, scheduler: Optional[LRSchedulerType] = None, sampler: Optional[CutSampler] = None, @@ -498,6 +521,8 @@ def save_checkpoint( 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: @@ -511,6 +536,7 @@ def save_checkpoint( save_checkpoint_impl( filename=filename, model=model, + model_avg=model_avg, params=params, optimizer=optimizer, scheduler=scheduler, @@ -667,6 +693,7 @@ def train_one_epoch( valid_dl: torch.utils.data.DataLoader, rng: random.Random, scaler: GradScaler, + model_avg: Optional[nn.Module] = None, tb_writer: Optional[SummaryWriter] = None, world_size: int = 1, rank: int = 0, @@ -696,6 +723,8 @@ def train_one_epoch( For selecting which dataset to use. 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: @@ -772,6 +801,17 @@ def train_one_epoch( 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 @@ -780,6 +820,7 @@ def train_one_epoch( out_dir=params.exp_dir, global_batch_idx=params.batch_idx_train, model=model, + model_avg=model_avg, params=params, optimizer=optimizer, scheduler=scheduler, @@ -915,7 +956,15 @@ def run(rank, world_size, args): num_param = sum([p.numel() for p in model.parameters()]) logging.info(f"Number of model parameters: {num_param}") - checkpoints = load_checkpoint_if_available(params=params, model=model) + 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) + + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) model.to(device) if world_size > 1: @@ -923,6 +972,10 @@ def run(rank, world_size, args): model = DDP(model, device_ids=[rank], find_unused_parameters=True) model.device = device + if rank == 0: + model_avg.to(device) + model_avg.device = device + optimizer = Eve(model.parameters(), lr=params.initial_lr) scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) @@ -1014,10 +1067,10 @@ def run(rank, world_size, args): logging.info("Loading grad scaler state dict") scaler.load_state_dict(checkpoints["grad_scaler"]) - for epoch in range(params.start_epoch, params.num_epochs): - scheduler.step_epoch(epoch) - fix_random_seed(params.seed + epoch) - train_dl.sampler.set_epoch(epoch) + 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) @@ -1027,6 +1080,7 @@ def run(rank, world_size, args): train_one_epoch( params=params, model=model, + model_avg=model_avg, optimizer=optimizer, scheduler=scheduler, sp=sp, @@ -1047,6 +1101,7 @@ def run(rank, world_size, args): save_checkpoint( params=params, model=model, + model_avg=model_avg, optimizer=optimizer, scheduler=scheduler, sampler=train_dl.sampler, @@ -1071,7 +1126,7 @@ def scan_pessimistic_batches_for_oom( from lhotse.dataset import find_pessimistic_batches logging.info( - "Sanity check -- see if any of the batches in epoch 0 would cause OOM." + "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(): diff --git a/icefall/checkpoint.py b/icefall/checkpoint.py index 5b562ccc8..2ca173663 100644 --- a/icefall/checkpoint.py +++ b/icefall/checkpoint.py @@ -346,7 +346,7 @@ def remove_checkpoints( for c in to_remove: os.remove(c) - +@torch.no_grad() def update_averaged_model( params: Dict[str, Tensor], model_cur: Union[nn.Module, DDP], @@ -442,7 +442,7 @@ def average_checkpoints_with_averaged_model( return avg - +@torch.no_grad() def average_state_dict( state_dict_1: Dict[str, Tensor], state_dict_2: Dict[str, Tensor],