From 374eacdd5ca02c92f6f1f6615b9b72c237172eb7 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Thu, 7 Apr 2022 21:32:59 +0800 Subject: [PATCH] First upload emformer_pruned_transducer_stateless recipe, refator emformer codes from torchaudio. --- .../asr_datamodule.py | 1 + .../beam_search.py | 1 + .../decode.py | 549 ++++++++++ .../decoder.py | 1 + .../emformer.py | 110 +- .../encoder_interface.py | 1 + .../joiner.py | 1 + .../model.py | 1 + .../noam.py | 104 ++ .../subsampling.py | 167 +-- .../test_emformer.py | 30 +- .../test_subsampling.py | 25 - .../train.py | 998 ++++++++++++++++++ 13 files changed, 1691 insertions(+), 298 deletions(-) create mode 120000 egs/librispeech/ASR/emformer_pruned_transducer_stateless/asr_datamodule.py create mode 120000 egs/librispeech/ASR/emformer_pruned_transducer_stateless/beam_search.py create mode 100755 egs/librispeech/ASR/emformer_pruned_transducer_stateless/decode.py create mode 120000 egs/librispeech/ASR/emformer_pruned_transducer_stateless/decoder.py create mode 120000 egs/librispeech/ASR/emformer_pruned_transducer_stateless/encoder_interface.py create mode 120000 egs/librispeech/ASR/emformer_pruned_transducer_stateless/joiner.py create mode 120000 egs/librispeech/ASR/emformer_pruned_transducer_stateless/model.py create mode 100644 egs/librispeech/ASR/emformer_pruned_transducer_stateless/noam.py mode change 100644 => 120000 egs/librispeech/ASR/emformer_pruned_transducer_stateless/subsampling.py delete mode 100644 egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_subsampling.py create mode 100755 egs/librispeech/ASR/emformer_pruned_transducer_stateless/train.py diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/asr_datamodule.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/asr_datamodule.py new file mode 120000 index 000000000..b4e5427e0 --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/beam_search.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/beam_search.py new file mode 120000 index 000000000..227d2247c --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/decode.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/decode.py new file mode 100755 index 000000000..c40b01dfa --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/decode.py @@ -0,0 +1,549 @@ +#!/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 +./transducer_emformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./transducer_emformer/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./transducer_emformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./transducer_emformer/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./transducer_emformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./transducer_emformer/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./transducer_emformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./transducer_emformer/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 LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + 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.", + ) + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--avg-last-n", + type=int, + default=0, + help="""If positive, --epoch and --avg are ignored and it + will use the last n checkpoints exp_dir/checkpoint-xxx.pt + where xxx is the number of processed batches while + saving that checkpoint. + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_emformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An interger indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = 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( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + ) + 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 + } + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if params.avg_last_n > 0: + filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n] + 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 + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/decoder.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/decoder.py new file mode 120000 index 000000000..0d5f10dc0 --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/decoder.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/emformer.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/emformer.py index 32498a2c1..edba2e0b3 100644 --- a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/emformer.py +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/emformer.py @@ -1,5 +1,6 @@ import math from typing import List, Optional, Tuple +import warnings import torch from torch import nn @@ -1051,7 +1052,6 @@ class EmformerEncoder(nn.Module): - output_lengths, with shape (B,), without containing the right_context at the end. """ - # assert x.size(0) == torch.max(lengths).item() right_context = self._gen_right_context(x) utterance = x[:x.size(0) - self.right_context_length] output_lengths = torch.clamp(lengths - self.right_context_length, min=0) @@ -1168,11 +1168,11 @@ class Emformer(EncoderInterface): ) if left_context_length != 0 and left_context_length % 4 != 0: raise NotImplementedError( - "left_context_length must be a mutiple of 4." + "left_context_length must be 0 or a mutiple of 4." ) if right_context_length != 0 and right_context_length % 4 != 0: raise NotImplementedError( - "right_context_length must be a mutiple of 4." + "right_context_length must be 0 or a mutiple of 4." ) # self.encoder_embed converts the input of shape (N, T, num_features) @@ -1185,8 +1185,6 @@ class Emformer(EncoderInterface): else: self.encoder_embed = Conv2dSubsampling(num_features, d_model) - self.encoder_pos = PositionalEncoding(d_model, dropout) - self.encoder = EmformerEncoder( chunk_length // 4, d_model, @@ -1228,19 +1226,20 @@ class Emformer(EncoderInterface): Returns: (Tensor, Tensor): - - output logits, with shape (B, U // 4, D). + - output logits, with shape (B, ((U - 1) // 2 - 1) // 2, D). - logits lengths, with shape (B,), without containing the right_context at the end. """ - # TODO: x.shape x = self.encoder_embed(x) - x = self.encoder_pos(x) x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) # Caution: We assume the subsampling factor is 4! - lengths = x_lens // 4 - assert x.size(0) == lengths.max().item() - output, output_lengths = self.encoder(x, lengths) # (T, N, C) + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + x_lens = ((x_lens - 1) // 2 - 1) // 2 + assert x.size(0) == x_lens.max().item() + + output, output_lengths = self.encoder(x, x_lens) # (T, N, C) logits = self.encoder_output_layer(output) logits = logits.permute(1, 0, 2) # (T, N, C) ->(N, T, C) @@ -1274,99 +1273,24 @@ class Emformer(EncoderInterface): (default: None) Returns: (Tensor, Tensor): - - output logits, with shape (B, U // 4, D). + - output logits, with shape (B, ((U - 1) // 2 - 1) // 2, D). - logits lengths, with shape (B,), without containing the right_context at the end. - updated states from current chunk's computation. """ x = self.encoder_embed(x) - x = self.encoder_pos(x) x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) # Caution: We assume the subsampling factor is 4! - lengths = x_lens // 4 - assert x.size(0) == lengths.max().item() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + x_lens = ((x_lens - 1) // 2 - 1) // 2 + assert x.size(0) == x_lens.max().item() + output, output_lengths, output_states = \ - self.encoder.infer(x, lengths, states) # (T, N, C) + self.encoder.infer(x, x_lens, states) # (T, N, C) logits = self.encoder_output_layer(output) logits = logits.permute(1, 0, 2) # (T, N, C) ->(N, T, C) return logits, output_lengths, output_states - - -class PositionalEncoding(nn.Module): - """This class implements the positional encoding - proposed in the following paper: - - - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf - - PE(pos, 2i) = sin(pos / (10000^(2i/d_modle)) - PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle)) - - Note:: - - 1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model))) - = exp(-1* 2i / d_model * log(100000)) - = exp(2i * -(log(10000) / d_model)) - """ - - def __init__(self, d_model: int, dropout: float = 0.1) -> None: - """ - Args: - d_model: - Embedding dimension. - dropout: - Dropout probability to be applied to the output of this module. - """ - super().__init__() - self.d_model = d_model - self.xscale = math.sqrt(self.d_model) - self.dropout = nn.Dropout(p=dropout) - # not doing: self.pe = None because of errors thrown by torchscript - self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32) - - def extend_pe(self, x: torch.Tensor) -> None: - """Extend the time t in the positional encoding if required. - - The shape of `self.pe` is (1, T1, d_model). The shape of the input x - is (N, T, d_model). If T > T1, then we change the shape of self.pe - to (N, T, d_model). Otherwise, nothing is done. - - Args: - x: - It is a tensor of shape (N, T, C). - Returns: - Return None. - """ - if self.pe is not None: - if self.pe.size(1) >= x.size(1): - self.pe = self.pe.to(dtype=x.dtype, device=x.device) - return - pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32) - position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) - div_term = torch.exp( - torch.arange(0, self.d_model, 2, dtype=torch.float32) - * -(math.log(10000.0) / self.d_model) - ) - pe[:, 0::2] = torch.sin(position * div_term) - pe[:, 1::2] = torch.cos(position * div_term) - pe = pe.unsqueeze(0) - # Now pe is of shape (1, T, d_model), where T is x.size(1) - self.pe = pe.to(device=x.device, dtype=x.dtype) - - def forward(self, x: torch.Tensor) -> torch.Tensor: - """ - Add positional encoding. - - Args: - x: - Its shape is (N, T, C) - - Returns: - Return a tensor of shape (N, T, C) - """ - self.extend_pe(x) - x = x * self.xscale + self.pe[:, : x.size(1), :] - return self.dropout(x) - diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/encoder_interface.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/encoder_interface.py new file mode 120000 index 000000000..aa5d0217a --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/encoder_interface.py @@ -0,0 +1 @@ +../transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/joiner.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/joiner.py new file mode 120000 index 000000000..81ad47c55 --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/joiner.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/model.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/model.py new file mode 120000 index 000000000..a61a0a23f --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/model.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/noam.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/noam.py new file mode 100644 index 000000000..e46bf35fb --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/noam.py @@ -0,0 +1,104 @@ +# Copyright 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch + + +class Noam(object): + """ + Implements Noam optimizer. + + Proposed in + "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf + + Modified from + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py # noqa + + Args: + params: + iterable of parameters to optimize or dicts defining parameter groups + model_size: + attention dimension of the transformer model + factor: + learning rate factor + warm_step: + warmup steps + """ + + def __init__( + self, + params, + model_size: int = 256, + factor: float = 10.0, + warm_step: int = 25000, + weight_decay=0, + ) -> None: + """Construct an Noam object.""" + self.optimizer = torch.optim.Adam( + params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay + ) + self._step = 0 + self.warmup = warm_step + self.factor = factor + self.model_size = model_size + self._rate = 0 + + @property + def param_groups(self): + """Return param_groups.""" + return self.optimizer.param_groups + + def step(self): + """Update parameters and rate.""" + self._step += 1 + rate = self.rate() + for p in self.optimizer.param_groups: + p["lr"] = rate + self._rate = rate + self.optimizer.step() + + def rate(self, step=None): + """Implement `lrate` above.""" + if step is None: + step = self._step + return ( + self.factor + * self.model_size ** (-0.5) + * min(step ** (-0.5), step * self.warmup ** (-1.5)) + ) + + def zero_grad(self): + """Reset gradient.""" + self.optimizer.zero_grad() + + def state_dict(self): + """Return state_dict.""" + return { + "_step": self._step, + "warmup": self.warmup, + "factor": self.factor, + "model_size": self.model_size, + "_rate": self._rate, + "optimizer": self.optimizer.state_dict(), + } + + def load_state_dict(self, state_dict): + """Load state_dict.""" + for key, value in state_dict.items(): + if key == "optimizer": + self.optimizer.load_state_dict(state_dict["optimizer"]) + else: + setattr(self, key, value) diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/subsampling.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/subsampling.py deleted file mode 100644 index 7d0ad44a6..000000000 --- a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/subsampling.py +++ /dev/null @@ -1,166 +0,0 @@ -# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) -# -# See ../../../../LICENSE for clarification regarding multiple authors -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -import torch -import torch.nn as nn - - -class 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 // 4. - - It is based on - https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa - """ - - def __init__(self, idim: int, odim: int) -> None: - """ - Args: - idim: - Input dim. The input shape is (N, T, idim). - Caution: It requires: T >= 4, idim >= 7 - odim: - Output dim. The output shape is (N, T // 4, odim) - """ - assert idim >= 7 - super().__init__() - self.conv_1 = nn.Sequential( - nn.Conv2d( - in_channels=1, out_channels=odim, kernel_size=3, stride=2 - ), - nn.ReLU(), - ) - self.conv_2 = nn.Sequential( - nn.Conv2d( - in_channels=odim, out_channels=odim, kernel_size=3, stride=2 - ), - nn.ReLU(), - ) - self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim) - - 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 // 4, 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 = nn.functional.pad(x, (0, 0, 0, 1), "constant", 0) - # x is of shape (N, 1, T + 1, idim) - x = self.conv_1(x) - # Now x is of shape (N, odim, T // 2, (idim - 1) // 2) - x = nn.functional.pad(x, (0, 0, 0, 1), "constant", 0) - # x is of shape (N, odim, T // 2 + 1, (idim - 1) // 2) - x = self.conv_2(x) - # Now x is of shape (N, odim, T // 4, ((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 // 4, odim) - return x - - -class VggSubsampling(nn.Module): - """Trying to follow the setup described in the following paper: - https://arxiv.org/pdf/1910.09799.pdf - - This paper is not 100% explicit so I am guessing to some extent, - and trying to compare with other VGG implementations. - - Convert an input of shape (N, T, idim) to an output - with shape (N, T', odim), where approximates T' = T//4. - """ - - def __init__(self, idim: int, odim: int) -> None: - """Construct a VggSubsampling object. - - This uses 2 VGG blocks with 2 Conv2d layers each, - subsampling its input by a factor of 4 in the time dimensions. - - Args: - idim: - Input dim. The input shape is (N, T, idim). - Caution: It requires: T >= 4, idim >= 4. - odim: - Output dim. The output shape is (N, T // 4, odim) - """ - super().__init__() - - cur_channels = 1 - layers = [] - block_dims = [32, 64] - - # The decision to use padding=1 for the 1st convolution, then padding=0 - # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by - # a back-compatibility concern so that the number of frames at the - # output would be equal to: - # (((T-1)//2)-1)//2. - # We can consider changing this by using padding=1 on the - # 2nd convolution, so the num-frames at the output would be T//4. - for block_dim in block_dims: - layers.append( - torch.nn.Conv2d( - in_channels=cur_channels, - out_channels=block_dim, - kernel_size=3, - padding=1, - stride=1, - ) - ) - layers.append(torch.nn.ReLU()) - layers.append( - torch.nn.Conv2d( - in_channels=block_dim, - out_channels=block_dim, - kernel_size=3, - padding=1, - stride=1, - ) - ) - layers.append( - torch.nn.MaxPool2d( - kernel_size=2, stride=2, padding=0, ceil_mode=False - ) - ) - cur_channels = block_dim - - self.layers = nn.Sequential(*layers) - - self.out = nn.Linear(block_dims[-1] * (idim // 4), odim) - - 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 // 4, odim) - """ - x = x.unsqueeze(1) - x = self.layers(x) - b, c, t, f = x.size() - x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) - return x diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/subsampling.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/subsampling.py new file mode 120000 index 000000000..6fee09e58 --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/subsampling.py @@ -0,0 +1 @@ +../conformer_ctc/subsampling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_emformer.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_emformer.py index ae93a4c8f..4c9cbba9c 100644 --- a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_emformer.py +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_emformer.py @@ -255,9 +255,9 @@ def test_emformer_forward(): from emformer import Emformer num_features = 80 output_dim = 1000 - chunk_length = 16 - L, R = 32, 16 - B, D, U = 2, 256, 48 + chunk_length = 8 + L, R = 128, 4 + B, D, U = 2, 256, 80 for use_memory in [True, False]: if use_memory: M = 3 @@ -274,13 +274,14 @@ def test_emformer_forward(): max_memory_size=M, vgg_frontend=False, ) - x = torch.randn(B, U + R, num_features) - x_lens = torch.randint(1, U + R + 1, (B,)) - x_lens[0] = U + R + x = torch.randn(B, U + R + 3, num_features) + x_lens = torch.randint(1, U + R + 3 + 1, (B,)) + x_lens[0] = U + R + 3 logits, output_lengths = model(x, x_lens) assert logits.shape == (B, U // 4, output_dim) assert torch.equal( - output_lengths, torch.clamp(x_lens // 4 - R // 4, min=0) + output_lengths, + torch.clamp(((x_lens - 1) // 2 - 1) // 2 - R // 4, min=0) ) @@ -288,9 +289,9 @@ def test_emformer_infer(): from emformer import Emformer num_features = 80 output_dim = 1000 - chunk_length = 16 + chunk_length = 8 U = chunk_length - L, R = 32, 16 + L, R = 128, 4 B, D = 2, 256 num_chunks = 3 num_encoder_layers = 2 @@ -313,14 +314,15 @@ def test_emformer_infer(): ) states = None for chunk_idx in range(num_chunks): - x = torch.randn(B, U + R, num_features) - x_lens = torch.randint(1, U + R + 1, (B,)) - x_lens[0] = U + R + x = torch.randn(B, U + R + 3, num_features) + x_lens = torch.randint(1, U + R + 3 + 1, (B,)) + x_lens[0] = U + R + 3 logits, output_lengths, states = \ model.infer(x, x_lens, states) assert logits.shape == (B, U // 4, output_dim) assert torch.equal( - output_lengths, torch.clamp(x_lens // 4 - R // 4, min=0) + output_lengths, + torch.clamp(((x_lens - 1) // 2 - 1) // 2 - R // 4, min=0) ) assert len(states) == num_encoder_layers for state in states: @@ -330,7 +332,7 @@ def test_emformer_infer(): assert state[2].shape == (L // 4, B, D) assert torch.equal( state[3], - (chunk_idx + 1) * U // 4 * torch.ones_like(state[3]) + U // 4 * (chunk_idx + 1) * torch.ones_like(state[3]) ) diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_subsampling.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_subsampling.py deleted file mode 100644 index 338688564..000000000 --- a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/test_subsampling.py +++ /dev/null @@ -1,25 +0,0 @@ -import torch -from subsampling import Conv2dSubsampling, VggSubsampling - - -def test_conv2d_subsampling(): - B, idim, odim = 1, 80, 512 - model = Conv2dSubsampling(idim, odim) - for t in range(4, 50): - x = torch.randn(B, t, idim) - outputs = model(x) - assert outputs.shape == (B, t // 4, odim) - - -def test_vgg_subsampling(): - B, idim, odim = 1, 80, 512 - model = VggSubsampling(idim, odim) - for t in range(4, 50): - x = torch.randn(B, t, idim) - outputs = model(x) - assert outputs.shape == (B, t // 4, odim) - - -if __name__ == "__main__": - test_conv2d_subsampling() - test_vgg_subsampling() diff --git a/egs/librispeech/ASR/emformer_pruned_transducer_stateless/train.py b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/train.py new file mode 100755 index 000000000..d7285f4a5 --- /dev/null +++ b/egs/librispeech/ASR/emformer_pruned_transducer_stateless/train.py @@ -0,0 +1,998 @@ +#!/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" + +./transducer_emformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --exp-dir transducer_emformer/exp \ + --full-libri 1 \ + --max-duration 300 +""" + + +import argparse +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from emformer import Emformer +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from noam import Noam +from torch import Tensor +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.nn.utils import clip_grad_norm_ +from torch.utils.tensorboard import SummaryWriter + +from 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, + measure_gradient_norms, + measure_weight_norms, + optim_step_and_measure_param_change, + setup_logger, + str2bool, +) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--attention-dim", + type=int, + default=512, + help="Attention dim for the Emformer", + ) + + parser.add_argument( + "--nhead", + type=int, + default=8, + help="Number of attention heads for the Emformer", + ) + + parser.add_argument( + "--dim-feedforward", + type=int, + default=2048, + help="Feed-forward dimension for the Emformer", + ) + + parser.add_argument( + "--num-encoder-layers", + type=int, + default=12, + help="Number of encoder layers for the Emformer", + ) + + parser.add_argument( + "--left-context-length", + type=int, + default=120, + help="Number of frames for the left context in the Emformer", + ) + + parser.add_argument( + "--chunk-length", + type=int, + default=16, + help="Number of frames for each segment in the Emformer", + ) + + parser.add_argument( + "--right-context-length", + type=int, + default=4, + help="Number of frames for right context in the Emformer", + ) + + parser.add_argument( + "--memory-size", + type=int, + default=0, + help="Number of entries in the memory for the Emformer", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=0, + help="""Resume training from from this epoch. + If it is positive, it will load checkpoint from + transducer_emformer/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="transducer_emformer/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( + "--lr-factor", + type=float, + default=5.0, + help="The lr_factor for Noam optimizer", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + 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( + "--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`. + """, + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - attention_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + "log_diagnostics": False, + # parameters for Emformer + "feature_dim": 80, + "subsampling_factor": 4, + "vgg_frontend": False, + # parameters for decoder + "embedding_dim": 512, + # parameters for Noam + "warm_step": 80000, # For the 100h subset, use 20000 + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Emformer( + num_features=params.feature_dim, + output_dim=params.vocab_size, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + left_context_length=params.left_context_length, + chunk_length=params.chunk_length, + right_context_length=params.right_context_length, + max_memory_size=params.memory_size, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.embedding_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + input_dim=params.vocab_size, + inner_dim=params.embedding_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = 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. + 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, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + sampler: Optional[CutSampler] = 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. + """ + 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, + sampler=sampler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Emformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + loss = params.simple_loss_scale * simple_loss + 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, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + 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. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + def maybe_log_gradients(tag: str): + if ( + params.log_diagnostics + and tb_writer is not None + and params.batch_idx_train % (params.log_interval * 5) == 0 + ): + tb_writer.add_scalars( + tag, + measure_gradient_norms(model, norm="l2"), + global_step=params.batch_idx_train, + ) + + def maybe_log_weights(tag: str): + if ( + params.log_diagnostics + and tb_writer is not None + and params.batch_idx_train % (params.log_interval * 5) == 0 + ): + tb_writer.add_scalars( + tag, + measure_weight_norms(model, norm="l2"), + global_step=params.batch_idx_train, + ) + + def maybe_log_param_relative_changes(): + if ( + params.log_diagnostics + and tb_writer is not None + and params.batch_idx_train % (params.log_interval * 5) == 0 + ): + deltas = optim_step_and_measure_param_change(model, optimizer) + tb_writer.add_scalars( + "train/relative_param_change_per_minibatch", + deltas, + global_step=params.batch_idx_train, + ) + else: + optimizer.step() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + + loss.backward() + + maybe_log_weights("train/param_norms") + maybe_log_gradients("train/grad_norms") + maybe_log_param_relative_changes() + + optimizer.zero_grad() + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + params=params, + optimizer=optimizer, + sampler=train_dl.sampler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}" + ) + + if tb_writer is not None: + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + + if batch_idx > 0 and batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + if params.full_libri is False: + params.valid_interval = 800 + params.warm_step = 20000 + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoints = load_checkpoint_if_available(params=params, model=model) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank]) + model.device = device + + optimizer = Noam( + model.parameters(), + model_size=params.attention_dim, + factor=params.lr_factor, + warm_step=params.warm_step, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + num_in_total = len(train_cuts) + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + num_left = len(train_cuts) + num_removed = num_in_total - num_left + removed_percent = num_removed / num_in_total * 100 + + logging.info(f"Before removing short and long utterances: {num_in_total}") + logging.info(f"After removing short and long utterances: {num_left}") + logging.info(f"Removed {num_removed} utterances ({removed_percent:.5f}%)") + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + for epoch in range(params.start_epoch, params.num_epochs): + fix_random_seed(params.seed + epoch) + train_dl.sampler.set_epoch(epoch) + + cur_lr = optimizer._rate + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + if rank == 0: + logging.info("epoch {}, learning rate {}".format(epoch, cur_lr)) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + optimizer=optimizer, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + save_checkpoint( + params=params, + model=model, + optimizer=optimizer, + sampler=train_dl.sampler, + 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: + optimizer.zero_grad() + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + clip_grad_norm_(model.parameters(), 5.0, 2.0) + optimizer.step() + except RuntimeError as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main()