diff --git a/egs/librispeech/ASR/zipformer_ctc_attn/attention_decoder.py b/egs/librispeech/ASR/zipformer_ctc_attn/attention_decoder.py index 2cf3fb83e..a215293d3 100644 --- a/egs/librispeech/ASR/zipformer_ctc_attn/attention_decoder.py +++ b/egs/librispeech/ASR/zipformer_ctc_attn/attention_decoder.py @@ -1,5 +1,5 @@ #!/usr/bin/env python3 -# Copyright 2022 Xiaomi Corp. (authors: Zengwei Yao) +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey, Zengwei Yao) # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -47,8 +47,7 @@ class AttentionDecoderModel(nn.Module): """ Args: vocab_size (int): Number of classes. - encoder_dim (int): - d_model: (int,int): embedding dimension of 2 encoder stacks + decoder_dim: (int,int): embedding dimension of 2 encoder stacks attention_dim: (int,int): attention dimension of 2 encoder stacks nhead (int, int): number of heads dim_feedforward (int, int): feedforward dimension in 2 encoder stacks @@ -62,15 +61,15 @@ class AttentionDecoderModel(nn.Module): def __init__( self, vocab_size: int, - d_model: int, + decoder_dim: int, unmasked_dim: int, num_decoder_layers: int, attention_dim: int, nhead: int, feedforward_dim: int, - dropout: float, sos_id: int, eos_id: int, + dropout: float = 0.1, ignore_id: int = -1, warmup_batches: float = 4000.0, label_smoothing: float = 0.1, @@ -84,7 +83,7 @@ class AttentionDecoderModel(nn.Module): # layer learn something. Then we start to warm up the other encoders. self.decoder = TransformerDecoder( vocab_size, - d_model, + decoder_dim, unmasked_dim, num_decoder_layers, attention_dim, @@ -103,7 +102,6 @@ class AttentionDecoderModel(nn.Module): def _pre_ys_in_out(self, token_ids: List[List[int]], device: torch.device): """Prepare ys_in_pad and ys_out_pad.""" ys = k2.RaggedTensor(token_ids).to(device=device) - row_splits = ys.shape.row_splits(1) ys_lens = row_splits[1:] - row_splits[:-1] @@ -168,10 +166,9 @@ class AttentionDecoderModel(nn.Module): decoder_out.view(-1, num_classes), ys_out_pad.view(-1), ignore_index=self.ignore_id, - reduction="None", + reduction="none", ) - nll = nll.view(batch_size, -1) - nll = nll.sum(1) + nll = nll.view(batch_size, -1).sum(1) return nll @@ -181,7 +178,7 @@ class TransformerDecoder(nn.Module): Args: vocab_size: output dim - d_model: equal to encoder_dim + d_model: decoder dimension num_decoder_layers: number of decoder layers attention_dim: total dimension of multi head attention n_head: number of attention heads @@ -715,7 +712,7 @@ def subsequent_mask(size, device="cpu", dtype=torch.bool): def _test_attention_decoder_model(): m = AttentionDecoderModel( vocab_size=500, - d_model=384, + decoder_dim=384, unmasked_dim=256, num_decoder_layers=6, attention_dim=192, @@ -733,6 +730,9 @@ def _test_attention_decoder_model(): loss = m.calc_att_loss(encoder_out, encoder_out_lens, token_ids) print(loss) + nll = m.nll(encoder_out, encoder_out_lens, token_ids) + print(nll) + if __name__ == "__main__": _test_attention_decoder_model() diff --git a/egs/librispeech/ASR/zipformer_ctc_attn/model.py b/egs/librispeech/ASR/zipformer_ctc_attn/model.py new file mode 100644 index 000000000..db22af8ad --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_attn/model.py @@ -0,0 +1,95 @@ +# Copyright 2023 Xiaomi Corp. (authors: 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. + + +from typing import List, Tuple + +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface + + +class CTCAttentionModel(nn.Module): + """Hybrid CTC & Attention decoder model.""" + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + encoder_dim: int, + vocab_size: int, + ): + """ + Args: + encoder: + It is the Zipformer encoder model. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + decoder: + It is the attention decoder. + encoder_dim: + The embedding dimension of encoder. + vocab_size: + The vocabulary size. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + + self.encoder = encoder + self.ctc_output = nn.Sequential( + nn.Dropout(p=0.1), + nn.Linear(encoder_dim, vocab_size), + nn.LogSoftmax(dim=-1), + ) + # Attention decoder + self.decoder = decoder + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + token_ids: List[List[int]], + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + token_ids: + A list of token id list. + + Returns: + - ctc_output, ctc log-probs + - att_loss, attention decoder loss + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert x.size(0) == x_lens.size(0) == len(token_ids) + + # encoder forward + encoder_out, x_lens = self.encoder(x, x_lens) + assert torch.all(x_lens > 0) + + # compute ctc log-probs + ctc_output = self.ctc_output(encoder_out) + + # compute attention decoder loss + att_loss = self.decoder.calc_att_loss(encoder_out, x_lens, token_ids) + + return ctc_output, att_loss diff --git a/egs/librispeech/ASR/zipformer_ctc_attn/train.py b/egs/librispeech/ASR/zipformer_ctc_attn/train.py new file mode 100755 index 000000000..27a9add0d --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_attn/train.py @@ -0,0 +1,1268 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 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" + +./zipformer_ctc_attn/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir zipformer_ctc_attn/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./zipformer_ctc_attn/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer_ctc_attn/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 attention_decoder import AttentionDecoderModel +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import CTCAttentionModel +from optim import Eden, ScaledAdam +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 zipformer import Zipformer + +from icefall import diagnostics +from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler +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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=384, + help="Embedding dimension of attention decoder. " + "It should be equal to encoder_dims[-1].", + ) + + parser.add_argument( + "--decoder-unmasked-dim", + type=int, + default=256, + help="Unmasked dimension in the attention decoder, relates to augmentation during training. " + "Must be <= each of decoder_dim. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--num-decoder-layers", + type=int, + default=6, + help="Number of attention decoder layers.", + ) + + parser.add_argument( + "--decoder-attention-dim", + type=int, + default=192, + help="Attention dimension in attention decoder.", + ) + + parser.add_argument( + "--decoder-nhead", + type=int, + default=8, + help="Number of attention heads in attention decoder.", + ) + + parser.add_argument( + "--decoder-feedforward-dim", + type=int, + default=2048, + help="Feedforward dimension in attention decoder.", + ) + + +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="zipformer_ctc_attn/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_bpe_500", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--att-loss-scale", + type=float, + default=0.8, + help="""The attention loss scale. + The total loss is (1 - att_loss_scale) * ctc_loss + att_loss_scale * att_loss. + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + 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=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + 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( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + 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=30, + 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=200, + 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 warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + 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 zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + # parameters for attention decoder loss + "ignore_id": -1, + "label_smoothing": 0.1, + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + encoder_dim = int(params.encoder_dims.split(",")[-1]) + assert params.decoder_dim == encoder_dim, (params.decoder_dim, encoder_dim) + + decoder = AttentionDecoderModel( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + unmasked_dim=params.decoder_unmasked_dim, + num_decoder_layers=params.num_decoder_layers, + attention_dim=params.decoder_attention_dim, + nhead=params.decoder_nhead, + feedforward_dim=params.decoder_feedforward_dim, + sos_id=params.sos_id, + eos_id=params.eos_id, + ignore_id=params.ignore_id, + label_smoothing=params.label_smoothing, + ) + return decoder + + +def get_ctc_attention_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + + model = CTCAttentionModel( + encoder=encoder, + decoder=decoder, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + 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], + graph_compiler: BpeCtcTrainingGraphCompiler, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer 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 Zipformer in our case. + graph_compiler: + It is used to build a decoding graph from a ctc topo and training + transcript. The training transcript is contained in the given `batch`, + while the ctc topo is built when this compiler is instantiated. + 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) + token_ids = graph_compiler.texts_to_ids(supervisions["text"]) + + with torch.set_grad_enabled(is_training): + ctc_output, att_loss = model( + x=feature, + x_lens=feature_lens, + token_ids=token_ids, + ) + assert att_loss.requires_grad == is_training + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + # token_ids has been sorted + decoding_graph = graph_compiler.compile(token_ids) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + # Compute ctc loss + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + + scale = params.att_loss_scale + assert 0 <= scale <= 1, scale + loss = (1 - scale) * ctc_loss + scale * att_loss + + 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["ctc_loss"] = ctc_loss.detach().cpu().item() + info["att_loss"] = att_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + graph_compiler: BpeCtcTrainingGraphCompiler, + 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, + graph_compiler=graph_compiler, + 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, + graph_compiler: BpeCtcTrainingGraphCompiler, + 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. + graph_compiler: + It is used to convert transcripts to FSAs. + 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"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, sp=graph_compiler.sp) + raise + + 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 % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + 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}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + 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 params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + 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}") + + assert "lang_bpe" in str(params.lang_dir), "Currently only supports bpe model." + graph_compiler = BpeCtcTrainingGraphCompiler( + params.lang_dir, + device=device, + sos_token="", + eos_token="", + ) + # sos_id, eos_id, ignore_id will be used in AttentionDecoderModel + params.sos_id = graph_compiler.sos_id + params.eos_id = graph_compiler.eos_id + # is defined in local/train_bpe_model.py + params.vocab_size = graph_compiler.sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_ctc_attention_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of total model parameters: {num_param}") + num_param = sum([p.numel() for p in model.decoder.parameters()]) + logging.info(f"Number of parameters in attention decoder: {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).to(torch.float64) + + 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], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + 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) + + if params.inf_check: + register_inf_check_hooks(model) + + 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 + if c.duration < 1.0 or c.duration > 20.0: + logging.warning( + f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = graph_compiler.sp.encode(c.supervisions[0].text, out_type=str) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + 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, + graph_compiler=graph_compiler, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + 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, + graph_compiler=graph_compiler, + 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 display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + graph_compiler: BpeCtcTrainingGraphCompiler, + 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: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception 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]}) ..." + ) + display_and_save_batch(batch, params=params, sp=graph_compiler.sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +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()