From f4912b283d71490d965e025b9ae966bc123dc906 Mon Sep 17 00:00:00 2001 From: marcoyang Date: Wed, 24 Jul 2024 15:02:50 +0800 Subject: [PATCH] remove the previous version --- egs/librispeech/ASR/zipformer/train_bf16.py | 1486 ---------- .../ASR/zipformer/zipformer_bf16.py | 2437 ----------------- 2 files changed, 3923 deletions(-) delete mode 100644 egs/librispeech/ASR/zipformer/train_bf16.py delete mode 100644 egs/librispeech/ASR/zipformer/zipformer_bf16.py diff --git a/egs/librispeech/ASR/zipformer/train_bf16.py b/egs/librispeech/ASR/zipformer/train_bf16.py deleted file mode 100644 index 9b6f4a93a..000000000 --- a/egs/librispeech/ASR/zipformer/train_bf16.py +++ /dev/null @@ -1,1486 +0,0 @@ -#!/usr/bin/env python3 -# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, -# Wei Kang, -# Mingshuang Luo, -# Zengwei Yao, -# Daniel Povey) -# -# 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" - -# For non-streaming model training: -./zipformer/train.py \ - --world-size 4 \ - --num-epochs 30 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir zipformer/exp \ - --full-libri 1 \ - --max-duration 1000 - -# For streaming model training: -./zipformer/train.py \ - --world-size 4 \ - --num-epochs 30 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir zipformer/exp \ - --causal 1 \ - --full-libri 1 \ - --max-duration 1000 - -It supports training with: - - transducer loss (default), with `--use-transducer True --use-ctc False` - - ctc loss (not recommended), with `--use-transducer False --use-ctc True` - - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` - - ctc loss & attention decoder loss, no transducer loss, - with `--use-transducer False --use-ctc True --use-attention-decoder True` -""" - - -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 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 AsrModel -from optim import Eden, ScaledAdam -from scaling import ScheduledFloat -from subsampling import Conv2dSubsampling -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 Zipformer2 - -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.err import raise_grad_scale_is_too_small_error -from icefall.hooks import register_inf_check_hooks -from icefall.utils import ( - AttributeDict, - MetricsTracker, - get_parameter_groups_with_lrs, - setup_logger, - str2bool, -) - -LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] - - -def get_adjusted_batch_count(params: AttributeDict) -> float: - # returns the number of batches we would have used so far if we had used the reference - # duration. This is for purposes of set_batch_count(). - return ( - params.batch_idx_train - * (params.max_duration * params.world_size) - / params.ref_duration - ) - - -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 name, module in model.named_modules(): - if hasattr(module, "batch_count"): - module.batch_count = batch_count - if hasattr(module, "name"): - module.name = name - - -def add_model_arguments(parser: argparse.ArgumentParser): - parser.add_argument( - "--num-encoder-layers", - type=str, - default="2,2,3,4,3,2", - help="Number of zipformer encoder layers per stack, comma separated.", - ) - - parser.add_argument( - "--downsampling-factor", - type=str, - default="1,2,4,8,4,2", - help="Downsampling factor for each stack of encoder layers.", - ) - - parser.add_argument( - "--feedforward-dim", - type=str, - default="512,768,1024,1536,1024,768", - help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", - ) - - parser.add_argument( - "--num-heads", - type=str, - default="4,4,4,8,4,4", - help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", - ) - - parser.add_argument( - "--encoder-dim", - type=str, - default="192,256,384,512,384,256", - help="Embedding dimension in encoder stacks: a single int or comma-separated list.", - ) - - parser.add_argument( - "--query-head-dim", - type=str, - default="32", - help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", - ) - - parser.add_argument( - "--value-head-dim", - type=str, - default="12", - help="Value dimension per head in encoder stacks: a single int or comma-separated list.", - ) - - parser.add_argument( - "--pos-head-dim", - type=str, - default="4", - help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", - ) - - parser.add_argument( - "--pos-dim", - type=int, - default="48", - help="Positional-encoding embedding dimension", - ) - - parser.add_argument( - "--encoder-unmasked-dim", - type=str, - default="192,192,256,256,256,192", - help="Unmasked dimensions in the encoders, relates to augmentation during training. " - "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", - ) - - parser.add_argument( - "--cnn-module-kernel", - type=str, - default="31,31,15,15,15,31", - help="Sizes of convolutional kernels in convolution modules in each encoder stack: " - "a single int or comma-separated list.", - ) - - parser.add_argument( - "--decoder-dim", - type=int, - default=512, - help="Embedding dimension in the decoder model.", - ) - - parser.add_argument( - "--joiner-dim", - type=int, - default=512, - help="""Dimension used in the joiner model. - Outputs from the encoder and decoder model are projected - to this dimension before adding. - """, - ) - - parser.add_argument( - "--attention-decoder-dim", - type=int, - default=512, - help="""Dimension used in the attention decoder""", - ) - - parser.add_argument( - "--attention-decoder-num-layers", - type=int, - default=6, - help="""Number of transformer layers used in attention decoder""", - ) - - parser.add_argument( - "--attention-decoder-attention-dim", - type=int, - default=512, - help="""Attention dimension used in attention decoder""", - ) - - parser.add_argument( - "--attention-decoder-num-heads", - type=int, - default=8, - help="""Number of attention heads used in attention decoder""", - ) - - parser.add_argument( - "--attention-decoder-feedforward-dim", - type=int, - default=2048, - help="""Feedforward dimension used in attention decoder""", - ) - - parser.add_argument( - "--causal", - type=str2bool, - default=False, - help="If True, use causal version of model.", - ) - - parser.add_argument( - "--chunk-size", - type=str, - default="16,32,64,-1", - help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " - " Must be just -1 if --causal=False", - ) - - parser.add_argument( - "--left-context-frames", - type=str, - default="64,128,256,-1", - help="Maximum left-contexts for causal training, measured in frames which will " - "be converted to a number of chunks. If splitting into chunks, " - "chunk left-context frames will be chosen randomly from this list; else not relevant.", - ) - - parser.add_argument( - "--use-transducer", - type=str2bool, - default=True, - help="If True, use Transducer head.", - ) - - parser.add_argument( - "--use-ctc", - type=str2bool, - default=False, - help="If True, use CTC head.", - ) - - parser.add_argument( - "--use-attention-decoder", - type=str2bool, - default=False, - help="If True, use attention-decoder head.", - ) - - -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/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( - "--base-lr", type=float, default=0.045, help="The base learning rate." - ) - - parser.add_argument( - "--lr-batches", - type=float, - default=7500, - 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( - "--ref-duration", - type=float, - default=600, - help="Reference batch duration for purposes of adjusting batch counts for setting various " - "schedules inside the model", - ) - - 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( - "--ctc-loss-scale", - type=float, - default=0.2, - help="Scale for CTC loss.", - ) - - parser.add_argument( - "--attention-decoder-loss-scale", - type=float, - default=0.8, - help="Scale for attention-decoder 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( - "--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=4000, - 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 1. - """, - ) - - 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. - - - 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 attention-decoder - "ignore_id": -1, - "label_smoothing": 0.1, - "warm_step": 2000, - "env_info": get_env_info(), - } - ) - - return params - - -def _to_int_tuple(s: str): - return tuple(map(int, s.split(","))) - - -def get_encoder_embed(params: AttributeDict) -> nn.Module: - # encoder_embed converts the input of shape (N, T, num_features) - # to the shape (N, (T - 7) // 2, encoder_dims). - # That is, it does two things simultaneously: - # (1) subsampling: T -> (T - 7) // 2 - # (2) embedding: num_features -> encoder_dims - # In the normal configuration, we will downsample once more at the end - # by a factor of 2, and most of the encoder stacks will run at a lower - # sampling rate. - encoder_embed = Conv2dSubsampling( - in_channels=params.feature_dim, - out_channels=_to_int_tuple(params.encoder_dim)[0], - dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), - ) - return encoder_embed - - -def get_encoder_model(params: AttributeDict) -> nn.Module: - encoder = Zipformer2( - output_downsampling_factor=2, - downsampling_factor=_to_int_tuple(params.downsampling_factor), - num_encoder_layers=_to_int_tuple(params.num_encoder_layers), - encoder_dim=_to_int_tuple(params.encoder_dim), - encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), - query_head_dim=_to_int_tuple(params.query_head_dim), - pos_head_dim=_to_int_tuple(params.pos_head_dim), - value_head_dim=_to_int_tuple(params.value_head_dim), - pos_dim=params.pos_dim, - num_heads=_to_int_tuple(params.num_heads), - feedforward_dim=_to_int_tuple(params.feedforward_dim), - cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), - dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), - warmup_batches=4000.0, - causal=params.causal, - chunk_size=_to_int_tuple(params.chunk_size), - left_context_frames=_to_int_tuple(params.left_context_frames), - ) - 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=max(_to_int_tuple(params.encoder_dim)), - decoder_dim=params.decoder_dim, - joiner_dim=params.joiner_dim, - vocab_size=params.vocab_size, - ) - return joiner - - -def get_attention_decoder_model(params: AttributeDict) -> nn.Module: - decoder = AttentionDecoderModel( - vocab_size=params.vocab_size, - decoder_dim=params.attention_decoder_dim, - num_decoder_layers=params.attention_decoder_num_layers, - attention_dim=params.attention_decoder_attention_dim, - num_heads=params.attention_decoder_num_heads, - feedforward_dim=params.attention_decoder_feedforward_dim, - memory_dim=max(_to_int_tuple(params.encoder_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_model(params: AttributeDict) -> nn.Module: - assert params.use_transducer or params.use_ctc, ( - f"At least one of them should be True, " - f"but got params.use_transducer={params.use_transducer}, " - f"params.use_ctc={params.use_ctc}" - ) - - encoder_embed = get_encoder_embed(params) - encoder = get_encoder_model(params) - - if params.use_transducer: - decoder = get_decoder_model(params) - joiner = get_joiner_model(params) - else: - decoder = None - joiner = None - - if params.use_attention_decoder: - attention_decoder = get_attention_decoder_model(params) - else: - attention_decoder = None - - model = AsrModel( - encoder_embed=encoder_embed, - encoder=encoder, - decoder=decoder, - joiner=joiner, - attention_decoder=attention_decoder, - encoder_dim=max(_to_int_tuple(params.encoder_dim)), - decoder_dim=params.decoder_dim, - vocab_size=params.vocab_size, - use_transducer=params.use_transducer, - use_ctc=params.use_ctc, - use_attention_decoder=params.use_attention_decoder, - ) - 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"] - - 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, -) -> Tuple[Tensor, MetricsTracker]: - """ - Compute 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. - 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) - - batch_idx_train = params.batch_idx_train - warm_step = params.warm_step - - texts = batch["supervisions"]["text"] - y = sp.encode(texts, out_type=int) - y = k2.RaggedTensor(y) - - with torch.set_grad_enabled(is_training): - simple_loss, pruned_loss, ctc_loss, attention_decoder_loss = model( - x=feature, - x_lens=feature_lens, - y=y, - prune_range=params.prune_range, - am_scale=params.am_scale, - lm_scale=params.lm_scale, - ) - - loss = 0.0 - - if params.use_transducer: - s = params.simple_loss_scale - # take down the scale on the simple loss from 1.0 at the start - # to params.simple_loss scale by warm_step. - simple_loss_scale = ( - s - if batch_idx_train >= warm_step - else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) - ) - pruned_loss_scale = ( - 1.0 - if batch_idx_train >= warm_step - else 0.1 + 0.9 * (batch_idx_train / warm_step) - ) - loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss - - if params.use_ctc: - loss += params.ctc_loss_scale * ctc_loss - - if params.use_attention_decoder: - loss += params.attention_decoder_loss_scale * attention_decoder_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() - if params.use_transducer: - info["simple_loss"] = simple_loss.detach().cpu().item() - info["pruned_loss"] = pruned_loss.detach().cpu().item() - if params.use_ctc: - info["ctc_loss"] = ctc_loss.detach().cpu().item() - if params.use_attention_decoder: - info["attn_decoder_loss"] = attention_decoder_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() - - saved_bad_model = False - - def save_bad_model(suffix: str = ""): - save_checkpoint_impl( - filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", - model=model, - model_avg=model_avg, - params=params, - optimizer=optimizer, - scheduler=scheduler, - sampler=train_dl.sampler, - scaler=scaler, - rank=0, - ) - - for batch_idx, batch in enumerate(train_dl): - if batch_idx % 10 == 0: - set_batch_count(model, get_adjusted_batch_count(params)) - - 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, - sp=sp, - batch=batch, - is_training=True, - ) - # summary stats - tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info - - # NOTE: We use reduction==sum and loss is computed over utterances - # in the batch and there is no normalization to it so far. - scaler.scale(loss).backward() - scheduler.step_batch(params.batch_idx_train) - - scaler.step(optimizer) - scaler.update() - optimizer.zero_grad() - except Exception as e: - logging.info( - f"Caught exception: {e}." - ) - save_bad_model() - display_and_save_batch(batch, params=params, sp=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 - ): - 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, - ) - 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 < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): - scaler.update(cur_grad_scale * 2.0) - if cur_grad_scale < 0.01: - if not saved_bad_model: - save_bad_model(suffix="-first-warning") - saved_bad_model = True - logging.warning(f"Grad scale is small: {cur_grad_scale}") - if cur_grad_scale < 1.0e-05: - save_bad_model() - raise_grad_scale_is_too_small_error(cur_grad_scale) - - if batch_idx % params.log_interval == 0: - cur_lr = max(scheduler.get_last_lr()) - 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, - sp=sp, - 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)) - - 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.sos_id = params.eos_id = sp.piece_to_id("") - params.vocab_size = sp.get_piece_size() - - if not params.use_transducer: - if not params.use_attention_decoder: - params.ctc_loss_scale = 1.0 - else: - assert params.ctc_loss_scale + params.attention_decoder_loss_scale == 1.0, ( - params.ctc_loss_scale, params.attention_decoder_loss_scale - ) - - logging.info(params) - - logging.info("About to create model") - model = get_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).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) - - optimizer = ScaledAdam( - get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), - lr=params.base_lr, # should have no effect - clipping_scale=2.0, - ) - - 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( - 512 - ) # allow 4 megabytes per sub-module - diagnostic = diagnostics.attach_diagnostics(model, opts) - - if params.inf_check: - register_inf_check_hooks(model) - - librispeech = LibriSpeechAsrDataModule(args) - - if params.full_libri: - train_cuts = librispeech.train_all_shuf_cuts() - - # previously we used the following code to load all training cuts, - # strictly speaking, shuffled training cuts should be used instead, - # but we leave the code here to demonstrate that there is an option - # like this to combine multiple cutsets - - # train_cuts = librispeech.train_clean_100_cuts() - # train_cuts += librispeech.train_clean_360_cuts() - # train_cuts += librispeech.train_other_500_cuts() - else: - train_cuts = librispeech.train_clean_100_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 = 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, - sp=sp, - 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, - 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 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, - 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: - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, _ = compute_loss( - params=params, - model=model, - sp=sp, - 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=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() diff --git a/egs/librispeech/ASR/zipformer/zipformer_bf16.py b/egs/librispeech/ASR/zipformer/zipformer_bf16.py deleted file mode 100644 index 69059287b..000000000 --- a/egs/librispeech/ASR/zipformer/zipformer_bf16.py +++ /dev/null @@ -1,2437 +0,0 @@ -#!/usr/bin/env python3 -# Copyright 2022-2023 Xiaomi Corp. (authors: Daniel Povey, -# 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. - -import copy -import logging -import math -import random -import warnings -from typing import List, Optional, Tuple, Union - -import torch -from encoder_interface import EncoderInterface -from scaling import ( - Identity, # more friendly to backward hooks than nn.Identity(), for diagnostic reasons. -) -from scaling import ( - ScaledLinear, # not as in other dirs.. just scales down initial parameter values. -) -from scaling import ( - ActivationDropoutAndLinear, - Balancer, - BiasNorm, - ChunkCausalDepthwiseConv1d, - Dropout2, - FloatLike, - ScheduledFloat, - Whiten, - convert_num_channels, - limit_param_value, - penalize_abs_values_gt, - softmax, -) -from torch import Tensor, nn - - -class Zipformer2(EncoderInterface): - """ - Args: - - Note: all "int or Tuple[int]" arguments below will be treated as lists of the same length - as downsampling_factor if they are single ints or one-element tuples. The length of - downsampling_factor defines the number of stacks. - - output_downsampling_factor (int): how much to downsample at the output. Note: - we also downsample by a factor of 2 in the Conv2dSubsampling encoder. - You should probably leave this at 2. - downsampling_factor (Tuple[int]): downsampling factor for each encoder stack. - Note: this is in addition to the downsampling factor of 2 that is applied in - the frontend (self.encoder_embed). - encoder_dim (Tuple[int]): embedding dimension of each of the encoder stacks, one per - encoder stack. - num_encoder_layers (int or Tuple[int])): number of encoder layers for each stack - encoder_unmasked_dim (int or Tuple[int]): unmasked dimension in each of - the encoder stacks for purposes of per-frame dropout (recommend 256 for - now). - query_head_dim (int or Tuple[int]): dimension of query and key per attention - head: per stack, if a tuple.. - pos_head_dim (int or Tuple[int]): dimension of positional-encoding projection per - attention head - value_head_dim (int or Tuple[int]): dimension of value in each attention head - num_heads: (int or Tuple[int]): number of heads in the self-attention mechanism. - Must be at least 4. - feedforward_dim (int or Tuple[int]): hidden dimension in feedforward modules - cnn_module_kernel (int or Tuple[int])): Kernel size of convolution module - - pos_dim (int): the dimension of each positional-encoding vector prior to projection, - e.g. 128. - - dropout (float): dropout rate - warmup_batches (float): number of batches to warm up over; this controls - dropout of encoder layers. - causal (bool): if True, support chunkwise causal convolution. This should - not hurt WER as no modeling power is lost, but the convolution modules will be - slightly slower and use more memory. Enables use of the chunk_size and - left_context_chunks options in forward(), which simulates streaming - decoding. - chunk_size: (list of int): only set this to other than [-1] if causal; - the chunk size will be randomly chosen from this list. -1 means no chunking. - left_context_frames: (list of int): determines the number of left- - context chunks for causal training; will be rounded to a number of - chunks. Must not be less than cnn_module_kernel (after factoring in - rounding and downsampling); an error will be thrown if this is violated. - """ - - def __init__( - self, - output_downsampling_factor: int = 2, - downsampling_factor: Tuple[int] = (2, 4), - encoder_dim: Union[int, Tuple[int]] = 384, - num_encoder_layers: Union[int, Tuple[int]] = 4, - encoder_unmasked_dim: Union[int, Tuple[int]] = 256, - query_head_dim: Union[int, Tuple[int]] = 24, - pos_head_dim: Union[int, Tuple[int]] = 4, - value_head_dim: Union[int, Tuple[int]] = 12, - num_heads: Union[int, Tuple[int]] = 8, - feedforward_dim: Union[int, Tuple[int]] = 1536, - cnn_module_kernel: Union[int, Tuple[int]] = 31, - pos_dim: int = 192, - dropout: FloatLike = None, # see code below for default - warmup_batches: float = 4000.0, - causal: bool = False, - chunk_size: Tuple[int] = [-1], - left_context_frames: Tuple[int] = [-1], - ) -> None: - super(Zipformer2, self).__init__() - - if dropout is None: - dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1)) - - def _to_tuple(x): - """Converts a single int or a 1-tuple of an int to a tuple with the same length - as downsampling_factor""" - if isinstance(x, int): - x = (x,) - if len(x) == 1: - x = x * len(downsampling_factor) - else: - assert len(x) == len(downsampling_factor) and isinstance(x[0], int) - return x - - self.output_downsampling_factor = output_downsampling_factor # int - self.downsampling_factor = downsampling_factor # tuple - self.encoder_dim = encoder_dim = _to_tuple(encoder_dim) # tuple - self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple( - encoder_unmasked_dim - ) # tuple - num_encoder_layers = _to_tuple(num_encoder_layers) - self.num_encoder_layers = num_encoder_layers - self.query_head_dim = query_head_dim = _to_tuple(query_head_dim) - self.value_head_dim = value_head_dim = _to_tuple(value_head_dim) - pos_head_dim = _to_tuple(pos_head_dim) - self.num_heads = num_heads = _to_tuple(num_heads) - feedforward_dim = _to_tuple(feedforward_dim) - self.cnn_module_kernel = cnn_module_kernel = _to_tuple(cnn_module_kernel) - - self.causal = causal - self.chunk_size = chunk_size - self.left_context_frames = left_context_frames - - for u, d in zip(encoder_unmasked_dim, encoder_dim): - assert u <= d - - # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder - encoders = [] - - num_encoders = len(downsampling_factor) - for i in range(num_encoders): - encoder_layer = Zipformer2EncoderLayer( - embed_dim=encoder_dim[i], - pos_dim=pos_dim, - num_heads=num_heads[i], - query_head_dim=query_head_dim[i], - pos_head_dim=pos_head_dim[i], - value_head_dim=value_head_dim[i], - feedforward_dim=feedforward_dim[i], - dropout=dropout, - cnn_module_kernel=cnn_module_kernel[i], - causal=causal, - ) - - # For the segment of the warmup period, we let the Conv2dSubsampling - # layer learn something. Then we start to warm up the other encoders. - encoder = Zipformer2Encoder( - encoder_layer, - num_encoder_layers[i], - pos_dim=pos_dim, - dropout=dropout, - warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1), - warmup_end=warmup_batches * (i + 2) / (num_encoders + 1), - final_layerdrop_rate=0.035 * (downsampling_factor[i] ** 0.5), - ) - - if downsampling_factor[i] != 1: - encoder = DownsampledZipformer2Encoder( - encoder, - dim=encoder_dim[i], - downsample=downsampling_factor[i], - dropout=dropout, - ) - - encoders.append(encoder) - - self.encoders = nn.ModuleList(encoders) - - self.downsample_output = SimpleDownsample( - max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout - ) - - def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]: - """ - In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of - randomized feature masks, one per encoder. - On e.g. 15% of frames, these masks will zero out all encoder dims larger than - some supplied number, e.g. >256, so in effect on those frames we are using - a smaller encoder dim. - - We generate the random masks at this level because we want the 2 masks to 'agree' - all the way up the encoder stack. This will mean that the 1st mask will have - mask values repeated self.zipformer_subsampling_factor times. - - Args: - x: the embeddings (needed for the shape and dtype and device), of shape - (1, batch_size, encoder_dims0) - """ - num_encoders = len(self.encoder_dim) - if not self.training: - return [1.0] * num_encoders - - (num_frames0, batch_size, _encoder_dims0) = x.shape - - assert self.encoder_dim[0] == _encoder_dims0, ( - self.encoder_dim[0], - _encoder_dims0, - ) - - feature_mask_dropout_prob = 0.125 - - # mask1 shape: (1, batch_size, 1) - mask1 = ( - torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob - ).to(x.dtype) - - # mask2 has additional sequences masked, about twice the number. - mask2 = torch.logical_and( - mask1, - ( - torch.rand(1, batch_size, 1, device=x.device) - > feature_mask_dropout_prob - ).to(x.dtype), - ) - - # dim: (1, batch_size, 2) - mask = torch.cat((mask1, mask2), dim=-1) - - feature_masks = [] - for i in range(num_encoders): - channels = self.encoder_dim[i] - feature_mask = torch.ones( - 1, batch_size, channels, dtype=x.dtype, device=x.device - ) - u1 = self.encoder_unmasked_dim[i] - u2 = u1 + (channels - u1) // 2 - - feature_mask[:, :, u1:u2] *= mask[..., 0:1] - feature_mask[:, :, u2:] *= mask[..., 1:2] - - feature_masks.append(feature_mask) - - return feature_masks - - def get_chunk_info(self) -> Tuple[int, int]: - """ - Returns chunk_size and left_context_chunks. - """ - if not self.causal: - return -1, -1 - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - assert len(self.chunk_size) == 1, self.chunk_size - chunk_size = self.chunk_size[0] - else: - chunk_size = random.choice(self.chunk_size) - - if chunk_size == -1: - left_context_chunks = -1 - else: - if torch.jit.is_scripting() or torch.jit.is_tracing(): - assert len(self.left_context_frames) == 1, self.left_context_frames - left_context_frames = self.left_context_frames[0] - else: - left_context_frames = random.choice(self.left_context_frames) - # Note: in Python, -1 // n == -1 for n > 0 - left_context_chunks = left_context_frames // chunk_size - if left_context_chunks == 0: - left_context_chunks = 1 - - return chunk_size, left_context_chunks - - def forward( - self, - x: Tensor, - x_lens: Tensor, - src_key_padding_mask: Optional[Tensor] = None, - ) -> Tuple[Tensor, Tensor]: - """ - Args: - x: - The input tensor. Its shape is (seq_len, batch_size, feature_dim). - x_lens: - A tensor of shape (batch_size,) containing the number of frames in - `x` before padding. - src_key_padding_mask: - The mask for padding, of shape (batch_size, seq_len); True means - masked position. May be None. - Returns: - Return a tuple containing 2 tensors: - - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim)) - - lengths, a tensor of shape (batch_size,) containing the number - of frames in `embeddings` before padding. - """ - outputs = [] - if torch.jit.is_scripting() or torch.jit.is_tracing(): - feature_masks = [1.0] * len(self.encoder_dim) - else: - feature_masks = self.get_feature_masks(x) - - chunk_size, left_context_chunks = self.get_chunk_info() - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - # Not support exporting a model for simulating streaming decoding - attn_mask = None - else: - attn_mask = self._get_attn_mask(x, chunk_size, left_context_chunks) - - for i, module in enumerate(self.encoders): - ds = self.downsampling_factor[i] - x = convert_num_channels(x, self.encoder_dim[i]) - - x = module( - x, - chunk_size=chunk_size, - feature_mask=feature_masks[i], - src_key_padding_mask=( - None - if src_key_padding_mask is None - else src_key_padding_mask[..., ::ds] - ), - attn_mask=attn_mask, - ) - outputs.append(x) - - # if the last output has the largest dimension, x will be unchanged, - # it will be the same as outputs[-1]. Otherwise it will be concatenated - # from different pieces of 'outputs', taking each dimension from the - # most recent output that has it present. - x = self._get_full_dim_output(outputs) - x = self.downsample_output(x) - # class Downsample has this rounding behavior.. - assert self.output_downsampling_factor == 2, self.output_downsampling_factor - if torch.jit.is_scripting() or torch.jit.is_tracing(): - lengths = (x_lens + 1) // 2 - else: - with warnings.catch_warnings(): - warnings.simplefilter("ignore") - lengths = (x_lens + 1) // 2 - - return x, lengths - - def _get_attn_mask( - self, x: Tensor, chunk_size: int, left_context_chunks: int - ) -> Optional[Tensor]: - """ - Return None if chunk_size == -1, else return attention mask of shape - (seq_len, seq_len), interpreted as (tgt_seq_len, src_seq_len). True - means a masked position. - Args: - x: embeddings after self.encoder_embed(), of shape (seq_len, batch_size, embed_dim). - chunk_size: chunk size, must divide - """ - if chunk_size <= 0: - return None - assert all(chunk_size % d == 0 for d in self.downsampling_factor) - if left_context_chunks >= 0: - num_encoders = len(self.encoder_dim) - assert all( - chunk_size * left_context_chunks - >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i] - for i in range(num_encoders) - ) - else: - left_context_chunks = 1000000 - - seq_len = x.shape[0] - - # t is frame index, shape (seq_len,) - t = torch.arange(seq_len, dtype=torch.int32, device=x.device) - # c is chunk index for each frame, shape (seq_len,) - if torch.jit.is_scripting() or torch.jit.is_tracing(): - c = t // chunk_size - else: - with warnings.catch_warnings(): - warnings.simplefilter("ignore") - c = t // chunk_size - src_c = c - tgt_c = c.unsqueeze(-1) - - attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks) - if __name__ == "__main__": - logging.info(f"attn_mask = {attn_mask}") - return attn_mask - - def _get_full_dim_output(self, outputs: List[Tensor]): - num_encoders = len(self.encoder_dim) - assert len(outputs) == num_encoders - output_dim = max(self.encoder_dim) - output_pieces = [outputs[-1]] - cur_dim = self.encoder_dim[-1] - for i in range(num_encoders - 2, -1, -1): - d = self.encoder_dim[i] - if d > cur_dim: - this_output = outputs[i] - output_pieces.append(this_output[..., cur_dim:d]) - cur_dim = d - assert cur_dim == output_dim - return torch.cat(output_pieces, dim=-1) - - def streaming_forward( - self, - x: Tensor, - x_lens: Tensor, - states: List[Tensor], - src_key_padding_mask: Tensor, - ) -> Tuple[Tensor, Tensor, List[Tensor]]: - """ - Args: - x: - The input tensor. Its shape is (seq_len, batch_size, feature_dim). - x_lens: - A tensor of shape (batch_size,) containing the number of frames in - `x` before padding. - states: list of cached tensors of all encoder layers. For layer-i, - states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, - cached_conv1, cached_conv2). - src_key_padding_mask: - The mask for padding, of shape (batch_size, seq_len); True means - masked position. May be None. - Returns: - Return a tuple containing 2 tensors: - - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim)) - - lengths, a tensor of shape (batch_size,) containing the number - of frames in `embeddings` before padding. - - updated states - """ - outputs = [] - new_states = [] - layer_offset = 0 - - for i, module in enumerate(self.encoders): - num_layers = module.num_layers - ds = self.downsampling_factor[i] - x = convert_num_channels(x, self.encoder_dim[i]) - - x, new_layer_states = module.streaming_forward( - x, - states=states[layer_offset * 6 : (layer_offset + num_layers) * 6], - left_context_len=self.left_context_frames[0] // ds, - src_key_padding_mask=src_key_padding_mask[..., ::ds], - ) - layer_offset += num_layers - outputs.append(x) - new_states += new_layer_states - - # if the last output has the largest dimension, x will be unchanged, - # it will be the same as outputs[-1]. Otherwise it will be concatenated - # from different pieces of 'outputs', taking each dimension from the - # most recent output that has it present. - x = self._get_full_dim_output(outputs) - x = self.downsample_output(x) - # class Downsample has this rounding behavior.. - assert self.output_downsampling_factor == 2 - if torch.jit.is_scripting() or torch.jit.is_tracing(): - lengths = (x_lens + 1) // 2 - else: - with warnings.catch_warnings(): - warnings.simplefilter("ignore") - lengths = (x_lens + 1) // 2 - - return x, lengths, new_states - - @torch.jit.export - def get_init_states( - self, - batch_size: int = 1, - device: torch.device = torch.device("cpu"), - ) -> List[Tensor]: - """Get initial states. - - A list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] - is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). - """ - states = [] - for i, module in enumerate(self.encoders): - num_layers = module.num_layers - embed_dim = self.encoder_dim[i] - ds = self.downsampling_factor[i] - num_heads = self.num_heads[i] - key_dim = self.query_head_dim[i] * num_heads - value_dim = self.value_head_dim[i] * num_heads - downsample_left = self.left_context_frames[0] // ds - nonlin_attn_head_dim = 3 * embed_dim // 4 - conv_left_pad = self.cnn_module_kernel[i] // 2 - for layer in range(num_layers): - cached_key = torch.zeros(downsample_left, batch_size, key_dim).to( - device - ) - cached_nonlin_attn = torch.zeros( - 1, batch_size, downsample_left, nonlin_attn_head_dim - ).to(device) - cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to( - device - ) - cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to( - device - ) - cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( - device - ) - cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( - device - ) - states += [ - cached_key, - cached_nonlin_attn, - cached_val1, - cached_val2, - cached_conv1, - cached_conv2, - ] - - return states - - -def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat: - return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x) - - -def _balancer_schedule(min_prob: float): - return ScheduledFloat((0.0, 0.4), (8000.0, min_prob)) - - -class Zipformer2EncoderLayer(nn.Module): - """ - Args: - embed_dim: the number of expected features in the input (required). - nhead: the number of heads in the multiheadattention models (required). - feedforward_dim: the dimension of the feedforward network model (required). - dropout: the dropout value (default=0.1). - cnn_module_kernel (int): Kernel size of convolution module (default=31). - - Examples:: - >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) - >>> src = torch.rand(10, 32, 512) - >>> pos_emb = torch.rand(32, 19, 512) - >>> out = encoder_layer(src, pos_emb) - """ - - def __init__( - self, - embed_dim: int, - pos_dim: int, - num_heads: int, - query_head_dim: int, - pos_head_dim: int, - value_head_dim: int, - feedforward_dim: int, - dropout: FloatLike = 0.1, - cnn_module_kernel: int = 31, - causal: bool = False, - attention_skip_rate: FloatLike = ScheduledFloat( - (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 - ), - conv_skip_rate: FloatLike = ScheduledFloat( - (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 - ), - const_attention_rate: FloatLike = ScheduledFloat( - (0.0, 0.25), (4000.0, 0.025), default=0 - ), - ff2_skip_rate: FloatLike = ScheduledFloat( - (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) - ), - ff3_skip_rate: FloatLike = ScheduledFloat( - (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) - ), - bypass_skip_rate: FloatLike = ScheduledFloat( - (0.0, 0.5), (4000.0, 0.02), default=0 - ), - ) -> None: - super(Zipformer2EncoderLayer, self).__init__() - self.embed_dim = embed_dim - - # self.bypass implements layer skipping as well as bypass; see its default values. - self.bypass = BypassModule( - embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0 - ) - # bypass_mid is bypass used in the middle of the layer. - self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0) - - # skip probability for dynamic modules (meaning: anything but feedforward). - self.attention_skip_rate = copy.deepcopy(attention_skip_rate) - # an additional skip probability that applies to ConvModule to stop it from - # contributing too much early on. - self.conv_skip_rate = copy.deepcopy(conv_skip_rate) - - # ff2_skip_rate is to prevent the ff2 module from having output that's too big - # compared to its residual. - self.ff2_skip_rate = copy.deepcopy(ff2_skip_rate) - self.ff3_skip_rate = copy.deepcopy(ff3_skip_rate) - - self.const_attention_rate = copy.deepcopy(const_attention_rate) - - self.self_attn_weights = RelPositionMultiheadAttentionWeights( - embed_dim, - pos_dim=pos_dim, - num_heads=num_heads, - query_head_dim=query_head_dim, - pos_head_dim=pos_head_dim, - dropout=0.0, - ) - - self.self_attn1 = SelfAttention(embed_dim, num_heads, value_head_dim) - - self.self_attn2 = SelfAttention(embed_dim, num_heads, value_head_dim) - - self.feed_forward1 = FeedforwardModule( - embed_dim, (feedforward_dim * 3) // 4, dropout - ) - - self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout) - - self.feed_forward3 = FeedforwardModule( - embed_dim, (feedforward_dim * 5) // 4, dropout - ) - - self.nonlin_attention = NonlinAttention( - embed_dim, hidden_channels=3 * embed_dim // 4 - ) - - self.conv_module1 = ConvolutionModule( - embed_dim, cnn_module_kernel, causal=causal - ) - - self.conv_module2 = ConvolutionModule( - embed_dim, cnn_module_kernel, causal=causal - ) - - # TODO: remove it - self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) - - self.norm = BiasNorm(embed_dim) - - self.balancer1 = Balancer( - embed_dim, - channel_dim=-1, - min_positive=0.45, - max_positive=0.55, - min_abs=0.2, - max_abs=4.0, - ) - - # balancer for output of NonlinAttentionModule - self.balancer_na = Balancer( - embed_dim, - channel_dim=-1, - min_positive=0.3, - max_positive=0.7, - min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)), - prob=0.05, # out of concern for memory usage - ) - - # balancer for output of feedforward2, prevent it from staying too - # small. give this a very small probability, even at the start of - # training, it's to fix a rare problem and it's OK to fix it slowly. - self.balancer_ff2 = Balancer( - embed_dim, - channel_dim=-1, - min_positive=0.3, - max_positive=0.7, - min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0), - max_abs=2.0, - prob=0.05, - ) - - self.balancer_ff3 = Balancer( - embed_dim, - channel_dim=-1, - min_positive=0.3, - max_positive=0.7, - min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0), - max_abs=4.0, - prob=0.05, - ) - - self.whiten = Whiten( - num_groups=1, - whitening_limit=_whitening_schedule(4.0, ratio=3.0), - prob=(0.025, 0.25), - grad_scale=0.01, - ) - - self.balancer2 = Balancer( - embed_dim, - channel_dim=-1, - min_positive=0.45, - max_positive=0.55, - min_abs=0.1, - max_abs=4.0, - ) - - def get_sequence_dropout_mask( - self, x: Tensor, dropout_rate: float - ) -> Optional[Tensor]: - if ( - dropout_rate == 0.0 - or not self.training - or torch.jit.is_scripting() - or torch.jit.is_tracing() - ): - return None - batch_size = x.shape[1] - mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype) - return mask - - def sequence_dropout(self, x: Tensor, dropout_rate: float) -> Tensor: - """ - Apply sequence-level dropout to x. - x shape: (seq_len, batch_size, embed_dim) - """ - dropout_mask = self.get_sequence_dropout_mask(x, dropout_rate) - if dropout_mask is None: - return x - else: - return x * dropout_mask - - def forward( - self, - src: Tensor, - pos_emb: Tensor, - chunk_size: int = -1, - attn_mask: Optional[Tensor] = None, - src_key_padding_mask: Optional[Tensor] = None, - ) -> Tensor: - """ - Pass the input through the encoder layer. - Args: - src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). - pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim) - chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. - feature_mask: something that broadcasts with src, that we'll multiply `src` - by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) - attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), - interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). - True means masked position. May be None. - src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means - masked position. May be None. - - Returns: - A tensor which has the same shape as src - """ - src_orig = src - - # dropout rate for non-feedforward submodules - if torch.jit.is_scripting() or torch.jit.is_tracing(): - attention_skip_rate = 0.0 - else: - attention_skip_rate = ( - float(self.attention_skip_rate) if self.training else 0.0 - ) - - # attn_weights: (num_heads, batch_size, seq_len, seq_len) - attn_weights = self.self_attn_weights( - src, - pos_emb=pos_emb, - attn_mask=attn_mask, - key_padding_mask=src_key_padding_mask, - ) - - src = src + self.feed_forward1(src) - - self_attn_dropout_mask = self.get_sequence_dropout_mask( - src, attention_skip_rate - ) - - selected_attn_weights = attn_weights[0:1] - if torch.jit.is_scripting() or torch.jit.is_tracing(): - pass - elif self.training and random.random() < float(self.const_attention_rate): - # Make attention weights constant. The intention is to - # encourage these modules to do something similar to an - # averaging-over-time operation. - # only need the mask, can just use the 1st one and expand later - selected_attn_weights = selected_attn_weights[0:1] - selected_attn_weights = (selected_attn_weights > 0.0).to( - selected_attn_weights.dtype - ) - selected_attn_weights = selected_attn_weights * ( - 1.0 / selected_attn_weights.sum(dim=-1, keepdim=True) - ) - - na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights)) - - src = src + ( - na if self_attn_dropout_mask is None else na * self_attn_dropout_mask - ) - - self_attn = self.self_attn1(src, attn_weights) - - src = src + ( - self_attn - if self_attn_dropout_mask is None - else self_attn * self_attn_dropout_mask - ) - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - conv_skip_rate = 0.0 - else: - conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout( - self.conv_module1( - src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask - ), - conv_skip_rate, - ) - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - ff2_skip_rate = 0.0 - else: - ff2_skip_rate = float(self.ff2_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout( - self.balancer_ff2(self.feed_forward2(src)), ff2_skip_rate - ) - - # bypass in the middle of the layer. - src = self.bypass_mid(src_orig, src) - - self_attn = self.self_attn2(src, attn_weights) - - src = src + ( - self_attn - if self_attn_dropout_mask is None - else self_attn * self_attn_dropout_mask - ) - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - conv_skip_rate = 0.0 - else: - conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout( - self.conv_module2( - src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask - ), - conv_skip_rate, - ) - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - ff3_skip_rate = 0.0 - else: - ff3_skip_rate = float(self.ff3_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout( - self.balancer_ff3(self.feed_forward3(src)), ff3_skip_rate - ) - - src = self.balancer1(src) - src = self.norm(src) - - src = self.bypass(src_orig, src) - - src = self.balancer2(src) - src = self.whiten(src) - - return src - - def streaming_forward( - self, - src: Tensor, - pos_emb: Tensor, - cached_key: Tensor, - cached_nonlin_attn: Tensor, - cached_val1: Tensor, - cached_val2: Tensor, - cached_conv1: Tensor, - cached_conv2: Tensor, - left_context_len: int, - src_key_padding_mask: Tensor, - ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]: - """Pass the input through the encoder layer in streaming forward mode. - - Args: - src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). - pos_emb: (1, left_context_len+2*seq_len-1, pos_emb_dim) or - (batch_size, left_context_len+2*seq_len-1, pos_emb_dim) - cached_key: cached attention key tensor of left context, - of shape (left_context_len, batch_size, key_dim) - cached_nonlin_attn: left context for nonlin_attention module, a Tensor of shape - (num_heads, batch_size, left_context_len, head_dim) - cached_val1: cached left context for the first attention module, - of shape (left_context_len, batch_size, value_dim) - cached_val2: cached left context for the second attention module, - of shape (left_context_len, batch_size, value_dim) - cached_conv1: cached left context for the first convolution module, - of shape (batch_size, channels, left_pad) - cached_conv2: cached left context for the second convolution module, - of shape (batch_size, channels, left_pad) - left_context_len: number of left context frames. - src_key_padding_mask: the mask for padding, of shape - (batch_size, left_context_len + seq_len); True means masked position. - May be None. - - Returns: - - x, with the same shape as src - - updated cached_key - - updated cached_nonlin_attn - - updated cached_val1 - - updated cached_val2 - - updated cached_conv1 - - updated cached_conv2 - """ - src_orig = src - - # attn_weights: (num_heads, batch_size, seq_len, seq_len) - attn_weights, cached_key = self.self_attn_weights.streaming_forward( - src, - pos_emb=pos_emb, - cached_key=cached_key, - left_context_len=left_context_len, - key_padding_mask=src_key_padding_mask, - ) - - src = src + self.feed_forward1(src) - - na, cached_nonlin_attn = self.nonlin_attention.streaming_forward( - src, - attn_weights[0:1], - cached_x=cached_nonlin_attn, - left_context_len=left_context_len, - ) - src = src + na - - self_attn, cached_val1 = self.self_attn1.streaming_forward( - src, - attn_weights=attn_weights, - cached_val=cached_val1, - left_context_len=left_context_len, - ) - src = src + self_attn - - src_conv, cached_conv1 = self.conv_module1.streaming_forward( - src, - cache=cached_conv1, - src_key_padding_mask=src_key_padding_mask[:, left_context_len:], - ) - src = src + src_conv - - src = src + self.feed_forward2(src) - - # bypass in the middle of the layer. - src = self.bypass_mid(src_orig, src) - - self_attn, cached_val2 = self.self_attn2.streaming_forward( - src, - attn_weights=attn_weights, - cached_val=cached_val2, - left_context_len=left_context_len, - ) - src = src + self_attn - - src_conv, cached_conv2 = self.conv_module2.streaming_forward( - src, - cache=cached_conv2, - src_key_padding_mask=src_key_padding_mask[:, left_context_len:], - ) - src = src + src_conv - - src = src + self.feed_forward3(src) - - src = self.norm(src) - - src = self.bypass(src_orig, src) - - return ( - src, - cached_key, - cached_nonlin_attn, - cached_val1, - cached_val2, - cached_conv1, - cached_conv2, - ) - - -class Zipformer2Encoder(nn.Module): - r"""Zipformer2Encoder is a stack of N encoder layers - - Args: - encoder_layer: an instance of the Zipformer2EncoderLayer() class (required). - num_layers: the number of sub-encoder-layers in the encoder (required). - pos_dim: the dimension for the relative positional encoding - - Examples:: - >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) - >>> zipformer_encoder = Zipformer2Encoder(encoder_layer, num_layers=6) - >>> src = torch.rand(10, 32, 512) - >>> out = zipformer_encoder(src) - """ - - def __init__( - self, - encoder_layer: nn.Module, - num_layers: int, - pos_dim: int, - dropout: float, - warmup_begin: float, - warmup_end: float, - initial_layerdrop_rate: float = 0.5, - final_layerdrop_rate: float = 0.05, - ) -> None: - super().__init__() - self.encoder_pos = CompactRelPositionalEncoding( - pos_dim, dropout_rate=0.15, length_factor=1.0 - ) - - self.layers = nn.ModuleList( - [copy.deepcopy(encoder_layer) for i in range(num_layers)] - ) - self.num_layers = num_layers - - assert 0 <= warmup_begin <= warmup_end, (warmup_begin, warmup_end) - - delta = (1.0 / num_layers) * (warmup_end - warmup_begin) - cur_begin = warmup_begin # interpreted as a training batch index - for i in range(num_layers): - cur_end = cur_begin + delta - self.layers[i].bypass.skip_rate = ScheduledFloat( - (cur_begin, initial_layerdrop_rate), - (cur_end, final_layerdrop_rate), - default=0.0, - ) - cur_begin = cur_end - - def forward( - self, - src: Tensor, - chunk_size: int = -1, - feature_mask: Union[Tensor, float] = 1.0, - attn_mask: Optional[Tensor] = None, - src_key_padding_mask: Optional[Tensor] = None, - ) -> Tensor: - r"""Pass the input through the encoder layers in turn. - - Args: - src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). - chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. - feature_mask: something that broadcasts with src, that we'll multiply `src` - by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) - attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), - interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). - True means masked position. May be None. - src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means - masked position. May be None. - - Returns: a Tensor with the same shape as src. - """ - pos_emb = self.encoder_pos(src) - output = src - - if not torch.jit.is_scripting() and not torch.jit.is_tracing(): - output = output * feature_mask - - for i, mod in enumerate(self.layers): - output = mod( - output, - pos_emb, - chunk_size=chunk_size, - attn_mask=attn_mask, - src_key_padding_mask=src_key_padding_mask, - ) - - if not torch.jit.is_scripting() and not torch.jit.is_tracing(): - output = output * feature_mask - - return output - - def streaming_forward( - self, - src: Tensor, - states: List[Tensor], - left_context_len: int, - src_key_padding_mask: Tensor, - ) -> Tuple[Tensor, List[Tensor]]: - r"""Pass the input through the encoder layers in turn. - - Args: - src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). - states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is - (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). - left_context_len: Number of left context frames. - src_key_padding_mask: the mask for padding, of shape - (batch_size, left_context_len + seq_len); True means masked position. - May be None. - - Returns: - - output, a Tensor with the same shape as src. - - updated states - """ - pos_emb = self.encoder_pos(src, left_context_len) - output = src - - new_states = [] - for i, mod in enumerate(self.layers): - ( - cached_key, - cached_nonlin_attn, - cached_val1, - cached_val2, - cached_conv1, - cached_conv2, - ) = states[i * 6 : (i + 1) * 6] - ( - output, - new_cached_key, - new_cached_nonlin_attn, - new_cached_val1, - new_cached_val2, - new_cached_conv1, - new_cached_conv2, - ) = mod.streaming_forward( - output, - pos_emb, - cached_key=cached_key, - cached_nonlin_attn=cached_nonlin_attn, - cached_val1=cached_val1, - cached_val2=cached_val2, - cached_conv1=cached_conv1, - cached_conv2=cached_conv2, - left_context_len=left_context_len, - src_key_padding_mask=src_key_padding_mask, - ) - new_states += [ - new_cached_key, - new_cached_nonlin_attn, - new_cached_val1, - new_cached_val2, - new_cached_conv1, - new_cached_conv2, - ] - - return output, new_states - - -class BypassModule(nn.Module): - """ - An nn.Module that implements a learnable bypass scale, and also randomized per-sequence - layer-skipping. The bypass is limited during early stages of training to be close to - "straight-through", i.e. to not do the bypass operation much initially, in order to - force all the modules to learn something. - """ - - def __init__( - self, - embed_dim: int, - skip_rate: FloatLike = 0.0, - straight_through_rate: FloatLike = 0.0, - scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0), - scale_max: FloatLike = 1.0, - ): - super().__init__() - self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) - self.skip_rate = copy.deepcopy(skip_rate) - self.straight_through_rate = copy.deepcopy(straight_through_rate) - self.scale_min = copy.deepcopy(scale_min) - self.scale_max = copy.deepcopy(scale_max) - - def _get_bypass_scale(self, batch_size: int): - # returns bypass-scale of shape (num_channels,), - # or (batch_size, num_channels,). This is actually the - # scale on the non-residual term, so 0 corresponds to bypassing - # this module. - if torch.jit.is_scripting() or torch.jit.is_tracing() or not self.training: - return self.bypass_scale - else: - ans = limit_param_value( - self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max) - ) - skip_rate = float(self.skip_rate) - if skip_rate != 0.0: - mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate - ans = ans * mask - # now ans is of shape (batch_size, num_channels), and is zero for sequences - # on which we have randomly chosen to do layer-skipping. - straight_through_rate = float(self.straight_through_rate) - if straight_through_rate != 0.0: - mask = ( - torch.rand((batch_size, 1), device=ans.device) - < straight_through_rate - ) - ans = torch.maximum(ans, mask.to(ans.dtype)) - return ans - - def forward(self, src_orig: Tensor, src: Tensor): - """ - Args: src_orig and src are both of shape (seq_len, batch_size, num_channels) - Returns: something with the same shape as src and src_orig - """ - bypass_scale = self._get_bypass_scale(src.shape[1]) - return src_orig + (src - src_orig) * bypass_scale - - -class DownsampledZipformer2Encoder(nn.Module): - r""" - DownsampledZipformer2Encoder is a zipformer encoder evaluated at a reduced frame rate, - after convolutional downsampling, and then upsampled again at the output, and combined - with the origin input, so that the output has the same shape as the input. - """ - - def __init__( - self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike - ): - super(DownsampledZipformer2Encoder, self).__init__() - self.downsample_factor = downsample - self.downsample = SimpleDownsample(dim, downsample, dropout) - self.num_layers = encoder.num_layers - self.encoder = encoder - self.upsample = SimpleUpsample(dim, downsample) - self.out_combiner = BypassModule(dim, straight_through_rate=0) - - def forward( - self, - src: Tensor, - chunk_size: int = -1, - feature_mask: Union[Tensor, float] = 1.0, - attn_mask: Optional[Tensor] = None, - src_key_padding_mask: Optional[Tensor] = None, - ) -> Tensor: - r"""Downsample, go through encoder, upsample. - - Args: - src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). - feature_mask: something that broadcasts with src, that we'll multiply `src` - by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) - attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), - interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). - True means masked position. May be None. - src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means - masked position. May be None. - - Returns: a Tensor with the same shape as src. - """ - src_orig = src - src = self.downsample(src) - ds = self.downsample_factor - if attn_mask is not None: - attn_mask = attn_mask[::ds, ::ds] - - src = self.encoder( - src, - chunk_size=chunk_size // ds, - feature_mask=feature_mask, - attn_mask=attn_mask, - src_key_padding_mask=src_key_padding_mask, - ) - src = self.upsample(src) - # remove any extra frames that are not a multiple of downsample_factor - src = src[: src_orig.shape[0]] - - return self.out_combiner(src_orig, src) - - def streaming_forward( - self, - src: Tensor, - states: List[Tensor], - left_context_len: int, - src_key_padding_mask: Tensor, - ) -> Tuple[Tensor, List[Tensor]]: - r"""Downsample, go through encoder, upsample, in streaming forward mode. - - Args: - src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). - states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is - (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). - left_context_len: Number of left context frames. - src_key_padding_mask: the mask for padding, of shape (batch_size, left_context_len+seq_len); - True means masked position. May be None. - - Returns: - - output, a Tensor with the same shape as src. - - updated states - """ - src_orig = src - src = self.downsample(src) - - src, new_states = self.encoder.streaming_forward( - src, - states=states, - left_context_len=left_context_len, - src_key_padding_mask=src_key_padding_mask, - ) - src = self.upsample(src) - # remove any extra frames that are not a multiple of downsample_factor - src = src[: src_orig.shape[0]] - - return self.out_combiner(src_orig, src), new_states - - -class SimpleDownsample(torch.nn.Module): - """ - Does downsampling with attention, by weighted sum, and a projection.. - """ - - def __init__(self, channels: int, downsample: int, dropout: FloatLike): - super(SimpleDownsample, self).__init__() - - self.bias = nn.Parameter(torch.zeros(downsample)) - - self.name = None # will be set from training code - self.dropout = copy.deepcopy(dropout) - - self.downsample = downsample - - def forward(self, src: Tensor) -> Tensor: - """ - x: (seq_len, batch_size, in_channels) - Returns a tensor of shape - ( (seq_len+downsample-1)//downsample, batch_size, channels) - """ - (seq_len, batch_size, in_channels) = src.shape - ds = self.downsample - d_seq_len = (seq_len + ds - 1) // ds - - # Pad to an exact multiple of self.downsample - # right-pad src, repeating the last element. - pad = d_seq_len * ds - seq_len - src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2]) - src = torch.cat((src, src_extra), dim=0) - assert src.shape[0] == d_seq_len * ds - - src = src.reshape(d_seq_len, ds, batch_size, in_channels) - - weights = self.bias.softmax(dim=0) - # weights: (downsample, 1, 1) - weights = weights.unsqueeze(-1).unsqueeze(-1) - - # ans1 is the first `in_channels` channels of the output - ans = (src * weights).sum(dim=1) - - return ans - - -class SimpleUpsample(torch.nn.Module): - """ - A very simple form of upsampling that mostly just repeats the input, but - also adds a position-specific bias. - """ - - def __init__(self, num_channels: int, upsample: int): - super(SimpleUpsample, self).__init__() - self.upsample = upsample - - def forward(self, src: Tensor) -> Tensor: - """ - x: (seq_len, batch_size, num_channels) - Returns a tensor of shape - ( (seq_len*upsample), batch_size, num_channels) - """ - upsample = self.upsample - (seq_len, batch_size, num_channels) = src.shape - src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels) - src = src.reshape(seq_len * upsample, batch_size, num_channels) - return src - - -class CompactRelPositionalEncoding(torch.nn.Module): - """ - Relative positional encoding module. This version is "compact" meaning it is able to encode - the important information about the relative position in a relatively small number of dimensions. - The goal is to make it so that small differences between large relative offsets (e.g. 1000 vs. 1001) - make very little difference to the embedding. Such differences were potentially important - when encoding absolute position, but not important when encoding relative position because there - is now no need to compare two large offsets with each other. - - Our embedding works by projecting the interval [-infinity,infinity] to a finite interval - using the atan() function, before doing the Fourier transform of that fixed interval. The - atan() function would compress the "long tails" too small, - making it hard to distinguish between different magnitudes of large offsets, so we use a logarithmic - function to compress large offsets to a smaller range before applying atan(). - Scalings are chosen in such a way that the embedding can clearly distinguish individual offsets as long - as they are quite close to the origin, e.g. abs(offset) <= about sqrt(embedding_dim) - - - Args: - embed_dim: Embedding dimension. - dropout_rate: Dropout rate. - max_len: Maximum input length: just a heuristic for initialization. - length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives - less weight to small differences of offset near the origin. - """ - - def __init__( - self, - embed_dim: int, - dropout_rate: FloatLike, - max_len: int = 1000, - length_factor: float = 1.0, - ) -> None: - """Construct a CompactRelPositionalEncoding object.""" - super(CompactRelPositionalEncoding, self).__init__() - self.embed_dim = embed_dim - assert embed_dim % 2 == 0, embed_dim - self.dropout = Dropout2(dropout_rate) - self.pe = None - assert length_factor >= 1.0, length_factor - self.length_factor = length_factor - self.extend_pe(torch.tensor(0.0).expand(max_len)) - - def extend_pe(self, x: Tensor, left_context_len: int = 0) -> None: - """Reset the positional encodings.""" - T = x.size(0) + left_context_len - - if self.pe is not None: - # self.pe contains both positive and negative parts - # the length of self.pe is 2 * input_len - 1 - if self.pe.size(0) >= T * 2 - 1: - self.pe = self.pe.to(dtype=x.dtype, device=x.device) - return - - # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ] - x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1) - - freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device) - - # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution - # for small time offsets but less resolution for large time offsets. - compression_length = self.embed_dim**0.5 - # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity; - # but it does so more slowly than T for large absolute values of T. - # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which - # is important. - x_compressed = ( - compression_length - * x.sign() - * ((x.abs() + compression_length).log() - math.log(compression_length)) - ) - - # if self.length_factor == 1.0, then length_scale is chosen so that the - # FFT can exactly separate points close to the origin (T == 0). So this - # part of the formulation is not really heuristic. - # But empirically, for ASR at least, length_factor > 1.0 seems to work better. - length_scale = self.length_factor * self.embed_dim / (2.0 * math.pi) - - # note for machine implementations: if atan is not available, we can use: - # x.sign() * ((1 / (x.abs() + 1)) - 1) * (-math.pi/2) - # check on wolframalpha.com: plot(sign(x) * (1 / ( abs(x) + 1) - 1 ) * -pi/2 , atan(x)) - x_atan = (x_compressed / length_scale).atan() # results between -pi and pi - - cosines = (x_atan * freqs).cos() - sines = (x_atan * freqs).sin() - - pe = torch.zeros(x.shape[0], self.embed_dim, device=x.device) - pe[:, 0::2] = cosines - pe[:, 1::2] = sines - pe[:, -1] = 1.0 # for bias. - - self.pe = pe.to(dtype=x.dtype) - - def forward(self, x: Tensor, left_context_len: int = 0) -> Tensor: - """Create positional encoding. - - Args: - x (Tensor): Input tensor (time, batch, `*`). - left_context_len: (int): Length of cached left context. - - Returns: - positional embedding, of shape (batch, left_context_len + 2*time-1, `*`). - """ - self.extend_pe(x, left_context_len) - x_size_left = x.size(0) + left_context_len - # length of positive side: x.size(0) + left_context_len - # length of negative side: x.size(0) - pos_emb = self.pe[ - self.pe.size(0) // 2 - - x_size_left - + 1 : self.pe.size(0) // 2 # noqa E203 - + x.size(0), - :, - ] - pos_emb = pos_emb.unsqueeze(0) - return self.dropout(pos_emb) - - -class RelPositionMultiheadAttentionWeights(nn.Module): - r"""Module that computes multi-head attention weights with relative position encoding. - Various other modules consume the resulting attention weights: see, for example, the - SimpleAttention module which allows you to compute conventional attention. - - This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context", - we have to write up the differences. - - - Args: - embed_dim: number of channels at the input to this module, e.g. 256 - pos_dim: dimension of the positional encoding vectors, e.g. 128. - num_heads: number of heads to compute weights for, e.g. 8 - query_head_dim: dimension of the query (and key), per head. e.g. 24. - pos_head_dim: dimension of the projected positional encoding per head, e.g. 4. - dropout: dropout probability for attn_output_weights. Default: 0.0. - pos_emb_skip_rate: probability for skipping the pos_emb part of the scores on - any given call to forward(), in training time. - """ - - def __init__( - self, - embed_dim: int, - pos_dim: int, - num_heads: int, - query_head_dim: int, - pos_head_dim: int, - dropout: float = 0.0, - pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)), - ) -> None: - super().__init__() - self.embed_dim = embed_dim - self.num_heads = num_heads - self.query_head_dim = query_head_dim - self.pos_head_dim = pos_head_dim - self.dropout = dropout - self.pos_emb_skip_rate = copy.deepcopy(pos_emb_skip_rate) - self.name = None # will be overwritten in training code; for diagnostics. - - key_head_dim = query_head_dim - in_proj_dim = (query_head_dim + key_head_dim + pos_head_dim) * num_heads - - # the initial_scale is supposed to take over the "scaling" factor of - # head_dim ** -0.5 that has been used in previous forms of attention, - # dividing it between the query and key. Note: this module is intended - # to be used with the ScaledAdam optimizer; with most other optimizers, - # it would be necessary to apply the scaling factor in the forward function. - self.in_proj = ScaledLinear( - embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25 - ) - - self.whiten_keys = Whiten( - num_groups=num_heads, - whitening_limit=_whitening_schedule(3.0), - prob=(0.025, 0.25), - grad_scale=0.025, - ) - - # add a balancer for the keys that runs with very small probability, and - # tries to enforce that all dimensions have mean around zero. The - # weights produced by this module are invariant to adding a constant to - # the keys, so the derivative of the bias is mathematically zero; but - # due to how Adam/ScaledAdam work, it can learn a fairly large nonzero - # bias because the small numerical roundoff tends to have a non-random - # sign. This module is intended to prevent that. Use a very small - # probability; that should be sufficient to fix the problem. - self.balance_keys = Balancer( - key_head_dim * num_heads, - channel_dim=-1, - min_positive=0.4, - max_positive=0.6, - min_abs=0.0, - max_abs=100.0, - prob=0.025, - ) - - # linear transformation for positional encoding. - self.linear_pos = ScaledLinear( - pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05 - ) - - # the following are for diagnostics only, see --print-diagnostics option - self.copy_pos_query = Identity() - self.copy_query = Identity() - - def forward( - self, - x: Tensor, - pos_emb: Tensor, - key_padding_mask: Optional[Tensor] = None, - attn_mask: Optional[Tensor] = None, - ) -> Tensor: - r""" - Args: - x: input of shape (seq_len, batch_size, embed_dim) - pos_emb: Positional embedding tensor, of shape (1, 2*seq_len - 1, pos_dim) - key_padding_mask: a bool tensor of shape (batch_size, seq_len). Positions that - are True in this mask will be ignored as sources in the attention weighting. - attn_mask: mask of shape (seq_len, seq_len) or (batch_size, seq_len, seq_len), - interpreted as ([batch_size,] tgt_seq_len, src_seq_len) - saying which positions are allowed to attend to which other positions. - Returns: - a tensor of attention weights, of shape (hum_heads, batch_size, seq_len, seq_len) - interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len). - """ - x = self.in_proj(x) - query_head_dim = self.query_head_dim - pos_head_dim = self.pos_head_dim - num_heads = self.num_heads - - seq_len, batch_size, _ = x.shape - - query_dim = query_head_dim * num_heads - - # self-attention - q = x[..., 0:query_dim] - k = x[..., query_dim : 2 * query_dim] - # p is the position-encoding query - p = x[..., 2 * query_dim :] - assert p.shape[-1] == num_heads * pos_head_dim, (p.shape[-1], num_heads, pos_head_dim) - - q = self.copy_query(q) # for diagnostics only, does nothing. - k = self.whiten_keys(self.balance_keys(k)) # does nothing in the forward pass. - p = self.copy_pos_query(p) # for diagnostics only, does nothing. - - q = q.reshape(seq_len, batch_size, num_heads, query_head_dim) - p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim) - k = k.reshape(seq_len, batch_size, num_heads, query_head_dim) - - # time1 refers to target, time2 refers to source. - q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) - p = p.permute(2, 1, 0, 3) # (head, batch, time1, pos_head_dim) - k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) - - attn_scores = torch.matmul(q, k) - - use_pos_scores = False - if torch.jit.is_scripting() or torch.jit.is_tracing(): - # We can't put random.random() in the same line - use_pos_scores = True - elif not self.training or random.random() >= float(self.pos_emb_skip_rate): - use_pos_scores = True - - if use_pos_scores: - pos_emb = self.linear_pos(pos_emb) - seq_len2 = 2 * seq_len - 1 - pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( - 2, 0, 3, 1 - ) - # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) - - # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) - # [where seq_len2 represents relative position.] - pos_scores = torch.matmul(p, pos_emb) - # the following .as_strided() expression converts the last axis of pos_scores from relative - # to absolute position. I don't know whether I might have got the time-offsets backwards or - # not, but let this code define which way round it is supposed to be. - if torch.jit.is_tracing(): - (num_heads, batch_size, time1, n) = pos_scores.shape - rows = torch.arange(start=time1 - 1, end=-1, step=-1) - cols = torch.arange(seq_len) - rows = rows.repeat(batch_size * num_heads).unsqueeze(-1) - indexes = rows + cols - pos_scores = pos_scores.reshape(-1, n) - pos_scores = torch.gather(pos_scores, dim=1, index=indexes) - pos_scores = pos_scores.reshape(num_heads, batch_size, time1, seq_len) - else: - pos_scores = pos_scores.as_strided( - (num_heads, batch_size, seq_len, seq_len), - ( - pos_scores.stride(0), - pos_scores.stride(1), - pos_scores.stride(2) - pos_scores.stride(3), - pos_scores.stride(3), - ), - storage_offset=pos_scores.stride(3) * (seq_len - 1), - ) - - attn_scores = attn_scores + pos_scores - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - pass - elif self.training and random.random() < 0.1: - # This is a harder way of limiting the attention scores to not be - # too large. It incurs a penalty if any of them has an absolute - # value greater than 50.0. this should be outside the normal range - # of the attention scores. We use this mechanism instead of, say, - # something added to the loss function involving the entropy, - # because once the entropy gets very small gradients through the - # softmax can become very small, and we'd get zero derivatives. The - # choices of 1.0e-04 as the scale on the penalty makes this - # mechanism vulnerable to the absolute scale of the loss function, - # but we view this as a failsafe to avoid "implausible" parameter - # values rather than a regularization method that should be active - # under normal circumstances. - attn_scores = penalize_abs_values_gt( - attn_scores, limit=25.0, penalty=1.0e-04, name=self.name - ) - - assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len) - - if attn_mask is not None: - assert attn_mask.dtype == torch.bool - # use -1000 to avoid nan's where attn_mask and key_padding_mask make - # all scores zero. It's important that this be large enough that exp(-1000) - # is exactly zero, for reasons related to const_attention_rate, it - # compares the final weights with zero. - attn_scores = attn_scores.masked_fill(attn_mask, -1000) - - if key_padding_mask is not None: - assert key_padding_mask.shape == ( - batch_size, - seq_len, - ), key_padding_mask.shape - attn_scores = attn_scores.masked_fill( - key_padding_mask.unsqueeze(1), - -1000, - ) - - # We use our own version of softmax, defined in scaling.py, which should - # save a little of the memory used in backprop by, if we are in - # automatic mixed precision mode (amp / autocast), by only storing the - # half-precision output for backprop purposes. - attn_weights = softmax(attn_scores, dim=-1) - - if torch.jit.is_scripting() or torch.jit.is_tracing(): - pass - elif random.random() < 0.001 and not self.training: - self._print_attn_entropy(attn_weights) - - attn_weights = nn.functional.dropout( - attn_weights, p=self.dropout, training=self.training - ) - - return attn_weights - - def streaming_forward( - self, - x: Tensor, - pos_emb: Tensor, - cached_key: Tensor, - left_context_len: int, - key_padding_mask: Tensor, - ) -> Tuple[Tensor, Tensor]: - r""" - Args: - x: input of shape (seq_len, batch_size, embed_dim) - pos_emb: Positional embedding tensor, of shape (1, left_context_len+2*seq_len-1, pos_dim) - cached_key: cached attention key tensor of left context, - of shape (left_context_len, batch_size, key_dim) - left_context_len: number of left context frames. - key_padding_mask: a bool tensor of shape (batch_size, seq_len). Positions that - are True in this mask will be ignored as sources in the attention weighting. - - Returns: - - attention weights, of shape (hum_heads, batch_size, seq_len, seq_len2), - interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len). - - updated cached attention key tensor of left context. - """ - x = self.in_proj(x) - query_head_dim = self.query_head_dim - pos_head_dim = self.pos_head_dim - num_heads = self.num_heads - - seq_len, batch_size, _ = x.shape - - query_dim = query_head_dim * num_heads - - # self-attention - q = x[..., 0:query_dim] - k = x[..., query_dim : 2 * query_dim] - # p is the position-encoding query - p = x[..., 2 * query_dim :] - assert p.shape[-1] == num_heads * pos_head_dim - - # Pad cached left contexts - assert cached_key.shape[0] == left_context_len, ( - cached_key.shape[0], - left_context_len, - ) - k = torch.cat([cached_key, k], dim=0) - # Update cached left contexts - cached_key = k[-left_context_len:, ...] - - # The length of key - k_len = k.shape[0] - - q = q.reshape(seq_len, batch_size, num_heads, query_head_dim) - p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim) - k = k.reshape(k_len, batch_size, num_heads, query_head_dim) - - # time1 refers to target, time2 refers to source. - q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) - p = p.permute(2, 1, 0, 3) # (head, batch, time1, pos_head_dim) - k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) - - attn_scores = torch.matmul(q, k) - - pos_emb = self.linear_pos(pos_emb) - seq_len2 = 2 * seq_len - 1 + left_context_len - pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( - 2, 0, 3, 1 - ) - # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) - - # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) - # [where seq_len2 represents relative position.] - pos_scores = torch.matmul(p, pos_emb) - - if torch.jit.is_tracing(): - (num_heads, batch_size, time1, n) = pos_scores.shape - rows = torch.arange(start=time1 - 1, end=-1, step=-1) - cols = torch.arange(k_len) - rows = rows.repeat(batch_size * num_heads).unsqueeze(-1) - indexes = rows + cols - pos_scores = pos_scores.reshape(-1, n) - pos_scores = torch.gather(pos_scores, dim=1, index=indexes) - pos_scores = pos_scores.reshape(num_heads, batch_size, time1, k_len) - # the following .as_strided() expression converts the last axis of pos_scores from relative - # to absolute position. I don't know whether I might have got the time-offsets backwards or - # not, but let this code define which way round it is supposed to be. - else: - pos_scores = pos_scores.as_strided( - (num_heads, batch_size, seq_len, k_len), - ( - pos_scores.stride(0), - pos_scores.stride(1), - pos_scores.stride(2) - pos_scores.stride(3), - pos_scores.stride(3), - ), - storage_offset=pos_scores.stride(3) * (seq_len - 1), - ) - - attn_scores = attn_scores + pos_scores - - assert attn_scores.shape == ( - num_heads, - batch_size, - seq_len, - k_len, - ), attn_scores.shape - - if key_padding_mask is not None: - assert key_padding_mask.shape == (batch_size, k_len), key_padding_mask.shape - attn_scores = attn_scores.masked_fill( - key_padding_mask.unsqueeze(1), - -1000, - ) - - attn_weights = attn_scores.softmax(dim=-1) - - return attn_weights, cached_key - - def _print_attn_entropy(self, attn_weights: Tensor): - # attn_weights: (num_heads, batch_size, seq_len, seq_len) - (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape - - with torch.no_grad(): - with torch.cuda.amp.autocast(enabled=False): - attn_weights = attn_weights.to(torch.float32) - attn_weights_entropy = ( - -((attn_weights + 1.0e-20).log() * attn_weights) - .sum(dim=-1) - .mean(dim=(1, 2)) - ) - logging.info( - f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}" - ) - - -class SelfAttention(nn.Module): - """ - The simplest possible attention module. This one works with already-computed attention - weights, e.g. as computed by RelPositionMultiheadAttentionWeights. - - Args: - embed_dim: the input and output embedding dimension - num_heads: the number of attention heads - value_head_dim: the value dimension per head - """ - - def __init__( - self, - embed_dim: int, - num_heads: int, - value_head_dim: int, - ) -> None: - super().__init__() - self.in_proj = nn.Linear(embed_dim, num_heads * value_head_dim, bias=True) - - self.out_proj = ScaledLinear( - num_heads * value_head_dim, embed_dim, bias=True, initial_scale=0.05 - ) - - self.whiten = Whiten( - num_groups=1, - whitening_limit=_whitening_schedule(7.5, ratio=3.0), - prob=(0.025, 0.25), - grad_scale=0.01, - ) - - def forward( - self, - x: Tensor, - attn_weights: Tensor, - ) -> Tensor: - """ - Args: - x: input tensor, of shape (seq_len, batch_size, embed_dim) - attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len), - with seq_len being interpreted as (tgt_seq_len, src_seq_len). Expect - attn_weights.sum(dim=-1) == 1. - Returns: - a tensor with the same shape as x. - """ - (seq_len, batch_size, embed_dim) = x.shape - num_heads = attn_weights.shape[0] - assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len) - - x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) - x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) - # now x: (num_heads, batch_size, seq_len, value_head_dim) - value_head_dim = x.shape[-1] - - # todo: see whether there is benefit in overriding matmul - x = torch.matmul(attn_weights, x) - # v: (num_heads, batch_size, seq_len, value_head_dim) - - x = ( - x.permute(2, 1, 0, 3) - .contiguous() - .view(seq_len, batch_size, num_heads * value_head_dim) - ) - - # returned value is of shape (seq_len, batch_size, embed_dim), like the input. - x = self.out_proj(x) - x = self.whiten(x) - - return x - - def streaming_forward( - self, - x: Tensor, - attn_weights: Tensor, - cached_val: Tensor, - left_context_len: int, - ) -> Tuple[Tensor, Tensor]: - """ - Args: - x: input tensor, of shape (seq_len, batch_size, embed_dim) - attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len), - with seq_len being interpreted as (tgt_seq_len, src_seq_len). Expect - attn_weights.sum(dim=-1) == 1. - cached_val: cached attention value tensor of left context, - of shape (left_context_len, batch_size, value_dim) - left_context_len: number of left context frames. - - Returns: - - attention weighted output, a tensor with the same shape as x. - - updated cached attention value tensor of left context. - """ - (seq_len, batch_size, embed_dim) = x.shape - num_heads = attn_weights.shape[0] - seq_len2 = seq_len + left_context_len - assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len2) - - x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) - - # Pad cached left contexts - assert cached_val.shape[0] == left_context_len, ( - cached_val.shape[0], - left_context_len, - ) - x = torch.cat([cached_val, x], dim=0) - # Update cached left contexts - cached_val = x[-left_context_len:, ...] - - x = x.reshape(seq_len2, batch_size, num_heads, -1).permute(2, 1, 0, 3) - # now x: (num_heads, batch_size, seq_len, value_head_dim) - value_head_dim = x.shape[-1] - - # todo: see whether there is benefit in overriding matmul - x = torch.matmul(attn_weights, x) - # v: (num_heads, batch_size, seq_len, value_head_dim) - - x = ( - x.permute(2, 1, 0, 3) - .contiguous() - .view(seq_len, batch_size, num_heads * value_head_dim) - ) - - # returned value is of shape (seq_len, batch_size, embed_dim), like the input. - x = self.out_proj(x) - - return x, cached_val - - -class FeedforwardModule(nn.Module): - """Feedforward module in Zipformer2 model.""" - - def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike): - super(FeedforwardModule, self).__init__() - self.in_proj = nn.Linear(embed_dim, feedforward_dim) - - self.hidden_balancer = Balancer( - feedforward_dim, - channel_dim=-1, - min_positive=0.3, - max_positive=1.0, - min_abs=0.75, - max_abs=5.0, - ) - - # shared_dim=0 means we share the dropout mask along the time axis - self.out_proj = ActivationDropoutAndLinear( - feedforward_dim, - embed_dim, - activation="SwooshL", - dropout_p=dropout, - dropout_shared_dim=0, - bias=True, - initial_scale=0.1, - ) - - self.out_whiten = Whiten( - num_groups=1, - whitening_limit=_whitening_schedule(7.5), - prob=(0.025, 0.25), - grad_scale=0.01, - ) - - def forward(self, x: Tensor): - x = self.in_proj(x) - x = self.hidden_balancer(x) - # out_proj contains SwooshL activation, then dropout, then linear. - x = self.out_proj(x) - x = self.out_whiten(x) - return x - - -class NonlinAttention(nn.Module): - """This is like the ConvolutionModule, but refactored so that we use multiplication by attention weights (borrowed - from the attention module) in place of actual convolution. We also took out the second nonlinearity, the - one after the attention mechanism. - - Args: - channels (int): The number of channels of conv layers. - """ - - def __init__( - self, - channels: int, - hidden_channels: int, - ) -> None: - super().__init__() - - self.hidden_channels = hidden_channels - - self.in_proj = nn.Linear(channels, hidden_channels * 3, bias=True) - - # balancer that goes before the sigmoid. Have quite a large min_abs value, at 2.0, - # because we noticed that well-trained instances of this module have abs-value before the sigmoid - # starting from about 3, and poorly-trained instances of the module have smaller abs values - # before the sigmoid. - self.balancer = Balancer( - hidden_channels, - channel_dim=-1, - min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)), - max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)), - min_abs=0.5, - max_abs=5.0, - ) - self.tanh = nn.Tanh() - - self.identity1 = Identity() # for diagnostics. - self.identity2 = Identity() # for diagnostics. - self.identity3 = Identity() # for diagnostics. - - self.out_proj = ScaledLinear( - hidden_channels, channels, bias=True, initial_scale=0.05 - ) - - self.whiten1 = Whiten( - num_groups=1, - whitening_limit=_whitening_schedule(5.0), - prob=(0.025, 0.25), - grad_scale=0.01, - ) - - self.whiten2 = Whiten( - num_groups=1, - whitening_limit=_whitening_schedule(5.0, ratio=3.0), - prob=(0.025, 0.25), - grad_scale=0.01, - ) - - def forward( - self, - x: Tensor, - attn_weights: Tensor, - ) -> Tensor: - """. - Args: - x: a Tensor of shape (seq_len, batch_size, num_channels) - attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) - Returns: - a Tensor with the same shape as x - """ - x = self.in_proj(x) - - (seq_len, batch_size, _) = x.shape - hidden_channels = self.hidden_channels - - s, x, y = x.chunk(3, dim=2) - - # s will go through tanh. - - s = self.balancer(s) - s = self.tanh(s) - - s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) - x = self.whiten1(x) - x = x * s - x = self.identity1(x) # diagnostics only, it's the identity. - - (seq_len, batch_size, embed_dim) = x.shape - num_heads = attn_weights.shape[0] - assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len) - - x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) - # now x: (num_heads, batch_size, seq_len, head_dim) - x = torch.matmul(attn_weights, x) - # now x: (num_heads, batch_size, seq_len, head_dim) - x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) - - y = self.identity2(y) - x = x * y - x = self.identity3(x) - - x = self.out_proj(x) - x = self.whiten2(x) - return x - - def streaming_forward( - self, - x: Tensor, - attn_weights: Tensor, - cached_x: Tensor, - left_context_len: int, - ) -> Tuple[Tensor, Tensor]: - """. - Args: - x: a Tensor of shape (seq_len, batch_size, num_channels) - attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) - cached_x: left context, a Tensor of shape - (num_heads, batch_size, left_context_len, head_dim) - left_context_len: number of left context frames. - Returns: - - a Tensor with the same shape as x - - updated left context with same shape as cached_x - """ - x = self.in_proj(x) - - (seq_len, batch_size, _) = x.shape - hidden_channels = self.hidden_channels - - s, x, y = x.chunk(3, dim=2) - - # s will go through tanh. - s = self.tanh(s) - - s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) - x = x * s - - (seq_len, batch_size, embed_dim) = x.shape - num_heads = attn_weights.shape[0] - assert attn_weights.shape == ( - num_heads, - batch_size, - seq_len, - left_context_len + seq_len, - ) - - x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) - # now x: (num_heads, batch_size, seq_len, head_dim) - - # Pad cached tensor - assert cached_x.shape[2] == left_context_len, ( - cached_x.shape[2], - left_context_len, - ) - x_pad = torch.cat([cached_x, x], dim=2) - # Update cached tensor - cached_x = x_pad[:, :, -left_context_len:, :] - - x = torch.matmul(attn_weights, x_pad) - # now x: (num_heads, batch_size, seq_len, head_dim) - x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) - - x = x * y - - x = self.out_proj(x) - return x, cached_x - - -class ConvolutionModule(nn.Module): - """ConvolutionModule in Zipformer2 model. - Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py - - Args: - channels (int): The number of channels of conv layers. - kernel_size (int): Kernerl size of conv layers. - bias (bool): Whether to use bias in conv layers (default=True). - - """ - - def __init__( - self, - channels: int, - kernel_size: int, - causal: bool, - ) -> None: - """Construct a ConvolutionModule object.""" - super(ConvolutionModule, self).__init__() - # kernerl_size should be a odd number for 'SAME' padding - assert (kernel_size - 1) % 2 == 0 - - bottleneck_dim = channels - self.causal = causal - - self.in_proj = nn.Linear( - channels, - 2 * bottleneck_dim, - ) - # the gradients on in_proj are a little noisy, likely to do with the - # sigmoid in glu. - - # after in_proj 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.balancer1 = Balancer( - bottleneck_dim, - channel_dim=-1, - min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)), - max_positive=1.0, - min_abs=1.5, - max_abs=ScheduledFloat((0.0, 5.0), (8000.0, 10.0), default=1.0), - ) - - self.activation1 = Identity() # for diagnostics - - self.sigmoid = nn.Sigmoid() - - self.activation2 = Identity() # for diagnostics - - assert kernel_size % 2 == 1 - - self.depthwise_conv = ( - ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size) - if causal - else nn.Conv1d( - in_channels=bottleneck_dim, - out_channels=bottleneck_dim, - groups=bottleneck_dim, - kernel_size=kernel_size, - padding=kernel_size // 2, - ) - ) - - self.balancer2 = Balancer( - bottleneck_dim, - channel_dim=1, - min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)), - max_positive=1.0, - min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)), - max_abs=10.0, - ) - - self.whiten = Whiten( - num_groups=1, - whitening_limit=_whitening_schedule(7.5), - prob=(0.025, 0.25), - grad_scale=0.01, - ) - - self.out_proj = ActivationDropoutAndLinear( - bottleneck_dim, - channels, - activation="SwooshR", - dropout_p=0.0, - initial_scale=0.05, - ) - - def forward( - self, - x: Tensor, - src_key_padding_mask: Optional[Tensor] = None, - chunk_size: int = -1, - ) -> Tensor: - """Compute convolution module. - - Args: - x: Input tensor (#time, batch, channels). - src_key_padding_mask: the mask for the src keys per batch (optional): - (batch, #time), contains True in masked positions. - - Returns: - Tensor: Output tensor (#time, batch, channels). - - """ - - x = self.in_proj(x) # (time, batch, 2*channels) - - x, s = x.chunk(2, dim=2) - s = self.balancer1(s) - s = self.sigmoid(s) - x = self.activation1(x) # identity. - x = x * s - x = self.activation2(x) # identity - - # (time, batch, channels) - - # exchange the temporal dimension and the feature dimension - x = x.permute(1, 2, 0) # (#batch, channels, time). - - if src_key_padding_mask is not None: - x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) - - if ( - not torch.jit.is_scripting() - and not torch.jit.is_tracing() - and chunk_size >= 0 - ): - # Not support exporting a model for simulated streaming decoding - assert ( - self.causal - ), "Must initialize model with causal=True if you use chunk_size" - x = self.depthwise_conv(x, chunk_size=chunk_size) - else: - x = self.depthwise_conv(x) - - x = self.balancer2(x) - x = x.permute(2, 0, 1) # (time, batch, channels) - - x = self.whiten(x) # (time, batch, channels) - x = self.out_proj(x) # (time, batch, channels) - - return x - - def streaming_forward( - self, - x: Tensor, - cache: Tensor, - src_key_padding_mask: Tensor, - ) -> Tuple[Tensor, Tensor]: - """Compute convolution module in streaming forward mode. - - Args: - x: Input tensor (#time, batch, channels). - cache: cached left context for depthwise_conv of shape - (#batch, channels, left_pad) - src_key_padding_mask: the mask for the src keys per batch (optional): - (batch, #time), contains True in masked positions. - - Returns: - - Output tensor (#time, batch, channels). - - Updated cache (#batch, channels, left_pad) - """ - - x = self.in_proj(x) # (time, batch, 2*channels) - - x, s = x.chunk(2, dim=2) - s = self.sigmoid(s) - x = x * s - # (time, batch, channels) - - # exchange the temporal dimension and the feature dimension - x = x.permute(1, 2, 0) # (#batch, channels, time). - - if src_key_padding_mask is not None: - x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) - - x, cache = self.depthwise_conv.streaming_forward(x, cache=cache) - - x = x.permute(2, 0, 1) # (time, batch, channels) - - x = self.out_proj(x) # (time, batch, channels) - - return x, cache - - -class ScalarMultiply(nn.Module): - def __init__(self, scale: float): - super().__init__() - self.scale = scale - - def forward(self, x): - return x * self.scale - - -def _test_zipformer_main(causal: bool = False): - batch_size = 5 - seq_len = 20 - # Just make sure the forward pass runs. - - c = Zipformer2( - encoder_dim=(64, 96), - encoder_unmasked_dim=(48, 64), - num_heads=(4, 4), - causal=causal, - chunk_size=(4,) if causal else (-1,), - left_context_frames=(64,), - ) - batch_size = 5 - seq_len = 20 - # Just make sure the forward pass runs. - f = c( - torch.randn(seq_len, batch_size, 64), - torch.full((batch_size,), seq_len, dtype=torch.int64), - ) - f[0].sum().backward() - c.eval() - f = c( - torch.randn(seq_len, batch_size, 64), - torch.full((batch_size,), seq_len, dtype=torch.int64), - ) - f # to remove flake8 warnings - - -if __name__ == "__main__": - logging.getLogger().setLevel(logging.INFO) - torch.set_num_threads(1) - torch.set_num_interop_threads(1) - _test_zipformer_main(False) - _test_zipformer_main(True)