From 8129470586be92cda3708cc49945fbfbd71ca176 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Sun, 10 Apr 2022 20:24:20 +0800 Subject: [PATCH] first upload the conv_emformer_transducer recipe, integrating convolution module into emformer layers. --- .../asr_datamodule.py | 1 + .../conv_emformer_transducer/beam_search.py | 1 + .../ASR/conv_emformer_transducer/decode.py | 549 +++++++ .../ASR/conv_emformer_transducer/decoder.py | 1 + .../ASR/conv_emformer_transducer/emformer.py | 1445 +++++++++++++++++ .../encoder_interface.py | 1 + .../ASR/conv_emformer_transducer/joiner.py | 1 + .../ASR/conv_emformer_transducer/model.py | 1 + .../ASR/conv_emformer_transducer/noam.py | 104 ++ .../conv_emformer_transducer/subsampling.py | 1 + .../conv_emformer_transducer/test_emformer.py | 359 ++++ .../ASR/conv_emformer_transducer/train.py | 1006 ++++++++++++ 12 files changed, 3470 insertions(+) create mode 120000 egs/librispeech/ASR/conv_emformer_transducer/asr_datamodule.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer/beam_search.py create mode 100755 egs/librispeech/ASR/conv_emformer_transducer/decode.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer/decoder.py create mode 100644 egs/librispeech/ASR/conv_emformer_transducer/emformer.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer/encoder_interface.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer/joiner.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer/model.py create mode 100644 egs/librispeech/ASR/conv_emformer_transducer/noam.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer/subsampling.py create mode 100644 egs/librispeech/ASR/conv_emformer_transducer/test_emformer.py create mode 100755 egs/librispeech/ASR/conv_emformer_transducer/train.py diff --git a/egs/librispeech/ASR/conv_emformer_transducer/asr_datamodule.py b/egs/librispeech/ASR/conv_emformer_transducer/asr_datamodule.py new file mode 120000 index 000000000..b4e5427e0 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer/beam_search.py new file mode 120000 index 000000000..227d2247c --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer/decode.py b/egs/librispeech/ASR/conv_emformer_transducer/decode.py new file mode 100755 index 000000000..c40b01dfa --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/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/conv_emformer_transducer/decoder.py b/egs/librispeech/ASR/conv_emformer_transducer/decoder.py new file mode 120000 index 000000000..0d5f10dc0 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/decoder.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer/emformer.py new file mode 100644 index 000000000..49e59bd00 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/emformer.py @@ -0,0 +1,1445 @@ +# Copyright 2022 Xiaomi Corporation (Author: Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# +# It is modified based on +# https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py. + +import math +import warnings +from typing import List, Optional, Tuple + +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from subsampling import Conv2dSubsampling, VggSubsampling + +from icefall.utils import make_pad_mask + + +def _gen_attention_mask_block( + col_widths: List[int], + col_mask: List[bool], + num_rows: int, + device: torch.device, +) -> torch.Tensor: + assert len(col_widths) == len( + col_mask + ), "Length of col_widths must match that of col_mask" + + mask_block = [ + torch.ones(num_rows, col_width, device=device) + if is_ones_col + else torch.zeros(num_rows, col_width, device=device) + for col_width, is_ones_col in zip(col_widths, col_mask) + ] + return torch.cat(mask_block, dim=1) + + +class EmformerAttention(nn.Module): + r"""Emformer layer attention module. + + Args: + embed_dim (int): + Embedding dimension. + nhead (int): + Number of attention heads in each Emformer layer. + dropout (float, optional): + Dropout probability. (Default: 0.0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + embed_dim: int, + nhead: int, + dropout: float = 0.0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + if embed_dim % nhead != 0: + raise ValueError( + f"embed_dim ({embed_dim}) is not a multiple of" + f"nhead ({nhead})." + ) + + self.embed_dim = embed_dim + self.nhead = nhead + self.dropout = dropout + self.tanh_on_mem = tanh_on_mem + self.negative_inf = negative_inf + + self.scaling = (self.embed_dim // self.nhead) ** -0.5 + + self.emb_to_key_value = nn.Linear(embed_dim, 2 * embed_dim, bias=True) + self.emb_to_query = nn.Linear(embed_dim, embed_dim, bias=True) + self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True) + + def _reset_parameters(self) -> None: + nn.init.xavier_uniform_(self.emb_to_key_value.weight) + nn.init.constant_(self.emb_to_key_value.bias, 0.0) + + nn.init.xavier_uniform_(self.emb_to_query.weight) + nn.init.constant_(self.emb_to_query.bias, 0.0) + + nn.init.xavier_uniform_(self.out_proj.weight) + nn.init.constant_(self.out_proj.bias, 0.0) + + def _gen_attention_probs( + self, + attention_weights: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor], + ) -> torch.Tensor: + """Given the entire attention weights, mask out unecessary connections + and optionally with padding positions, to obtain underlying chunk-wise + attention probabilities. + + B: batch size; + Q: length of query; + KV: length of key and value. + + Args: + attention_weights (torch.Tensor): + Attention weights computed on the entire concatenated tensor + with shape (B * nhead, Q, KV). + attention_mask (torch.Tensor): + Mask tensor where chunk-wise connections are filled with `False`, + and other unnecessary connections are filled with `True`, + with shape (Q, KV). + padding_mask (torch.Tensor, optional): + Mask tensor where the padding positions are fill with `True`, + and other positions are filled with `False`, with shapa `(B, KV)`. + + Returns: + A tensor of shape (B * nhead, Q, KV). + """ + attention_weights_float = attention_weights.float() + attention_weights_float = attention_weights_float.masked_fill( + attention_mask.unsqueeze(0), self.negative_inf + ) + if padding_mask is not None: + Q = attention_weights.size(1) + B = attention_weights.size(0) // self.nhead + attention_weights_float = attention_weights_float.view( + B, self.nhead, Q, -1 + ) + attention_weights_float = attention_weights_float.masked_fill( + padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + self.negative_inf, + ) + attention_weights_float = attention_weights_float.view( + B * self.nhead, Q, -1 + ) + + attention_probs = nn.functional.softmax( + attention_weights_float, dim=-1 + ).type_as(attention_weights) + attention_probs = nn.functional.dropout( + attention_probs, p=float(self.dropout), training=self.training + ) + return attention_probs + + def _forward_impl( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + left_context_key: Optional[torch.Tensor] = None, + left_context_val: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Underlying chunk-wise attention implementation. + + L: length of left_context; + S: length of summary; + M: length of memory; + Q: length of attention query; + KV: length of attention key and value. + + 1) Concat right_context, utterance, summary, + and compute query tensor with length Q = R + U + S. + 2) Concat memory, right_context, utterance, + and compute key, value tensors with length KV = M + R + U; + optionally with left_context_key and left_context_val (inference mode), + then KV = M + R + L + U. + 3) Compute entire attention scores with query, key, and value, + then apply attention_mask to get underlying chunk-wise attention scores. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing + number of valid frames for i-th batch element in utterance. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + summary (torch.Tensor): + Summary elements, with shape (S, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + attention_mask (torch.Tensor): + Attention mask for underlying attention, with shape (Q, KV). + left_context_key (torch,Tensor, optional): + Cached attention key of left context from preceding computation, + with shape (L, B, D). + left_context_val (torch.Tensor, optional): + Cached attention value of left context from preceding computation, + with shape (L, B, D). + + Returns: + A tuple containing 4 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (S, B, D). + - attention key, with shape (KV, B, D). + - attention value, with shape (KV, B, D). + """ + B = utterance.size(1) + + # Compute query with [right context, utterance, summary]. + query = self.emb_to_query( + torch.cat([right_context, utterance, summary]) + ) + # Compute key and value with [mems, right context, utterance]. + key, value = self.emb_to_key_value( + torch.cat([memory, right_context, utterance]) + ).chunk(chunks=2, dim=2) + + if left_context_key is not None and left_context_val is not None: + # This is for inference mode. Now compute key and value with + # [mems, right context, left context, uttrance] + M = memory.size(0) + R = right_context.size(0) + right_context_end_idx = M + R + key = torch.cat( + [ + key[:right_context_end_idx], + left_context_key, + key[right_context_end_idx:], + ] + ) + value = torch.cat( + [ + value[:right_context_end_idx], + left_context_val, + value[right_context_end_idx:], + ] + ) + + # Compute attention weights from query, key, and value. + reshaped_query, reshaped_key, reshaped_value = [ + tensor.contiguous() + .view(-1, B * self.nhead, self.embed_dim // self.nhead) + .transpose(0, 1) + for tensor in [query, key, value] + ] + attention_weights = torch.bmm( + reshaped_query * self.scaling, reshaped_key.transpose(1, 2) + ) + + # Compute padding mask + if B == 1: + padding_mask = None + else: + KV = key.size(0) + U = utterance.size(0) + padding_mask = make_pad_mask(KV - U + lengths) + + # Compute attention probabilities. + attention_probs = self._gen_attention_probs( + attention_weights, attention_mask, padding_mask + ) + + # Compute attention. + attention = torch.bmm(attention_probs, reshaped_value) + Q = query.size(0) + assert attention.shape == ( + B * self.nhead, + Q, + self.embed_dim // self.nhead, + ) + attention = ( + attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim) + ) + + # Apply output projection. + outputs = self.out_proj(attention) + + S = summary.size(0) + summary_start_idx = Q - S + output_right_context_utterance = outputs[:summary_start_idx] + output_memory = outputs[summary_start_idx:] + if self.tanh_on_mem: + output_memory = torch.tanh(output_memory) + else: + output_memory = torch.clamp(output_memory, min=-10, max=10) + + return output_right_context_utterance, output_memory, key, value + + def forward( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + # TODO: Modify docs. + """Forward pass for training. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + S: length of summary; + M: length of memory. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing + number of valid frames for i-th batch element in utterance. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + summary (torch.Tensor): + Summary elements, with shape (S, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + attention_mask (torch.Tensor): + Attention mask for underlying chunk-wise attention, + with shape (Q, KV), where Q = R + U + S, KV = M + R + U. + + Returns: + A tuple containing 2 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (M, B, D), where M = S - 1 or M = 0. + """ + ( + output_right_context_utterance, + output_memory, + _, + _, + ) = self._forward_impl( + utterance, lengths, right_context, summary, memory, attention_mask + ) + return output_right_context_utterance, output_memory[:-1] + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + left_context_key: torch.Tensor, + left_context_val: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + L: length of left_context; + S: length of summary; + M: length of memory; + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing + number of valid frames for i-th batch element in utterance. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + summary (torch.Tensor): + Summary element, with shape (1, B, D), or empty. + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + left_context_key (torch,Tensor): + Cached attention key of left context from preceding computation, + with shape (L, B, D). + left_context_val (torch.Tensor): + Cached attention value of left context from preceding computation, + with shape (L, B, D). + + Returns: + A tuple containing 4 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (1, B, D) or (0, B, D). + - attention key of left context and utterance, which would be cached + for next computation, with shape (L + U, B, D). + - attention value of left context and utterance, which would be + cached for next computation, with shape (L + U, B, D). + """ + # query: [right context, utterance, summary] + Q = right_context.size(0) + utterance.size(0) + summary.size(0) + # key, value: [memory, right context, left context, uttrance] + KV = ( + memory.size(0) + + right_context.size(0) + + left_context_key.size(0) + + utterance.size(0) + ) + attention_mask = torch.zeros(Q, KV).to( + dtype=torch.bool, device=utterance.device + ) + # Disallow attention bettween the summary vector with the memory bank + attention_mask[-1, : memory.size(0)] = True + ( + output_right_context_utterance, + output_memory, + key, + value, + ) = self._forward_impl( + utterance, + lengths, + right_context, + summary, + memory, + attention_mask, + left_context_key=left_context_key, + left_context_val=left_context_val, + ) + right_context_end_idx = memory.size(0) + right_context.size(0) + return ( + output_right_context_utterance, + output_memory, + key[right_context_end_idx:], + value[right_context_end_idx:], + ) + + +class EmformerLayer(nn.Module): + """Emformer layer that constitutes Emformer. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads. + dim_feedforward (int): + Hidden layer dimension of feedforward network. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (Default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (Default: 0) + max_memory_size (int, optional): + Maximum number of memory elements to use. (Default: 0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int, + chunk_length: int, + dropout: float = 0.0, + cnn_module_kernel: int = 3, + left_context_length: int = 0, + max_memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.attention = EmformerAttention( + embed_dim=d_model, + nhead=nhead, + dropout=0.0, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + self.summary_op = nn.AvgPool1d( + kernel_size=chunk_length, stride=chunk_length, ceil_mode=True + ) + + self.feed_forward_macaron = nn.Sequential( + nn.Linear(d_model, dim_feedforward), + Swish(), + nn.Dropout(dropout), + nn.Linear(dim_feedforward, d_model), + ) + + self.feed_forward = nn.Sequential( + nn.Linear(d_model, dim_feedforward), + Swish(), + nn.Dropout(dropout), + nn.Linear(dim_feedforward, d_model), + ) + + self.conv_module = ConvolutionModule(d_model, cnn_module_kernel) + + self.norm_ff_macaron = nn.LayerNorm(d_model) + self.norm_ff = nn.LayerNorm(d_model) + self.norm_mha = nn.LayerNorm(d_model) + self.norm_conv = nn.LayerNorm(d_model) + self.norm_final = nn.LayerNorm(d_model) + + self.dropout = nn.Dropout(dropout) + + self.ff_scale = 0.5 + self.left_context_length = left_context_length + self.chunk_length = chunk_length + self.max_memory_size = max_memory_size + self.d_model = d_model + self.use_memory = max_memory_size > 0 + + def _init_state( + self, batch_size: int, device: Optional[torch.device] + ) -> List[torch.Tensor]: + """Initialize states with zeros.""" + empty_memory = torch.zeros( + self.max_memory_size, batch_size, self.d_model, device=device + ) + left_context_key = torch.zeros( + self.left_context_length, batch_size, self.d_model, device=device + ) + left_context_val = torch.zeros( + self.left_context_length, batch_size, self.d_model, device=device + ) + past_length = torch.zeros( + 1, batch_size, dtype=torch.int32, device=device + ) + return [empty_memory, left_context_key, left_context_val, past_length] + + def _unpack_state( + self, state: List[torch.Tensor] + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Unpack cached states including: + 1) output memory from previous chunks in the lower layer; + 2) attention key and value of left context from proceeding chunk's + computation. + """ + past_length = state[3][0][0].item() + past_left_context_length = min(self.left_context_length, past_length) + past_memory_length = min( + self.max_memory_size, math.ceil(past_length / self.chunk_length) + ) + memory_start_idx = self.max_memory_size - past_memory_length + pre_memory = state[0][memory_start_idx:] + left_context_start_idx = ( + self.left_context_length - past_left_context_length + ) + left_context_key = state[1][left_context_start_idx:] + left_context_val = state[2][left_context_start_idx:] + return pre_memory, left_context_key, left_context_val + + def _pack_state( + self, + next_key: torch.Tensor, + next_val: torch.Tensor, + update_length: int, + memory: torch.Tensor, + state: List[torch.Tensor], + ) -> List[torch.Tensor]: + """Pack updated states including: + 1) output memory of current chunk in the lower layer; + 2) attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + 3) length of current chunk. + """ + new_memory = torch.cat([state[0], memory]) + new_key = torch.cat([state[1], next_key]) + new_val = torch.cat([state[2], next_val]) + memory_start_idx = new_memory.size(0) - self.max_memory_size + state[0] = new_memory[memory_start_idx:] + key_start_idx = new_key.size(0) - self.left_context_length + state[1] = new_key[key_start_idx:] + val_start_idx = new_val.size(0) - self.left_context_length + state[2] = new_val[val_start_idx:] + state[3] = state[3] + update_length + return state + + def _apply_macaron_feed_foward_module( + self, right_context_utterance: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply macaron style feed forward module.""" + residual = right_context_utterance + right_context_utterance = self.norm_ff_macaron(right_context_utterance) + right_context_utterance = residual + self.ff_scale * self.dropout( + self.feed_forward_macaron(right_context_utterance) + ) + return right_context_utterance + + def _apply_conv_module( + self, + right_context_utterance: torch.Tensor, + right_context_end_idx: int, + ) -> torch.Tensor: + """Apply convolution module on utterance.""" + utterance = right_context_utterance[right_context_end_idx:] + right_context = right_context_utterance[:right_context_end_idx] + + residual = utterance + utterance = self.norm_conv(utterance) + utterance = residual + self.dropout(self.conv_module(utterance)) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance + + def _apply_feed_forward_module( + self, + right_context_utterance: torch.Tensor, + ) -> torch.Tensor: + """Apply feed forward module.""" + residual = right_context_utterance + right_context_utterance = self.norm_ff(right_context_utterance) + right_context_utterance = residual + self.ff_scale * self.dropout( + self.feed_forward(right_context_utterance) + ) + return right_context_utterance + + def _apply_attention_module_forward( + self, + right_context_utterance: torch.Tensor, + right_context_end_idx: int, + lengths: torch.Tensor, + memory: torch.Tensor, + attention_mask: Optional[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply attention module in non-infer mode.""" + if attention_mask is None: + raise ValueError( + "attention_mask must be not None in non-infer mode. " + ) + + residual = right_context_utterance + right_context_utterance = self.norm_mha(right_context_utterance) + utterance = right_context_utterance[right_context_end_idx:] + right_context = right_context_utterance[:right_context_end_idx] + + if self.use_memory: + summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + ) + else: + summary = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + output_right_context_utterance, output_memory = self.attention( + utterance=utterance, + lengths=lengths, + right_context=right_context, + summary=summary, + memory=memory, + attention_mask=attention_mask, + ) + right_context_utterance = residual + self.dropout( + output_right_context_utterance + ) + + return right_context_utterance, output_memory + + def _apply_attention_module_infer( + self, + right_context_utterance: torch.Tensor, + right_context_end_idx: int, + lengths: torch.Tensor, + memory: torch.Tensor, + state: Optional[List[torch.Tensor]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]: + """Apply attention in infer mode. + 1) Unpack cached states including: + - memory from previous chunks in the lower layer; + - attention key and value of left context from proceeding + chunk's compuation; + 2) Apply attention computation; + 3) Pack updated states including: + - output memory of current chunk in the lower layer; + - attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + - length of current chunk. + """ + residual = right_context_utterance + right_context_utterance = self.norm_mha(right_context_utterance) + utterance = right_context_utterance[right_context_end_idx:] + right_context = right_context_utterance[:right_context_end_idx] + + if state is None: + state = self._init_state(utterance.size(1), device=utterance.device) + pre_memory, left_context_key, left_context_val = self._unpack_state( + state + ) + if self.use_memory: + summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + ) + summary = summary[:1] + else: + summary = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + ( + output_right_context_utterance, + output_memory, + next_key, + next_val, + ) = self.attention.infer( + utterance=utterance, + lengths=lengths, + right_context=right_context, + summary=summary, + memory=pre_memory, + left_context_key=left_context_key, + left_context_val=left_context_val, + ) + right_context_utterance = residual + self.dropout( + output_right_context_utterance + ) + state = self._pack_state( + next_key, next_val, utterance.size(0), memory, state + ) + return right_context_utterance, output_memory, state + + def forward( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + r"""Forward pass for training. + 1) Apply layer normalization on input utterance and right context + before attention; + 2) Apply attention module, compute updated utterance, right context, + and memory; + 3) Apply feed forward module and layer normalization on output utterance + and right context. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + M: length of memory. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing + number of valid frames for i-th batch element in utterance. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + attention_mask (torch.Tensor): + Attention mask for underlying attention module, + with shape (Q, KV), where Q = R + U + S, KV = M + R + U. + + Returns: + A tuple containing 3 tensors: + - output utterance, with shape (U, B, D). + - output right context, with shape (R, B, D). + - output memory, with shape (M, B, D). + """ + right_context_utterance = torch.cat([right_context, utterance]) + right_context_end_idx = right_context.size(0) + + right_context_utterance = self._apply_macaron_feed_foward_module( + right_context_utterance + ) + + ( + right_context_utterance, + output_memory, + ) = self._apply_attention_module_forward( + right_context_utterance, + right_context_end_idx, + lengths, + memory, + attention_mask, + ) + + right_context_utterance = self._apply_conv_module( + right_context_utterance, right_context_end_idx + ) + + right_context_utterance = self._apply_feed_forward_module( + right_context_utterance + ) + + right_context_utterance = self.norm_final(right_context_utterance) + + output_utterance = right_context_utterance[right_context_end_idx:] + output_right_context = right_context_utterance[:right_context_end_idx] + return output_utterance, output_right_context, output_memory + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + state: Optional[List[torch.Tensor]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor], torch.Tensor]: + """Forward pass for inference. + + 1) Apply layer normalization on input utterance and right context + before attention; + 2) Apply attention module with cached state, compute updated utterance, + right context, and memory, and update state; + 3) Apply feed forward module and layer normalization on output + utterance and right context. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + M: length of memory. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing + number of valid frames for i-th batch element in utterance. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + state (List[torch.Tensor], optional): + List of tensors representing layer internal state generated in + preceding computation. (default=None) + + Returns: + (Tensor, Tensor, List[torch.Tensor], Tensor): + - output utterance, with shape (U, B, D); + - output right_context, with shape (R, B, D); + - output memory, with shape (1, B, D) or (0, B, D). + - output state. + """ + right_context_utterance = torch.cat([right_context, utterance]) + right_context_end_idx = right_context.size(0) + + right_context_utterance = self._apply_macaron_feed_foward_module( + right_context_utterance + ) + + ( + right_context_utterance, + output_memory, + output_state, + ) = self._apply_attention_module_infer( + right_context_utterance, + right_context_end_idx, + lengths, + memory, + state, + ) + + right_context_utterance = self._apply_conv_module( + right_context_utterance, right_context_end_idx + ) + + right_context_utterance = self._apply_feed_forward_module( + right_context_utterance + ) + + right_context_utterance = self.norm_final(right_context_utterance) + + output_utterance = right_context_utterance[right_context_end_idx:] + output_right_context = right_context_utterance[:right_context_end_idx] + return ( + output_utterance, + output_right_context, + output_memory, + output_state, + ) + + +class EmformerEncoder(nn.Module): + """Implements the Emformer architecture introduced in + *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency + Streaming Speech Recognition* + [:footcite:`shi2021emformer`]. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads in each emformer layer. + dim_feedforward (int): + Hidden layer dimension of each emformer layer's feedforward network. + num_encoder_layers (int): + Number of emformer layers to instantiate. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (default: 0.0) + left_context_length (int, optional): + Length of left context. (default: 0) + right_context_length (int, optional): + Length of right context. (default: 0) + max_memory_size (int, optional): + Maximum number of memory elements to use. (default: 0) + tanh_on_mem (bool, optional): + If ``true``, applies tanh to memory elements. (default: ``false``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (default: -1e8) + """ + + def __init__( + self, + chunk_length: int, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + cnn_module_kernel: int = 3, + left_context_length: int = 0, + right_context_length: int = 0, + max_memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.use_memory = max_memory_size > 0 + self.init_memory_op = nn.AvgPool1d( + kernel_size=chunk_length, + stride=chunk_length, + ceil_mode=True, + ) + + self.emformer_layers = nn.ModuleList( + [ + EmformerLayer( + d_model, + nhead, + dim_feedforward, + chunk_length, + dropout=dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length, + max_memory_size=max_memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + for layer_idx in range(num_encoder_layers) + ] + ) + + self.left_context_length = left_context_length + self.right_context_length = right_context_length + self.chunk_length = chunk_length + self.max_memory_size = max_memory_size + + def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor: + """Hard copy each chunk's right context and concat them.""" + T = x.shape[0] + num_segs = math.ceil( + (T - self.right_context_length) / self.chunk_length + ) + right_context_blocks = [] + for seg_idx in range(num_segs - 1): + start = (seg_idx + 1) * self.chunk_length + end = start + self.right_context_length + right_context_blocks.append(x[start:end]) + last_right_context_start_idx = T - self.right_context_length + right_context_blocks.append(x[last_right_context_start_idx:]) + return torch.cat(right_context_blocks) + + def _gen_attention_mask_col_widths( + self, chunk_idx: int, U: int + ) -> List[int]: + """Calculate column widths (key, value) in attention mask for the + chunk_idx chunk.""" + num_chunks = math.ceil(U / self.chunk_length) + rc = self.right_context_length + lc = self.left_context_length + rc_start = chunk_idx * rc + rc_end = rc_start + rc + chunk_start = max(chunk_idx * self.chunk_length - lc, 0) + chunk_end = min((chunk_idx + 1) * self.chunk_length, U) + R = rc * num_chunks + + if self.use_memory: + m_start = max(chunk_idx - self.max_memory_size, 0) + M = num_chunks - 1 + col_widths = [ + m_start, # before memory + chunk_idx - m_start, # memory + M - chunk_idx, # after memory + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + else: + col_widths = [ + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + + return col_widths + + def _gen_attention_mask(self, utterance: torch.Tensor) -> torch.Tensor: + """Generate attention mask for underlying chunk-wise attention + computation, where chunk-wise connections are filled with `False`, + and other unnecessary connections beyond chunk are filled with `True`. + + R: length of right_context; + U: length of utterance; + S: length of summary; + M: length of memory; + Q: length of attention query; + KV: length of attention key and value. + + The shape of attention mask is (Q, KV). + If self.use_memory is `True`: + query = [right_context, utterance, summary]; + key, value = [memory, right_context, utterance]; + Q = R + U + S, KV = M + R + U. + Otherwise: + query = [right_context, utterance] + key, value = [right_context, utterance] + Q = R + U, KV = R + U. + """ + U = utterance.size(0) + num_chunks = math.ceil(U / self.chunk_length) + + right_context_mask = [] + utterance_mask = [] + summary_mask = [] + + if self.use_memory: + num_cols = 9 + # right context and utterance both attend to memory, right context, + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4, 7] for idx in range(num_cols) + ] + # summary attends to right context, utterance + summary_cols_mask = [idx in [4, 7] for idx in range(num_cols)] + masks_to_concat = [right_context_mask, utterance_mask, summary_mask] + else: + num_cols = 6 + # right context and utterance both attend to right context and + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4] for idx in range(num_cols) + ] + summary_cols_mask = None + masks_to_concat = [right_context_mask, utterance_mask] + + for chunk_idx in range(num_chunks): + col_widths = self._gen_attention_mask_col_widths(chunk_idx, U) + + right_context_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + self.right_context_length, + utterance.device, + ) + right_context_mask.append(right_context_mask_block) + + utterance_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + min( + self.chunk_length, + U - chunk_idx * self.chunk_length, + ), + utterance.device, + ) + utterance_mask.append(utterance_mask_block) + + if summary_cols_mask is not None: + summary_mask_block = _gen_attention_mask_block( + col_widths, summary_cols_mask, 1, utterance.device + ) + summary_mask.append(summary_mask_block) + + attention_mask = ( + 1 - torch.cat([torch.cat(mask) for mask in masks_to_concat]) + ).to(torch.bool) + return attention_mask + + def forward( + self, x: torch.Tensor, lengths: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and non-streaming inference. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + + Returns: + A tuple of 2 tensors: + - output utterance frames, with shape (U, B, D). + - output_lengths, with shape (B,), without containing the + right_context at the end. + """ + 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) + attention_mask = self._gen_attention_mask(utterance) + memory = ( + self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[ + :-1 + ] + if self.use_memory + else torch.empty(0).to(dtype=x.dtype, device=x.device) + ) + output = utterance + for layer in self.emformer_layers: + output, right_context, memory = layer( + output, output_lengths, right_context, memory, attention_mask + ) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + lengths: torch.Tensor, + states: Optional[List[List[torch.Tensor]]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]: + """Forward pass for streaming inference. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + states (List[List[torch.Tensor]], optional): + Cached states from proceeding chunk's computation, where each + element (List[torch.Tensor]) corresponding to each emformer layer. + (default: None) + + Returns: + (Tensor, Tensor, List[List[torch.Tensor]]): + - output utterance frames, with shape (U, B, D). + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + """ + assert x.size(0) == self.chunk_length + self.right_context_length, ( + "Per configured chunk_length and right_context_length, " + f"expected size of {self.chunk_length + self.right_context_length} " + f"for dimension 1 of x, but got {x.size(1)}." + ) + right_context_start_idx = x.size(0) - self.right_context_length + right_context = x[right_context_start_idx:] + utterance = x[:right_context_start_idx] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + memory = ( + self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1) + if self.use_memory + else torch.empty(0).to(dtype=x.dtype, device=x.device) + ) + output = utterance + output_states: List[List[torch.Tensor]] = [] + for layer_idx, layer in enumerate(self.emformer_layers): + output, right_context, memory, output_state = layer.infer( + output, + output_lengths, + right_context, + memory, + None if states is None else states[layer_idx], + ) + output_states.append(output_state) + + return output, output_lengths, output_states + + +class Emformer(EncoderInterface): + def __init__( + self, + num_features: int, + output_dim: int, + chunk_length: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + cnn_module_kernel: int = 3, + vgg_frontend: bool = False, + left_context_length: int = 0, + right_context_length: int = 0, + max_memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.subsampling_factor = subsampling_factor + self.right_context_length = right_context_length + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + if chunk_length % 4 != 0: + raise NotImplementedError("chunk_length must be a mutiple of 4.") + if left_context_length != 0 and left_context_length % 4 != 0: + raise NotImplementedError( + "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 0 or a mutiple of 4." + ) + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + if vgg_frontend: + self.encoder_embed = VggSubsampling(num_features, d_model) + else: + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder = EmformerEncoder( + chunk_length // 4, + d_model, + nhead, + dim_feedforward, + num_encoder_layers, + dropout, + cnn_module_kernel, + left_context_length=left_context_length // 4, + right_context_length=right_context_length // 4, + max_memory_size=max_memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + + # TODO(fangjun): remove dropout + self.encoder_output_layer = nn.Sequential( + nn.Dropout(p=dropout), nn.Linear(d_model, output_dim) + ) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and non-streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + x_lens (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + + Returns: + (Tensor, Tensor): + - output logits, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - logits lengths, with shape (B,), without containing the + right_context at the end. + """ + x = self.encoder_embed(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + 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) + + return logits, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + states: Optional[List[List[torch.Tensor]]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]: + """Forward pass for streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + states (List[List[torch.Tensor]], optional): + Cached states from proceeding chunk's computation, where each + element (List[torch.Tensor]) corresponding to each emformer layer. + (default: None) + Returns: + (Tensor, Tensor): + - output logits, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - 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 = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + 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, 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 ConvolutionModule(nn.Module): + """ConvolutionModule in Conformer model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py # noqa + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__( + self, channels: int, kernel_size: int, bias: bool = True + ) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + self.pointwise_conv1 = nn.Conv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.depthwise_conv = nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + padding=(kernel_size - 1) // 2, + groups=channels, + bias=bias, + ) + self.norm = nn.LayerNorm(channels) + self.pointwise_conv2 = nn.Conv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.activation = Swish() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + # x is (batch, channels, time) + x = x.permute(0, 2, 1) + x = self.norm(x) + x = x.permute(0, 2, 1) + + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + return x.permute(2, 0, 1) + + +class Swish(torch.nn.Module): + """Construct an Swish object.""" + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Return Swich activation function.""" + return x * torch.sigmoid(x) diff --git a/egs/librispeech/ASR/conv_emformer_transducer/encoder_interface.py b/egs/librispeech/ASR/conv_emformer_transducer/encoder_interface.py new file mode 120000 index 000000000..aa5d0217a --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/encoder_interface.py @@ -0,0 +1 @@ +../transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer/joiner.py b/egs/librispeech/ASR/conv_emformer_transducer/joiner.py new file mode 120000 index 000000000..81ad47c55 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/joiner.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer/model.py b/egs/librispeech/ASR/conv_emformer_transducer/model.py new file mode 120000 index 000000000..a61a0a23f --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/model.py @@ -0,0 +1 @@ +../pruned_transducer_stateless/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer/noam.py b/egs/librispeech/ASR/conv_emformer_transducer/noam.py new file mode 100644 index 000000000..e46bf35fb --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/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/conv_emformer_transducer/subsampling.py b/egs/librispeech/ASR/conv_emformer_transducer/subsampling.py new file mode 120000 index 000000000..6fee09e58 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/subsampling.py @@ -0,0 +1 @@ +../conformer_ctc/subsampling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer/test_emformer.py new file mode 100644 index 000000000..1f735637f --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/test_emformer.py @@ -0,0 +1,359 @@ +import torch + + +def test_emformer_attention_forward(): + from emformer import EmformerAttention + + B, D = 2, 256 + U, R = 12, 2 + chunk_length = 2 + attention = EmformerAttention(embed_dim=D, nhead=8) + + for use_memory in [True, False]: + if use_memory: + S = U // chunk_length + M = S - 1 + else: + S, M = 0, 0 + + Q, KV = R + U + S, M + R + U + utterance = torch.randn(U, B, D) + lengths = torch.randint(1, U + 1, (B,)) + lengths[0] = U + right_context = torch.randn(R, B, D) + summary = torch.randn(S, B, D) + memory = torch.randn(M, B, D) + attention_mask = torch.rand(Q, KV) >= 0.5 + + output_right_context_utterance, output_memory = attention( + utterance, + lengths, + right_context, + summary, + memory, + attention_mask, + ) + assert output_right_context_utterance.shape == (R + U, B, D) + assert output_memory.shape == (M, B, D) + + +def test_emformer_attention_infer(): + from emformer import EmformerAttention + + B, D = 2, 256 + R, L = 4, 2 + chunk_length = 2 + U = chunk_length + attention = EmformerAttention(embed_dim=D, nhead=8) + + for use_memory in [True, False]: + if use_memory: + S, M = 1, 3 + else: + S, M = 0, 0 + + utterance = torch.randn(U, B, D) + lengths = torch.randint(1, U + 1, (B,)) + lengths[0] = U + right_context = torch.randn(R, B, D) + summary = torch.randn(S, B, D) + memory = torch.randn(M, B, D) + left_context_key = torch.randn(L, B, D) + left_context_val = torch.randn(L, B, D) + + ( + output_right_context_utterance, + output_memory, + next_key, + next_val, + ) = attention.infer( + utterance, + lengths, + right_context, + summary, + memory, + left_context_key, + left_context_val, + ) + assert output_right_context_utterance.shape == (R + U, B, D) + assert output_memory.shape == (S, B, D) + assert next_key.shape == (L + U, B, D) + assert next_val.shape == (L + U, B, D) + + +def test_emformer_layer_forward(): + from emformer import EmformerLayer + + B, D = 2, 256 + U, R, L = 12, 2, 5 + chunk_length = 2 + + for use_memory in [True, False]: + if use_memory: + S = U // chunk_length + M = S - 1 + else: + S, M = 0, 0 + + layer = EmformerLayer( + d_model=D, + nhead=8, + dim_feedforward=1024, + chunk_length=chunk_length, + cnn_module_kernel=3, + left_context_length=L, + max_memory_size=M, + ) + + Q, KV = R + U + S, M + R + U + utterance = torch.randn(U, B, D) + lengths = torch.randint(1, U + 1, (B,)) + lengths[0] = U + right_context = torch.randn(R, B, D) + memory = torch.randn(M, B, D) + attention_mask = torch.rand(Q, KV) >= 0.5 + + output_utterance, output_right_context, output_memory = layer( + utterance, + lengths, + right_context, + memory, + attention_mask, + ) + assert output_utterance.shape == (U, B, D) + assert output_right_context.shape == (R, B, D) + assert output_memory.shape == (M, B, D) + + +def test_emformer_layer_infer(): + from emformer import EmformerLayer + + B, D = 2, 256 + R, L = 2, 5 + chunk_length = 2 + U = chunk_length + + for use_memory in [True, False]: + if use_memory: + M = 3 + else: + M = 0 + + layer = EmformerLayer( + d_model=D, + nhead=8, + dim_feedforward=1024, + chunk_length=chunk_length, + cnn_module_kernel=3, + left_context_length=L, + max_memory_size=M, + ) + + utterance = torch.randn(U, B, D) + lengths = torch.randint(1, U + 1, (B,)) + lengths[0] = U + right_context = torch.randn(R, B, D) + memory = torch.randn(M, B, D) + state = None + ( + output_utterance, + output_right_context, + output_memory, + output_state, + ) = layer.infer( + utterance, + lengths, + right_context, + memory, + state, + ) + assert output_utterance.shape == (U, B, D) + assert output_right_context.shape == (R, B, D) + if use_memory: + assert output_memory.shape == (1, B, D) + else: + assert output_memory.shape == (0, B, D) + assert len(output_state) == 4 + assert output_state[0].shape == (M, B, D) + assert output_state[1].shape == (L, B, D) + assert output_state[2].shape == (L, B, D) + assert output_state[3].shape == (1, B) + + +def test_emformer_encoder_forward(): + from emformer import EmformerEncoder + + B, D = 2, 256 + U, R, L = 12, 2, 5 + chunk_length = 2 + + for use_memory in [True, False]: + if use_memory: + S = U // chunk_length + M = S - 1 + else: + S, M = 0, 0 + + encoder = EmformerEncoder( + chunk_length=chunk_length, + d_model=D, + dim_feedforward=1024, + num_encoder_layers=2, + cnn_module_kernel=3, + left_context_length=L, + right_context_length=R, + max_memory_size=M, + ) + + x = torch.randn(U + R, B, D) + lengths = torch.randint(1, U + R + 1, (B,)) + lengths[0] = U + R + + output, output_lengths = encoder(x, lengths) + assert output.shape == (U, B, D) + assert torch.equal(output_lengths, torch.clamp(lengths - R, min=0)) + + +def test_emformer_encoder_infer(): + from emformer import EmformerEncoder + + B, D = 2, 256 + R, L = 2, 5 + chunk_length = 2 + U = chunk_length + num_chunks = 3 + num_encoder_layers = 2 + + for use_memory in [True, False]: + if use_memory: + M = 3 + else: + M = 0 + + encoder = EmformerEncoder( + chunk_length=chunk_length, + d_model=D, + dim_feedforward=1024, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=3, + left_context_length=L, + right_context_length=R, + max_memory_size=M, + ) + + states = None + for chunk_idx in range(num_chunks): + x = torch.randn(U + R, B, D) + lengths = torch.randint(1, U + R + 1, (B,)) + lengths[0] = U + R + output, output_lengths, states = encoder.infer(x, lengths, states) + assert output.shape == (U, B, D) + assert torch.equal(output_lengths, torch.clamp(lengths - R, min=0)) + assert len(states) == num_encoder_layers + for state in states: + assert len(state) == 4 + assert state[0].shape == (M, B, D) + assert state[1].shape == (L, B, D) + assert state[2].shape == (L, B, D) + assert torch.equal( + state[3], (chunk_idx + 1) * U * torch.ones_like(state[3]) + ) + + +def test_emformer_forward(): + from emformer import Emformer + + num_features = 80 + output_dim = 1000 + chunk_length = 8 + L, R = 128, 4 + B, D, U = 2, 256, 80 + for use_memory in [True, False]: + if use_memory: + M = 3 + else: + M = 0 + model = Emformer( + num_features=num_features, + output_dim=output_dim, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=D, + cnn_module_kernel=3, + left_context_length=L, + right_context_length=R, + max_memory_size=M, + vgg_frontend=False, + ) + 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 - 1) // 2 - 1) // 2 - R // 4, min=0), + ) + + +def test_emformer_infer(): + from emformer import Emformer + + num_features = 80 + output_dim = 1000 + chunk_length = 8 + U = chunk_length + L, R = 128, 4 + B, D = 2, 256 + num_chunks = 3 + num_encoder_layers = 2 + for use_memory in [True, False]: + if use_memory: + M = 3 + else: + M = 0 + model = Emformer( + num_features=num_features, + output_dim=output_dim, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=D, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=3, + left_context_length=L, + right_context_length=R, + max_memory_size=M, + vgg_frontend=False, + ) + states = None + for chunk_idx in range(num_chunks): + 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 - 1) // 2 - 1) // 2 - R // 4, min=0), + ) + assert len(states) == num_encoder_layers + for state in states: + assert len(state) == 4 + assert state[0].shape == (M, B, D) + assert state[1].shape == (L // 4, B, D) + assert state[2].shape == (L // 4, B, D) + assert torch.equal( + state[3], + U // 4 * (chunk_idx + 1) * torch.ones_like(state[3]), + ) + + +if __name__ == "__main__": + test_emformer_attention_forward() + test_emformer_attention_infer() + test_emformer_layer_forward() + test_emformer_layer_infer() + test_emformer_encoder_forward() + test_emformer_encoder_infer() + test_emformer_forward() + test_emformer_infer() diff --git a/egs/librispeech/ASR/conv_emformer_transducer/train.py b/egs/librispeech/ASR/conv_emformer_transducer/train.py new file mode 100755 index 000000000..bdb541ac6 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer/train.py @@ -0,0 +1,1006 @@ +#!/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( + "--cnn-module-kernel", + type=int, + default=3, + help="Kernel size for the convolution module.", + ) + + 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, + cnn_module_kernel=params.cnn_module_kernel, + 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()