diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py new file mode 120000 index 000000000..a074d6085 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py new file mode 100755 index 000000000..dd6885b19 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py @@ -0,0 +1,634 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./pruned_transducer_stateless4/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/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, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless4/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + 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, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i + 1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): hyps + } + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{params.decoding_method}-beam-size-{params.beam_size}" + ) + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0 + start = params.epoch - params.avg + assert start >= 1 + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py new file mode 120000 index 000000000..0793c5709 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decoder.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py new file mode 100644 index 000000000..8d989730b --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py @@ -0,0 +1,2181 @@ +# 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. # noqa + +import math +import warnings +from typing import List, Optional, Tuple + +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv1d, + ScaledConv2d, + ScaledLinear, +) + +from icefall.utils import make_pad_mask + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" # noqa + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py # noqa + + Suppose: + i -> position of query, + j -> position of key(value), + we use positive relative position embedding when key(value) is to the + left of query(i.e., i > j) and negative embedding otherwise. + + Args: + d_model: Embedding dimension. + dropout: Dropout rate. + max_len: Maximum input length. + """ + + def __init__( + self, d_model: int, dropout: float, max_len: int = 5000 + ) -> None: + """Construct an PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.xscale = math.sqrt(self.d_model) + self.dropout = torch.nn.Dropout(p=dropout) + self.pe = None + self.pos_len = max_len + self.neg_len = max_len + self.gen_pe_positive() + self.gen_pe_negative() + + def gen_pe_positive(self) -> None: + """Generate the positive positional encodings.""" + pe_positive = torch.zeros(self.pos_len, self.d_model) + position_positive = torch.arange( + 0, self.pos_len, dtype=torch.float32 + ).unsqueeze(1) + div_term = torch.exp( + torch.arange(0, self.d_model, 2, dtype=torch.float32) + * -(math.log(10000.0) / self.d_model) + ) + pe_positive[:, 0::2] = torch.sin(position_positive * div_term) + pe_positive[:, 1::2] = torch.cos(position_positive * div_term) + # Reserve the order of positive indices and concat both positive and + # negative indices. This is used to support the shifting trick + # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" # noqa + self.pe_positive = torch.flip(pe_positive, [0]) + + def gen_pe_negative(self) -> None: + """Generate the negative positional encodings.""" + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use positive relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i torch.Tensor: + """Get positional encoding given positive length and negative length.""" + if self.pe_positive.dtype != dtype or str( + self.pe_positive.device + ) != str(device): + self.pe_positive = self.pe_positive.to(dtype=dtype, device=device) + if self.pe_negative.dtype != dtype or str( + self.pe_negative.device + ) != str(device): + self.pe_negative = self.pe_negative.to(dtype=dtype, device=device) + pe = torch.cat( + [ + self.pe_positive[self.pos_len - pos_len :], + self.pe_negative[1:neg_len], + ], + dim=0, + ) + return pe + + def forward( + self, + x: torch.Tensor, + pos_len: int, + neg_len: int, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Scale input x and get positional encoding. + Args: + x (torch.Tensor): Input tensor (`*`). + + Returns: + torch.Tensor: + Encoded tensor of shape (`*`). + torch.Tensor: + Position embedding of shape (pos_len + neg_len - 1, `*`). + """ + x = x * self.xscale + if pos_len > self.pos_len: + self.pos_len = pos_len + self.gen_pe_positive() + if neg_len > self.neg_len: + self.neg_len = neg_len + self.gen_pe_negative() + pos_emb = self.get_pe(pos_len, neg_len, x.device, x.dtype) + return self.dropout(x), self.dropout(pos_emb) + + +class ConvolutionModule(nn.Module): + """ConvolutionModule. + + Modified from https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa + + Args: + chunk_length (int): + Length of each chunk. + right_context_length (int): + Length of right context. + channels (int): + The number of channels of conv layers. + kernel_size (int): + Kernerl size of conv layers. + bias (bool): + Whether to use bias in conv layers (default=True). + """ + + def __init__( + self, + chunk_length: int, + right_context_length: int, + channels: int, + kernel_size: int, + bias: bool = True, + ) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + self.chunk_length = chunk_length + self.right_context_length = right_context_length + self.channels = channels + + self.pointwise_conv1 = ScaledConv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + # After pointwise_conv1 we put x through a gated linear unit + # (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in + # the range 1 to 4, but sometimes, for some reason, for layer 0 the rms + # ends up being very large, between 50 and 100 for different channels. + # This will cause very peaky and sparse derivatives for the sigmoid + # gating function, which will tend to make the loss function not learn + # effectively. (for most layers the average absolute values are in the + # range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for + # different layers, which likely breaks down as 0.5 for the "linear" + # half and 0.2 to 0.3 for the part that goes into the sigmoid. + # The idea is that if we constrain the rms values to a reasonable range + # via a constraint of max_abs=10.0, it will be in a better position to + # start learning something, i.e. to latch onto the correct range. + self.deriv_balancer1 = ActivationBalancer( + channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0 + ) + + # make it causal by padding cached (kernel_size - 1) frames on the left + self.cache_size = kernel_size - 1 + self.depthwise_conv = ScaledConv1d( + channels, + channels, + kernel_size, + stride=1, + padding=0, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channel_dim=1, min_positive=0.05, max_positive=1.0 + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.25, + ) + + def _split_right_context( + self, + pad_utterance: torch.Tensor, + right_context: torch.Tensor, + ) -> torch.Tensor: + """ + Args: + pad_utterance: + Its shape is (cache_size + U, B, D). + right_context: + Its shape is (R, B, D). + + Returns: + Right context segments padding with corresponding context. + Its shape is (num_segs * B, D, cache_size + right_context_length). + """ + U_, B, D = pad_utterance.size() + R = right_context.size(0) + assert self.right_context_length != 0 + assert R % self.right_context_length == 0 + num_chunks = R // self.right_context_length + right_context = right_context.reshape( + num_chunks, self.right_context_length, B, D + ) + right_context = right_context.permute(0, 2, 1, 3).reshape( + num_chunks * B, self.right_context_length, D + ) + padding = [] + for idx in range(num_chunks): + end_idx = min(U_, self.cache_size + (idx + 1) * self.chunk_length) + start_idx = end_idx - self.cache_size + padding.append(pad_utterance[start_idx:end_idx]) + padding = torch.cat(padding, dim=1).permute(1, 0, 2) + # (num_segs * B, cache_size, D) + pad_right_context = torch.cat([padding, right_context], dim=1) + # (num_segs * B, cache_size + right_context_length, D) + return pad_right_context.permute(0, 2, 1) + + def _merge_right_context( + self, right_context: torch.Tensor, B: int + ) -> torch.Tensor: + """ + Args: + right_context: + Right context segments. + It shape is (num_segs * B, D, right_context_length). + B: + Batch size. + + Returns: + A tensor of shape (B, D, R), where + R = num_segs * right_context_length. + """ + right_context = right_context.reshape( + -1, B, self.channels, self.right_context_length + ) + right_context = right_context.permute(1, 2, 0, 3) + right_context = right_context.reshape(B, self.channels, -1) + return right_context + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + cache: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Causal convolution module. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + cache (torch.Tensor, optional): + Cached tensor for left padding of shape (B, D, cache_size). + + Returns: + A tuple of 3 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + - updated cache tensor of shape (B, D, cache_size). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv and GLU mechanism + x = torch.cat([right_context, utterance], dim=0) # (R + U, B, D) + x = x.permute(1, 2, 0) # (B, D, R + U) + x = self.pointwise_conv1(x) # (B, 2 * D, R + U) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, R + U) + utterance = x[:, :, R:] # (B, D, U) + right_context = x[:, :, :R] # (B, D, R) + + if cache is None: + cache = torch.zeros( + B, D, self.cache_size, device=x.device, dtype=x.dtype + ) + else: + assert cache.shape == (B, D, self.cache_size), cache.shape + pad_utterance = torch.cat( + [cache, utterance], dim=2 + ) # (B, D, cache + U) + # update cache + new_cache = pad_utterance[:, :, -self.cache_size :] + + # depth-wise conv on utterance + utterance = self.depthwise_conv(pad_utterance) # (B, D, U) + + if self.right_context_length > 0: + # depth-wise conv on right_context + pad_right_context = self._split_right_context( + pad_utterance.permute(2, 0, 1), right_context.permute(2, 0, 1) + ) # (num_segs * B, D, cache_size + right_context_length) + right_context = self.depthwise_conv( + pad_right_context + ) # (num_segs * B, D, right_context_length) + right_context = self._merge_right_context( + right_context, B + ) # (B, D, R) + + x = torch.cat([right_context, utterance], dim=2) # (B, D, R + U) + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, R + U) + + right_context = x[:, :, :R] # (B, D, R) + utterance = x[:, :, R:] # (B, D, U) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + new_cache, + ) + + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + cache: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Causal convolution module applied on both utterance and right_context. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + cache (torch.Tensor, optional): + Cached tensor for left padding of shape (B, D, cache_size). + + Returns: + A tuple of 3 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + - updated cache tensor of shape (B, D, cache_size). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv + x = torch.cat([utterance, right_context], dim=0) # (U + R, B, D) + x = x.permute(1, 2, 0) # (B, D, U + R) + x = self.pointwise_conv1(x) # (B, 2 * D, U + R) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, U + R) + + if cache is None: + cache = torch.zeros( + B, D, self.cache_size, device=x.device, dtype=x.dtype + ) + else: + assert cache.shape == (B, D, self.cache_size), cache.shape + x = torch.cat([cache, x], dim=2) # (B, D, cache_size + U + R) + # update cache + x_length = x.size(2) + new_cache = x[:, :, x_length - R - self.cache_size : x_length - R] + + # 1-D depth-wise conv + x = self.depthwise_conv(x) # (B, D, U + R) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, U + R) + + utterance = x[:, :, :U] # (B, D, U) + right_context = x[:, :, U:] # (B, D, R) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + new_cache, + ) + + +class EmformerAttention(nn.Module): + r"""Emformer layer attention module. + + Relative positional encoding is applied in this module, which is difference + from https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py # noqa + + Args: + embed_dim (int): + Embedding dimension. + nhead (int): + Number of attention heads in each Emformer layer. + chunk_length (int): + Length of each input chunk. + right_context_length (int): + Length of right context. + 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, + chunk_length: int, + right_context_length: int, + dropout: float = 0.0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + if embed_dim % nhead != 0: + raise ValueError( + f"embed_dim ({embed_dim}) is not a multiple of" + f"nhead ({nhead})." + ) + + self.embed_dim = embed_dim + self.nhead = nhead + self.tanh_on_mem = tanh_on_mem + self.negative_inf = negative_inf + self.head_dim = embed_dim // nhead + self.chunk_length = chunk_length + self.right_context_length = right_context_length + self.dropout = dropout + + self.emb_to_key_value = ScaledLinear( + embed_dim, 2 * embed_dim, bias=True + ) + self.emb_to_query = ScaledLinear(embed_dim, embed_dim, bias=True) + self.out_proj = ScaledLinear( + embed_dim, embed_dim, bias=True, initial_scale=0.25 + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False) + # these two learnable bias are used in matrix c and matrix d + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 # noqa + self.pos_bias_u = nn.Parameter(torch.Tensor(nhead, self.head_dim)) + self.pos_bias_v = nn.Parameter(torch.Tensor(nhead, self.head_dim)) + self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach()) + self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach()) + + self._reset_parameters() + + def _pos_bias_u(self): + return self.pos_bias_u * self.pos_bias_u_scale.exp() + + def _pos_bias_v(self): + return self.pos_bias_v * self.pos_bias_v_scale.exp() + + def _reset_parameters(self) -> None: + nn.init.normal_(self.pos_bias_u, std=0.01) + nn.init.normal_(self.pos_bias_v, std=0.01) + + def _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=self.dropout, training=self.training + ) + return attention_probs + + def _rel_shift(self, x: torch.Tensor) -> torch.Tensor: + """Compute relative positional encoding. + + Args: + x: Input tensor, of shape (B, nhead, U, PE). + U is the length of query vector. + For training and validation mode, + PE = 2 * U + right_context_length - 1. + For inference mode, + PE = tot_left_length + 2 * U + right_context_length - 1, + where tot_left_length = M * chunk_length. + + Returns: + A tensor of shape (B, nhead, U, out_len). + For training and validation mode, out_len = U + right_context_length. + For inference mode, out_len = tot_left_length + U + right_context_length. # noqa + """ + B, nhead, U, PE = x.size() + B_stride = x.stride(0) + nhead_stride = x.stride(1) + U_stride = x.stride(2) + PE_stride = x.stride(3) + out_len = PE - (U - 1) + return x.as_strided( + size=(B, nhead, U, out_len), + stride=(B_stride, nhead_stride, U_stride - PE_stride, PE_stride), + storage_offset=PE_stride * (U - 1), + ) + + def _get_right_context_part( + self, matrix_bd_utterance: torch.Tensor + ) -> torch.Tensor: + """ + Args: + matrix_bd_utterance: + (B * nhead, U, U + right_context_length) + + Returns: + A tensor of shape (B * nhead, U, R), + where R = num_chunks * right_context_length. + """ + assert self.right_context_length > 0 + U = matrix_bd_utterance.size(1) + num_chunks = math.ceil(U / self.chunk_length) + right_context_blocks = [] + for i in range(num_chunks - 1): + start_idx = (i + 1) * self.chunk_length + end_idx = start_idx + self.right_context_length + right_context_blocks.append( + matrix_bd_utterance[:, :, start_idx:end_idx] + ) + right_context_blocks.append( + matrix_bd_utterance[:, :, -self.right_context_length :] + ) + return torch.cat(right_context_blocks, dim=2) + + def _forward_impl( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + pos_emb: torch.Tensor, + left_context_key: Optional[torch.Tensor] = None, + left_context_val: Optional[torch.Tensor] = None, + need_weights=False, + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + torch.Tensor, + torch.Tensor, + torch.Tensor, + torch.Tensor, + ]: + """Underlying chunk-wise attention implementation.""" + U, B, _ = utterance.size() + R = right_context.size(0) + M = memory.size(0) + scaling = float(self.head_dim) ** -0.5 + + # compute query with [right_context, utterance, summary]. + query = self.emb_to_query( + torch.cat([right_context, utterance, summary]) + ) + # compute key and value with [memory, right_context, utterance]. + key, value = self.emb_to_key_value( + torch.cat([memory, right_context, utterance]) + ).chunk(chunks=2, dim=2) + + if left_context_key is not None and left_context_val is not None: + # now compute key and value with + # [memory, right context, left context, uttrance] + # this is used in inference mode + key = torch.cat([key[: M + R], left_context_key, key[M + R :]]) + value = torch.cat( + [value[: M + R], left_context_val, value[M + R :]] + ) + Q = query.size(0) + KV = key.size(0) + + reshaped_key, reshaped_value = [ + tensor.contiguous() + .view(KV, B * self.nhead, self.head_dim) + .transpose(0, 1) + for tensor in [key, value] + ] # both of shape (B * nhead, KV, head_dim) + reshaped_query = ( + query.contiguous().view(Q, B, self.nhead, self.head_dim) * scaling + ) + + # compute attention score + # first, compute attention matrix a and matrix c + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 # noqa + query_with_bais_u = ( + (reshaped_query + self._pos_bias_u()) + .view(Q, B * self.nhead, self.head_dim) + .transpose(0, 1) + ) # (B * nhead, Q, head_dim) + matrix_ac = torch.bmm( + query_with_bais_u, reshaped_key.transpose(1, 2) + ) # (B * nhead, Q, KV) + + # second, compute attention matrix b and matrix d + # relative positional encoding is applied on the part of attention + # between chunk (in query) and itself as well as its left context + # (in key) + utterance_with_bais_v = ( + reshaped_query[R : R + U] + self._pos_bias_v() + ).permute(1, 2, 0, 3) + # (B, nhead, U, head_dim) + PE = pos_emb.size(0) + if left_context_key is not None and left_context_val is not None: + # inference mode + L = left_context_key.size(0) + tot_left_length = M * self.chunk_length if M > 0 else L + assert tot_left_length >= L + assert PE == tot_left_length + 2 * U + self.right_context_length - 1 + else: + # training and validation mode + assert PE == 2 * U + self.right_context_length - 1 + pos_emb = ( + self.linear_pos(pos_emb) + .view(PE, self.nhead, self.head_dim) + .transpose(0, 1) + .unsqueeze(0) + ) # (1, nhead, PE, head_dim) + matrix_bd_utterance = torch.matmul( + utterance_with_bais_v, pos_emb.transpose(-2, -1) + ) # (B, nhead, U, PE) + # rel-shift operation + matrix_bd_utterance = self._rel_shift(matrix_bd_utterance) + # (B, nhead, U, U + right_context_length) for training and validation mode; # noqa + # (B, nhead, U, tot_left_length + U + right_context_length) for inference mode. # noqa + matrix_bd_utterance = matrix_bd_utterance.contiguous().view( + B * self.nhead, U, -1 + ) + matrix_bd = torch.zeros_like(matrix_ac) + if left_context_key is not None and left_context_val is not None: + # inference mode + # key: [memory, right context, left context, utterance] + # for memory + if M > 0: + # take average over the chunk frames for the memory vector + matrix_bd[:, R : R + U, :M] = torch.nn.functional.avg_pool2d( + matrix_bd_utterance[:, :, :tot_left_length].unsqueeze(1), + kernel_size=(1, self.chunk_length), + stride=(1, self.chunk_length), + ).squeeze(1) + # for right_context + if R > 0: + matrix_bd[:, R : R + U, M : M + R] = matrix_bd_utterance[ + :, :, tot_left_length + U : + ] + # for left_context and utterance + matrix_bd[:, R : R + U, M + R :] = matrix_bd_utterance[ + :, :, tot_left_length - L : tot_left_length + U + ] + else: + # training and validation mode + # key: [memory, right context, utterance] + # for memory + if M > 0: + # take average over the chunk frames for the memory vector + matrix_bd[:, R : R + U, :M] = torch.nn.functional.avg_pool2d( + matrix_bd_utterance[:, :, :U].unsqueeze(1), + kernel_size=(1, self.chunk_length), + stride=(1, self.chunk_length), + ceil_mode=True, + ).squeeze(1)[:, :, :-1] + # for right_context + if R > 0: + matrix_bd[ + :, R : R + U, M : M + R + ] = self._get_right_context_part(matrix_bd_utterance) + # for utterance + matrix_bd[:, R : R + U, M + R :] = matrix_bd_utterance[:, :, :U] + + attention_weights = matrix_ac + matrix_bd + + # compute padding mask + if B == 1: + padding_mask = None + else: + 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 outputs + attention = torch.bmm(attention_probs, reshaped_value) + assert attention.shape == (B * self.nhead, Q, self.head_dim) + attention = ( + attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim) + ) + + # apply output projection + outputs = self.out_proj(attention) + + output_right_context_utterance = outputs[: R + U] + output_memory = outputs[R + U :] + if self.tanh_on_mem: + output_memory = torch.tanh(output_memory) + else: + output_memory = torch.clamp(output_memory, min=-10, max=10) + + if need_weights: + # average over attention heads + attention_probs = attention_probs.reshape(B, self.nhead, Q, KV) + attention_probs = attention_probs.sum(dim=1) / self.nhead + probs_memory = attention_probs[:, R : R + U, :M].sum(dim=2) + probs_frames = attention_probs[:, R : R + U, M:].sum(dim=2) + return ( + output_right_context_utterance, + output_memory, + key, + value, + probs_memory, + probs_frames, + ) + + return ( + output_right_context_utterance, + output_memory, + key, + value, + None, + None, + ) + + def forward( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + pos_emb: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + # TODO: Modify docs. + """Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of the hard-copied right contexts; + U: length of full utterance; + S: length of summary vectors; + M: length of memory vectors. + + It computes a `big` attention matrix on full utterance and + then utilizes a pre-computed mask to simulate chunk-wise attention. + + It concatenates three blocks: hard-copied right contexts, + full utterance, and summary vectors, as a `big` block, + to compute the query tensor: + query = [right_context, utterance, summary], + with length Q = R + U + S. + It concatenates the three blocks: memory vectors, + hard-copied right contexts, and full utterance as another `big` block, + to compute the key and value tensors: + key & value = [memory, right_context, utterance], + with length KV = M + R + U. + Attention scores is computed with above `big` query and key. + + Then the underlying chunk-wise attention is obtained by applying + the attention mask. Suppose + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + s_i: summary vector of c_i; + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); + s_i (in query) -> l_i, c_i, r_i (in key). + + Relative positional encoding is applied on the part of attention between + [utterance] (in query) and [memory, right_context, utterance] (in key). + Actually, it is applied on the part of attention between each chunk + (in query) and itself, its memory vectors, left context, and right + context (in key), after applying the mask: + c_i (in query) -> l_i, c_i, r_i, m_i (in key). + + Args: + utterance (torch.Tensor): + Full 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): + Hard-copied right context frames, with shape (R, B, D), + where R = num_chunks * right_context_length + summary (torch.Tensor): + Summary elements with shape (S, B, D), where S = num_chunks. + It is an empty tensor without using memory. + memory (torch.Tensor): + Memory elements, with shape (M, B, D), where M = num_chunks - 1. + It is an empty tensor without using memory. + attention_mask (torch.Tensor): + Pre-computed attention mask to simulate underlying chunk-wise + attention, with shape (Q, KV). + pos_emb (torch.Tensor): + Position encoding embedding, with shape (PE, D). + where PE = 2 * U + right_context_length - 1. + + 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. + - summary of attention weights on memory, with shape (B, U). + - summary of attention weights on left context, utterance, and + right context, with shape (B, U). + """ + ( + output_right_context_utterance, + output_memory, + _, + _, + probs_memory, + probs_frames, + ) = self._forward_impl( + utterance, + lengths, + right_context, + summary, + memory, + attention_mask, + pos_emb, + need_weights=True, + ) + return ( + output_right_context_utterance, + output_memory[:-1], + probs_memory, + probs_frames, + ) + + 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, + pos_emb: 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, i.e., current chunk; + L: length of cached left context; + S: length of summary vectors, S = 1; + M: length of cached memory vectors. + + It concatenates the right context, utterance (i.e., current chunk) + and summary vector of current chunk, to compute the query tensor: + query = [right_context, utterance, summary], + with length Q = R + U + S. + It concatenates the memory vectors, right context, left context, and + current chunk, to compute the key and value tensors: + key & value = [memory, right_context, left_context, utterance], + with length KV = M + R + L + U. + + The chunk-wise attention is: + chunk, right context (in query) -> + left context, chunk, right context, memory vectors (in key); + summary (in query) -> left context, chunk, right context (in key). + + Relative positional encoding is applied on the part of attention: + chunk (in query) -> + left context, chunk, right context, memory vectors (in key); + + Args: + utterance (torch.Tensor): + Current chunk frames, with shape (U, B, D), where U = chunk_length. + 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), + where R = right_context_length. + summary (torch.Tensor): + Summary vector with shape (1, B, D), or empty tensor. + memory (torch.Tensor): + Memory vectors, with shape (M, B, D), or empty tensor. + left_context_key (torch,Tensor): + Cached attention key of left context from preceding computation, + with shape (L, B, D), where L <= left_context_length. + left_context_val (torch.Tensor): + Cached attention value of left context from preceding computation, + with shape (L, B, D), where L <= left_context_length. + pos_emb (torch.Tensor): + Position encoding embedding, with shape (PE, D), + where PE = M * chunk_length + 2 * U - 1 if M > 0 else L + 2 * U - 1. + + Returns: + A tuple containing 4 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (1, B, D) or (0, B, D). + - attention key of left context and utterance, which would be cached + for next computation, with shape (L + U, B, D). + - attention value of left context and utterance, which would be + cached for next computation, with shape (L + U, B, D). + """ + U = utterance.size(0) + R = right_context.size(0) + L = left_context_key.size(0) + S = summary.size(0) + M = memory.size(0) + + # query = [right context, utterance, summary] + Q = R + U + S + # key, value = [memory, right context, left context, uttrance] + KV = M + R + L + U + attention_mask = torch.zeros(Q, KV).to( + dtype=torch.bool, device=utterance.device + ) + # disallow attention bettween the summary vector with the memory bank + attention_mask[-1, :M] = True + ( + output_right_context_utterance, + output_memory, + key, + value, + _, + _, + ) = self._forward_impl( + utterance, + lengths, + right_context, + summary, + memory, + attention_mask, + pos_emb, + left_context_key=left_context_key, + left_context_val=left_context_val, + ) + return ( + output_right_context_utterance, + output_memory, + key[M + R :], + value[M + R :], + ) + + +class EmformerEncoderLayer(nn.Module): + """Emformer layer that constitutes Emformer. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads. + dim_feedforward (int): + Hidden layer dimension of feedforward network. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (Default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (Default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (Default: 0) + right_context_length (int, optional): + Length of right context. (Default: 0) + 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 = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + 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.attention = EmformerAttention( + embed_dim=d_model, + nhead=nhead, + chunk_length=chunk_length, + right_context_length=right_context_length, + dropout=dropout, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + self.summary_op = nn.AvgPool1d( + kernel_size=chunk_length, stride=chunk_length, ceil_mode=True + ) + + self.feed_forward_macaron = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.feed_forward = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.conv_module = ConvolutionModule( + chunk_length, + right_context_length, + d_model, + cnn_module_kernel, + ) + + self.norm_final = BasicNorm(d_model) + + # try to ensure the output is close to zero-mean + # (or at least, zero-median). + self.balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0 + ) + + self.dropout = nn.Dropout(dropout) + + self.layer_dropout = layer_dropout + self.left_context_length = left_context_length + self.chunk_length = chunk_length + self.right_context_length = right_context_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_conv_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + ) -> torch.Tensor: + """Apply convolution module in training and validation mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context, _ = self.conv_module(utterance, right_context) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance + + def _apply_conv_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + conv_cache: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Apply convolution module on utterance in inference mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context, conv_cache = self.conv_module.infer( + utterance, right_context, conv_cache + ) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance, conv_cache + + def _apply_attention_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + lengths: torch.Tensor, + memory: torch.Tensor, + pos_emb: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply attention module in training and validation mode.""" + if attention_mask is None: + raise ValueError( + "attention_mask must be not None in training or validation mode." # noqa + ) + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + if self.use_memory: + summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + ) + else: + summary = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + output_right_context_utterance, output_memory, _, _ = self.attention( + utterance=utterance, + lengths=lengths, + right_context=right_context, + summary=summary, + memory=memory, + attention_mask=attention_mask, + pos_emb=pos_emb, + ) + + return output_right_context_utterance, output_memory + + def _apply_attention_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + lengths: torch.Tensor, + memory: torch.Tensor, + pos_emb: torch.Tensor, + state: Optional[List[torch.Tensor]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]: + """Apply attention module in inference mode. + 1) Unpack cached states including: + - memory from previous chunks in the lower layer; + - attention key and value of left context from 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. + """ + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + 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 + ) + U = utterance.size(0) + # pos_emb is of shape [PE, D], where PE = M * chunk_length + 2 * U - 1, + # for query of [utterance] (i), key-value [memory vectors, left context, utterance, right context] (j) # noqa + # the max relative distance i - j is M * chunk_length + U - 1 + # the min relative distance i - j is -(U + right_context_length - 1) + M = pre_memory.size(0) # M <= max_memory_size + if self.max_memory_size > 0: + PE = M * self.chunk_length + 2 * U + self.right_context_length - 1 + tot_PE = ( + self.max_memory_size * self.chunk_length + + 2 * U + + self.right_context_length + - 1 + ) + else: + L = left_context_key.size(0) + PE = L + 2 * U + self.right_context_length - 1 + tot_PE = ( + self.left_context_length + 2 * U + self.right_context_length - 1 + ) + assert pos_emb.size(0) == tot_PE + pos_emb = pos_emb[tot_PE - PE :] + ( + 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, + pos_emb=pos_emb, + ) + state = self._pack_state( + next_key, next_val, utterance.size(0), memory, state + ) + return output_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, + pos_emb: torch.Tensor, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + r"""Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of hard-copied right contexts; + U: length of full utterance; + M: length of memory vectors. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + 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). + It is an empty tensor without using memory. + attention_mask (torch.Tensor): + Attention mask for underlying attention module, + with shape (Q, KV), where Q = R + U + S, KV = M + R + U. + pos_emb (torch.Tensor): + Position encoding embedding, with shape (PE, D). + For training mode, P = 2*U-1. + + Returns: + A tuple containing 3 tensors: + - output utterance, with shape (U, B, D). + - output right context, with shape (R, B, D). + - output memory, with shape (M, B, D). + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + src_orig = src + + warmup_scale = min(0.1 + warmup, 1.0) + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + if self.training: + alpha = ( + warmup_scale + if torch.rand(()).item() <= (1.0 - self.layer_dropout) + else 0.1 + ) + else: + alpha = 1.0 + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + src_att, output_memory = self._apply_attention_module_forward( + src, R, lengths, memory, pos_emb, attention_mask + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv = self._apply_conv_module_forward(src, R) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + output_utterance = src[R:] + output_right_context = src[:R] + return output_utterance, output_right_context, output_memory + + def infer( + self, + utterance: torch.Tensor, + lengths: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + pos_emb: torch.Tensor, + state: Optional[List[torch.Tensor]] = None, + conv_cache: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor], torch.Tensor]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + M: length of memory. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + 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) + pos_emb (torch.Tensor): + Position encoding embedding, with shape (PE, D). + For infer mode, PE = L+2*U-1. + conv_cache (torch.Tensor, optional): + Cache tensor of left context for causal convolution. + + Returns: + (Tensor, Tensor, List[torch.Tensor], Tensor): + - output utterance, with shape (U, B, D); + - output right_context, with shape (R, B, D); + - output memory, with shape (1, B, D) or (0, B, D). + - output state. + - updated conv_cache. + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + ( + src_att, + output_memory, + output_state, + ) = self._apply_attention_module_infer( + src, R, lengths, memory, pos_emb, state + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv, conv_cache = self._apply_conv_module_infer(src, R, conv_cache) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + output_utterance = src[R:] + output_right_context = src[:R] + return ( + output_utterance, + output_right_context, + output_memory, + output_state, + conv_cache, + ) + + +def _gen_attention_mask_block( + col_widths: List[int], + col_mask: List[bool], + num_rows: int, + device: torch.device, +) -> torch.Tensor: + assert len(col_widths) == len( + col_mask + ), "Length of col_widths must match that of col_mask" + + mask_block = [ + torch.ones(num_rows, col_width, device=device) + if is_ones_col + else torch.zeros(num_rows, col_width, device=device) + for col_width, is_ones_col in zip(col_widths, col_mask) + ] + return torch.cat(mask_block, dim=1) + + +class EmformerEncoder(nn.Module): + """Implements the Emformer architecture introduced in + *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency + Streaming Speech Recognition* + [:footcite:`shi2021emformer`]. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads in each emformer layer. + dim_feedforward (int): + Hidden layer dimension of each emformer layer's feedforward network. + num_encoder_layers (int): + Number of emformer layers to instantiate. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (default: 0) + right_context_length (int, optional): + Length of right context. (default: 0) + 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, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + 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( + [ + EmformerEncoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + chunk_length=chunk_length, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length, + right_context_length=right_context_length, + 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.encoder_pos = RelPositionalEncoding(d_model, dropout) + + 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_chunks = math.ceil( + (T - self.right_context_length) / self.chunk_length + ) + right_context_blocks = [] + for seg_idx in range(num_chunks - 1): + start = (seg_idx + 1) * self.chunk_length + end = start + self.right_context_length + right_context_blocks.append(x[start:end]) + right_context_blocks.append(x[T - self.right_context_length :]) + 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 to simulate underlying chunk-wise attention + computation, where chunk-wise connections are filled with `False`, + and other unnecessary connections beyond chunk are filled with `True`. + + R: length of hard-copied right contexts; + U: length of full utterance; + S: length of summary vectors; + M: length of memory vectors; + Q: length of attention query; + KV: length of attention key and value. + + The shape of attention mask is (Q, KV). + If self.use_memory is `True`: + query = [right_context, utterance, summary]; + key, value = [memory, right_context, utterance]; + Q = R + U + S, KV = M + R + U. + Otherwise: + query = [right_context, utterance] + key, value = [right_context, utterance] + Q = R + U, KV = R + U. + + Suppose: + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + s_i: summary vector of c_i. + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); + s_i (in query) -> l_i, c_i, r_i (in key). + """ + U = utterance.size(0) + num_chunks = math.ceil(U / self.chunk_length) + + right_context_mask = [] + utterance_mask = [] + summary_mask = [] + + if self.use_memory: + num_cols = 9 + # right context and utterance both attend to memory, right context, + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4, 7] for idx in range(num_cols) + ] + # summary attends to right context, utterance + summary_cols_mask = [idx in [4, 7] for idx in range(num_cols)] + masks_to_concat = [right_context_mask, utterance_mask, summary_mask] + else: + num_cols = 6 + # right context and utterance both attend to right context and + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4] for idx in range(num_cols) + ] + summary_cols_mask = None + masks_to_concat = [right_context_mask, utterance_mask] + + for chunk_idx in range(num_chunks): + col_widths = self._gen_attention_mask_col_widths(chunk_idx, U) + + right_context_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + self.right_context_length, + utterance.device, + ) + right_context_mask.append(right_context_mask_block) + + utterance_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + min( + self.chunk_length, + U - chunk_idx * self.chunk_length, + ), + utterance.device, + ) + utterance_mask.append(utterance_mask_block) + + if summary_cols_mask is not None: + summary_mask_block = _gen_attention_mask_block( + col_widths, summary_cols_mask, 1, utterance.device + ) + summary_mask.append(summary_mask_block) + + attention_mask = ( + 1 - torch.cat([torch.cat(mask) for mask in masks_to_concat]) + ).to(torch.bool) + return attention_mask + + def forward( + self, x: torch.Tensor, lengths: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and validation mode. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + + Returns: + A tuple of 2 tensors: + - output utterance frames, with shape (U, B, D). + - output_lengths, with shape (B,), without containing the + right_context at the end. + """ + U = x.size(0) - self.right_context_length + x, pos_emb = self.encoder_pos( + x, pos_len=U, neg_len=U + self.right_context_length + ) + + right_context = self._gen_right_context(x) + utterance = x[:U] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + attention_mask = self._gen_attention_mask(utterance) + memory = ( + self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[ + :-1 + ] + if self.use_memory + else torch.empty(0).to(dtype=x.dtype, device=x.device) + ) + + output = utterance + for layer in self.emformer_layers: + output, right_context, memory = layer( + output, + output_lengths, + right_context, + memory, + attention_mask, + pos_emb, + warmup=warmup, + ) + + return output, output_lengths + + def infer( + self, + x: torch.Tensor, + lengths: torch.Tensor, + states: Optional[List[List[torch.Tensor]]] = None, + conv_caches: Optional[List[torch.Tensor]] = None, + ) -> Tuple[ + torch.Tensor, torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor] + ]: + """Forward pass for streaming inference. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + states (List[List[torch.Tensor]], optional): + Cached states from proceeding chunk's computation, where each + element (List[torch.Tensor]) corresponds to each emformer layer. + (default: None) + conv_caches (List[torch.Tensor], optional): + Cached tensors of left context for causal convolution, where each + element (Tensor) corresponds to each convolutional layer. + + Returns: + (Tensor, Tensor, List[List[torch.Tensor]], List[torch.Tensor]): + - output utterance frames, with shape (U, B, D). + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + - updated convolution caches from current chunk. + """ + 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)}." + ) + + pos_len = ( + self.max_memory_size * self.chunk_length + self.chunk_length + if self.max_memory_size > 0 + else self.left_context_length + self.chunk_length + ) + neg_len = self.chunk_length + self.right_context_length + x, pos_emb = self.encoder_pos(x, pos_len=pos_len, neg_len=neg_len) + + 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]] = [] + output_conv_caches: List[torch.Tensor] = [] + for layer_idx, layer in enumerate(self.emformer_layers): + ( + output, + right_context, + memory, + output_state, + output_conv_cache, + ) = layer.infer( + output, + output_lengths, + right_context, + memory, + pos_emb, + None if states is None else states[layer_idx], + None if conv_caches is None else conv_caches[layer_idx], + ) + output_states.append(output_state) + output_conv_caches.append(output_conv_cache) + + return output, output_lengths, output_states, output_conv_caches + + +class Emformer(EncoderInterface): + def __init__( + self, + num_features: int, + chunk_length: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 3, + left_context_length: int = 0, + right_context_length: int = 0, + 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." + ) + if ( + max_memory_size > 0 + and max_memory_size * chunk_length < left_context_length + ): + raise NotImplementedError( + "max_memory_size * chunk_length can not be less than left_context_length" # noqa + ) + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder = EmformerEncoder( + chunk_length=chunk_length // 4, + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + num_encoder_layers=num_encoder_layers, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length // 4, + right_context_length=right_context_length // 4, + max_memory_size=max_memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and non-streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + x_lens (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + """ + x = self.encoder_embed(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # 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, warmup=warmup + ) # (T, N, C) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + states: Optional[List[List[torch.Tensor]]] = None, + conv_caches: Optional[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]) corresponds to each emformer layer. + (default: None) + conv_caches (List[torch.Tensor], optional): + Cached tensors of left context for causal convolution, where each + element (Tensor) corresponds to each convolutional layer. + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + - updated convolution caches from current chunk. + """ + 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, + output_conv_caches, + ) = self.encoder.infer(x, x_lens, states, conv_caches) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths, output_states, output_conv_caches + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/4 length). + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = ((T-1)//2 - 1)//2, which approximates T' == T//4 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + layer1_channels: int = 8, + layer2_channels: int = 32, + layer3_channels: int = 128, + ) -> None: + """ + Args: + in_channels: + Number of channels in. The input shape is (N, T, in_channels). + Caution: It requires: T >=7, in_channels >=7 + out_channels + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + """ + assert in_channels >= 7 + super().__init__() + + self.conv = nn.Sequential( + ScaledConv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=1, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ) + self.out = ScaledLinear( + layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels + ) + # set learn_eps=False because out_norm is preceded by `out`, and `out` + # itself has learned scale, so the extra degree of freedom is not + # needed. + self.out_norm = BasicNorm(out_channels, learn_eps=False) + # constrain median of output to be close to zero. + self.out_balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55 + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + x = self.conv(x) + # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + x = self.out_norm(x) + x = self.out_balancer(x) + return x diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py new file mode 120000 index 000000000..b9aa0ae08 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/encoder_interface.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py new file mode 120000 index 000000000..815fd4bb6 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/joiner.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py new file mode 120000 index 000000000..ebb6d774d --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/model.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py new file mode 120000 index 000000000..e2deb4492 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/optim.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py new file mode 120000 index 000000000..09d802cc4 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/scaling.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py new file mode 100644 index 000000000..57ad3b4ec --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/test_emformer.py @@ -0,0 +1,585 @@ +import torch + + +def test_rel_positional_encoding(): + from emformer import RelPositionalEncoding + + D = 256 + pos_enc = RelPositionalEncoding(D, dropout=0.1) + pos_len = 100 + neg_len = 100 + x = torch.randn(2, D) + x, pos_emb = pos_enc(x, pos_len, neg_len) + assert pos_emb.shape == (pos_len + neg_len - 1, D) + + +def test_emformer_attention_forward(): + from emformer import EmformerAttention + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 3 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + attention = EmformerAttention( + embed_dim=D, + nhead=8, + chunk_length=chunk_length, + right_context_length=right_context_length, + ) + + for use_memory in [True, False]: + if use_memory: + S = num_chunks + 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 + PE = 2 * U + right_context_length - 1 + pos_emb = torch.randn(PE, D) + + ( + output_right_context_utterance, + output_memory, + probs_memory, + probs_frames, + ) = attention( + utterance, + lengths, + right_context, + summary, + memory, + attention_mask, + pos_emb, + ) + assert output_right_context_utterance.shape == (R + U, B, D) + assert output_memory.shape == (M, B, D) + assert probs_memory.shape == (B, U) + assert probs_frames.shape == (B, U) + + +def test_emformer_attention_infer(): + from emformer import EmformerAttention + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 1 + U = chunk_length * num_chunks + R = right_context_length * num_chunks + L = 3 + attention = EmformerAttention( + embed_dim=D, + nhead=8, + chunk_length=chunk_length, + right_context_length=right_context_length, + ) + + 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) + PE = ( + 2 * U + + right_context_length + - 1 + + (M * chunk_length if M > 0 else L) + ) + pos_emb = torch.randn(PE, 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, + pos_emb, + ) + 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_convolution_module_forward(): + from emformer import ConvolutionModule + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 3 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + cache = torch.randn(B, D, kernel_size - 1) + + utterance, right_context, new_cache = conv_module( + utterance, right_context, cache + ) + assert utterance.shape == (U, B, D) + assert right_context.shape == (R, B, D) + assert new_cache.shape == (B, D, kernel_size - 1) + + +def test_convolution_module_infer(): + from emformer import ConvolutionModule + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 1 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + cache = torch.randn(B, D, kernel_size - 1) + + utterance, right_context, new_cache = conv_module.infer( + utterance, right_context, cache + ) + assert utterance.shape == (U, B, D) + assert right_context.shape == (R, B, D) + assert new_cache.shape == (B, D, kernel_size - 1) + + +def test_emformer_encoder_layer_forward(): + from emformer import EmformerEncoderLayer + + B, D = 2, 256 + chunk_length = 8 + right_context_length = 2 + left_context_length = 8 + kernel_size = 31 + num_chunks = 3 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + + for use_memory in [True, False]: + if use_memory: + S = num_chunks + M = S - 1 + else: + S, M = 0, 0 + + layer = EmformerEncoderLayer( + d_model=D, + nhead=8, + dim_feedforward=1024, + chunk_length=chunk_length, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + 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 + PE = 2 * U + right_context_length - 1 + pos_emb = torch.randn(PE, D) + + output_utterance, output_right_context, output_memory = layer( + utterance, + lengths, + right_context, + memory, + attention_mask, + pos_emb, + ) + 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_encoder_layer_infer(): + from emformer import EmformerEncoderLayer + + B, D = 2, 256 + chunk_length = 8 + right_context_length = 2 + left_context_length = 8 + kernel_size = 31 + num_chunks = 1 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + + for use_memory in [True, False]: + if use_memory: + max_memory_size = 3 + M = 1 + else: + max_memory_size = 0 + M = 0 + + layer = EmformerEncoderLayer( + d_model=D, + nhead=8, + dim_feedforward=1024, + chunk_length=chunk_length, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + max_memory_size=max_memory_size, + ) + + 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 + PE = ( + 2 * U + + right_context_length + - 1 + + ( + max_memory_size * chunk_length + if max_memory_size > 0 + else left_context_length + ) + ) + pos_emb = torch.randn(PE, D) + conv_cache = None + ( + output_utterance, + output_right_context, + output_memory, + output_state, + conv_cache, + ) = layer.infer( + utterance, + lengths, + right_context, + memory, + pos_emb, + state, + conv_cache, + ) + 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 == (max_memory_size, B, D) + assert output_state[1].shape == (left_context_length, B, D) + assert output_state[2].shape == (left_context_length, B, D) + assert output_state[3].shape == (1, B) + assert conv_cache.shape == (B, D, kernel_size - 1) + + +def test_emformer_encoder_forward(): + from emformer import EmformerEncoder + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + left_context_length = 2 + num_chunks = 3 + U = num_chunks * chunk_length + kernel_size = 31 + num_encoder_layers = 2 + + for use_memory in [True, False]: + if use_memory: + S = num_chunks + M = S - 1 + else: + S, M = 0, 0 + + encoder = EmformerEncoder( + chunk_length=chunk_length, + d_model=D, + dim_feedforward=1024, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + max_memory_size=M, + ) + + x = torch.randn(U + right_context_length, B, D) + lengths = torch.randint(1, U + right_context_length + 1, (B,)) + lengths[0] = U + right_context_length + + output, output_lengths = encoder(x, lengths) + assert output.shape == (U, B, D) + assert torch.equal( + output_lengths, torch.clamp(lengths - right_context_length, min=0) + ) + + +def test_emformer_encoder_infer(): + from emformer import EmformerEncoder + + B, D = 2, 256 + num_encoder_layers = 2 + chunk_length = 4 + right_context_length = 2 + left_context_length = 2 + num_chunks = 3 + kernel_size = 31 + + for use_memory in [True, False]: + if use_memory: + max_memory_size = 3 + else: + max_memory_size = 0 + + encoder = EmformerEncoder( + chunk_length=chunk_length, + d_model=D, + dim_feedforward=1024, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + max_memory_size=max_memory_size, + ) + + states = None + conv_caches = None + for chunk_idx in range(num_chunks): + x = torch.randn(chunk_length + right_context_length, B, D) + lengths = torch.randint( + 1, chunk_length + right_context_length + 1, (B,) + ) + lengths[0] = chunk_length + right_context_length + output, output_lengths, states, conv_caches = encoder.infer( + x, lengths, states, conv_caches + ) + assert output.shape == (chunk_length, B, D) + assert torch.equal( + output_lengths, + torch.clamp(lengths - right_context_length, min=0), + ) + assert len(states) == num_encoder_layers + for state in states: + assert len(state) == 4 + assert state[0].shape == (max_memory_size, B, D) + assert state[1].shape == (left_context_length, B, D) + assert state[2].shape == (left_context_length, B, D) + assert torch.equal( + state[3], + (chunk_idx + 1) * chunk_length * torch.ones_like(state[3]), + ) + for conv_cache in conv_caches: + assert conv_cache.shape == (B, D, kernel_size - 1) + + +def test_emformer_encoder_forward_infer_consistency(): + from emformer import EmformerEncoder + + chunk_length = 4 + num_chunks = 3 + U = chunk_length * num_chunks + left_context_length, right_context_length = 1, 2 + D = 256 + num_encoder_layers = 3 + kernel_size = 31 + memory_sizes = [0, 3] + + for max_memory_size in memory_sizes: + encoder = EmformerEncoder( + chunk_length=chunk_length, + d_model=D, + dim_feedforward=1024, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + max_memory_size=max_memory_size, + ) + encoder.eval() + + x = torch.randn(U + right_context_length, 1, D) + lengths = torch.tensor([U + right_context_length]) + + # training mode with full utterance + forward_output, forward_output_lengths = encoder(x, lengths) + + # streaming inference mode with individual chunks + states = None + conv_caches = None + for chunk_idx in range(num_chunks): + start_idx = chunk_idx * chunk_length + end_idx = start_idx + chunk_length + chunk = x[start_idx : end_idx + right_context_length] # noqa + chunk_length = torch.tensor([chunk_length]) + ( + infer_output_chunk, + infer_output_lengths, + states, + conv_caches, + ) = encoder.infer(chunk, chunk_length, states, conv_caches) + forward_output_chunk = forward_output[start_idx:end_idx] + assert torch.allclose( + infer_output_chunk, + forward_output_chunk, + atol=1e-4, + rtol=0.0, + ), ( + infer_output_chunk - forward_output_chunk + ) + + +def test_emformer_forward(): + from emformer import Emformer + + num_features = 80 + chunk_length = 16 + right_context_length = 8 + left_context_length = 8 + num_chunks = 3 + U = num_chunks * chunk_length + B, D = 2, 256 + kernel_size = 31 + + for use_memory in [True, False]: + if use_memory: + max_memory_size = 3 + else: + max_memory_size = 0 + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=D, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + max_memory_size=max_memory_size, + ) + x = torch.randn(B, U + right_context_length + 3, num_features) + x_lens = torch.randint(1, U + right_context_length + 3 + 1, (B,)) + x_lens[0] = U + right_context_length + 3 + output, output_lengths = model(x, x_lens) + assert output.shape == (B, U // 4, D) + assert torch.equal( + output_lengths, + torch.clamp( + ((x_lens - 1) // 2 - 1) // 2 - right_context_length // 4, min=0 + ), + ) + + +def test_emformer_infer(): + from emformer import Emformer + + num_features = 80 + chunk_length = 8 + U = chunk_length + left_context_length, right_context_length = 32, 4 + B, D = 2, 256 + num_chunks = 3 + num_encoder_layers = 2 + kernel_size = 31 + + for use_memory in [True, False]: + if use_memory: + max_memory_size = 32 + else: + max_memory_size = 0 + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=D, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + max_memory_size=max_memory_size, + ) + states = None + conv_caches = None + for chunk_idx in range(num_chunks): + x = torch.randn(B, U + right_context_length + 3, num_features) + x_lens = torch.randint(1, U + right_context_length + 3 + 1, (B,)) + x_lens[0] = U + right_context_length + 3 + output, output_lengths, states, conv_caches = model.infer( + x, x_lens, states, conv_caches + ) + assert output.shape == (B, U // 4, D) + assert torch.equal( + output_lengths, + torch.clamp( + ((x_lens - 1) // 2 - 1) // 2 - right_context_length // 4, + min=0, + ), + ) + assert len(states) == num_encoder_layers + for state in states: + assert len(state) == 4 + assert state[0].shape == (max_memory_size, B, D) + assert state[1].shape == (left_context_length // 4, B, D) + assert state[2].shape == (left_context_length // 4, B, D) + assert torch.equal( + state[3], + U // 4 * (chunk_idx + 1) * torch.ones_like(state[3]), + ) + for conv_cache in conv_caches: + assert conv_cache.shape == (B, D, kernel_size - 1) + + +if __name__ == "__main__": + test_rel_positional_encoding() + test_emformer_attention_forward() + test_emformer_attention_infer() + test_convolution_module_forward() + test_convolution_module_infer() + test_emformer_encoder_layer_forward() + test_emformer_encoder_layer_infer() + test_emformer_encoder_forward() + test_emformer_encoder_infer() + test_emformer_encoder_forward_infer_consistency() + test_emformer_forward() + test_emformer_infer() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py new file mode 100755 index 000000000..eef4b0f49 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py @@ -0,0 +1,1121 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless4/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless2/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless4/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless2/exp \ + --full-libri 1 \ + --max-duration 550 + +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from emformer import Emformer +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, Eve +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--encoder-dim", + type=int, + default=512, + help="Attention dim for the Emformer", + ) + + parser.add_argument( + "--nhead", + type=int, + default=8, + help="Number of attention heads for the Emformer", + ) + + parser.add_argument( + "--dim-feedforward", + type=int, + default=2048, + help="Feed-forward dimension for the Emformer", + ) + + parser.add_argument( + "--num-encoder-layers", + type=int, + default=12, + help="Number of encoder layers for the Emformer", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=int, + default=31, + help="Kernel size for the convolution module.", + ) + + parser.add_argument( + "--left-context-length", + type=int, + default=32, + help="Number of frames for the left context in the Emformer", + ) + + parser.add_argument( + "--chunk-length", + type=int, + default=32, + help="Number of frames for each chuunk in the Emformer", + ) + + parser.add_argument( + "--right-context-length", + type=int, + default=8, + 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=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--initial-lr", + type=float, + default=0.003, + help="""The initial learning rate. This value should not need to be + changed.""", + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate decreases. + We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" + "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=8000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=20, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for Emformer + "feature_dim": 80, + "subsampling_factor": 4, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Emformer( + num_features=params.feature_dim, + chunk_length=params.chunk_length, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + cnn_module_kernel=params.cnn_module_kernel, + left_context_length=params.left_context_length, + right_context_length=params.right_context_length, + max_memory_size=params.memory_size, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = ( + model.device + if isinstance(model, DDP) + else next(model.parameters()).device + ) + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = ( + (feature_lens // params.subsampling_factor).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}" + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + + if batch_idx > 0 and batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + if params.full_libri is False: + params.valid_interval = 1600 + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank]) + + optimizer = Eve(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2 ** 22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + # warmup = 0.0 is so that the derivs for the pruned loss stay zero + # (i.e. are not remembered by the decaying-average in adam), because + # we want to avoid these params being subject to shrinkage in adam. + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=0.0, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + except RuntimeError as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main()