# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. # # 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. """Contains U2 model.""" import json import os import time from collections import defaultdict from collections import OrderedDict from contextlib import nullcontext from typing import Optional import jsonlines import numpy as np import paddle from paddle import distributed as dist from paddle.io import DataLoader from yacs.config import CfgNode from deepspeech.io.collator import SpeechCollator from deepspeech.io.dataset import ManifestDataset from deepspeech.io.sampler import SortagradBatchSampler from deepspeech.io.sampler import SortagradDistributedBatchSampler from deepspeech.models.u2 import U2Model from deepspeech.training.optimizer import OptimizerFactory from deepspeech.training.reporter import ObsScope from deepspeech.training.reporter import report from deepspeech.training.scheduler import LRSchedulerFactory from deepspeech.training.timer import Timer from deepspeech.training.trainer import Trainer from deepspeech.utils import ctc_utils from deepspeech.utils import error_rate from deepspeech.utils import layer_tools from deepspeech.utils import mp_tools from deepspeech.utils.log import Log from deepspeech.utils.utility import UpdateConfig logger = Log(__name__).getlog() class U2Trainer(Trainer): @classmethod def params(cls, config: Optional[CfgNode]=None) -> CfgNode: # training config default = CfgNode( dict( n_epoch=50, # train epochs log_interval=100, # steps accum_grad=1, # accum grad by # steps global_grad_clip=5.0, # the global norm clip )) default.optim = 'adam' default.optim_conf = CfgNode( dict( lr=5e-4, # learning rate weight_decay=1e-6, # the coeff of weight decay )) default.scheduler = 'warmuplr' default.scheduler_conf = CfgNode( dict( warmup_steps=25000, lr_decay=1.0, # learning rate decay )) if config is not None: config.merge_from_other_cfg(default) return default def __init__(self, config, args): super().__init__(config, args) def train_batch(self, batch_index, batch_data, msg): train_conf = self.config.training start = time.time() # forward utt, audio, audio_len, text, text_len = batch_data loss, attention_loss, ctc_loss = self.model(audio, audio_len, text, text_len) # loss div by `batch_size * accum_grad` loss /= train_conf.accum_grad losses_np = {'loss': float(loss) * train_conf.accum_grad} if attention_loss: losses_np['att_loss'] = float(attention_loss) if ctc_loss: losses_np['ctc_loss'] = float(ctc_loss) # loss backward if (batch_index + 1) % train_conf.accum_grad != 0: # Disable gradient synchronizations across DDP processes. # Within this context, gradients will be accumulated on module # variables, which will later be synchronized. # When using cpu w/o DDP, model does not have `no_sync` context = self.model.no_sync if (hasattr(self.model, "no_sync") and self.parallel) else nullcontext else: # Used for single gpu training and DDP gradient synchronization # processes. context = nullcontext with context(): loss.backward() layer_tools.print_grads(self.model, print_func=None) # optimizer step if (batch_index + 1) % train_conf.accum_grad == 0: self.optimizer.step() self.optimizer.clear_grad() self.lr_scheduler.step() self.iteration += 1 iteration_time = time.time() - start for k, v in losses_np.items(): report(k, v) report("batch_size", self.config.collator.batch_size) report("accum", train_conf.accum_grad) report("step_cost", iteration_time) if (batch_index + 1) % train_conf.accum_grad == 0: if dist.get_rank() == 0 and self.visualizer: losses_np_v = losses_np.copy() losses_np_v.update({"lr": self.lr_scheduler()}) self.visualizer.add_scalars("step", losses_np_v, self.iteration - 1) @paddle.no_grad() def valid(self): self.model.eval() logger.info(f"Valid Total Examples: {len(self.valid_loader.dataset)}") valid_losses = defaultdict(list) num_seen_utts = 1 total_loss = 0.0 for i, batch in enumerate(self.valid_loader): utt, audio, audio_len, text, text_len = batch loss, attention_loss, ctc_loss = self.model(audio, audio_len, text, text_len) if paddle.isfinite(loss): num_utts = batch[1].shape[0] num_seen_utts += num_utts total_loss += float(loss) * num_utts valid_losses['val_loss'].append(float(loss)) if attention_loss: valid_losses['val_att_loss'].append(float(attention_loss)) if ctc_loss: valid_losses['val_ctc_loss'].append(float(ctc_loss)) if (i + 1) % self.config.training.log_interval == 0: valid_dump = {k: np.mean(v) for k, v in valid_losses.items()} valid_dump['val_history_loss'] = total_loss / num_seen_utts # logging msg = f"Valid: Rank: {dist.get_rank()}, " msg += "epoch: {}, ".format(self.epoch) msg += "step: {}, ".format(self.iteration) msg += "batch: {}/{}, ".format(i + 1, len(self.valid_loader)) msg += ', '.join('{}: {:>.6f}'.format(k, v) for k, v in valid_dump.items()) logger.info(msg) logger.info('Rank {} Val info val_loss {}'.format( dist.get_rank(), total_loss / num_seen_utts)) return total_loss, num_seen_utts def train(self): """The training process control by step.""" # !!!IMPORTANT!!! # Try to export the model by script, if fails, we should refine # the code to satisfy the script export requirements # script_model = paddle.jit.to_static(self.model) # script_model_path = str(self.checkpoint_dir / 'init') # paddle.jit.save(script_model, script_model_path) self.before_train() logger.info(f"Train Total Examples: {len(self.train_loader.dataset)}") while self.epoch < self.config.training.n_epoch: with Timer("Epoch-Train Time Cost: {}"): self.model.train() try: data_start_time = time.time() for batch_index, batch in enumerate(self.train_loader): dataload_time = time.time() - data_start_time msg = "Train:" observation = OrderedDict() with ObsScope(observation): report("Rank", dist.get_rank()) report("epoch", self.epoch) report('step', self.iteration) report("lr", self.lr_scheduler()) self.train_batch(batch_index, batch, msg) self.after_train_batch() report('iter', batch_index + 1) report('total', len(self.train_loader)) report('reader_cost', dataload_time) observation['batch_cost'] = observation[ 'reader_cost'] + observation['step_cost'] observation['samples'] = observation['batch_size'] observation['ips[sent./sec]'] = observation[ 'batch_size'] / observation['batch_cost'] for k, v in observation.items(): msg += f" {k}: " msg += f"{v:>.8f}" if isinstance(v, float) else f"{v}" msg += "," msg = msg[:-1] # remove the last "," if (batch_index + 1 ) % self.config.training.log_interval == 0: logger.info(msg) data_start_time = time.time() except Exception as e: logger.error(e) raise e with Timer("Eval Time Cost: {}"): total_loss, num_seen_utts = self.valid() if dist.get_world_size() > 1: num_seen_utts = paddle.to_tensor(num_seen_utts) # the default operator in all_reduce function is sum. dist.all_reduce(num_seen_utts) total_loss = paddle.to_tensor(total_loss) dist.all_reduce(total_loss) cv_loss = total_loss / num_seen_utts cv_loss = float(cv_loss) else: cv_loss = total_loss / num_seen_utts logger.info( 'Epoch {} Val info val_loss {}'.format(self.epoch, cv_loss)) if self.visualizer: self.visualizer.add_scalars( 'epoch', {'cv_loss': cv_loss, 'lr': self.lr_scheduler()}, self.epoch) self.save(tag=self.epoch, infos={'val_loss': cv_loss}) self.new_epoch() def setup_dataloader(self): config = self.config.clone() config.defrost() config.collator.keep_transcription_text = False # train/valid dataset, return token ids config.data.manifest = config.data.train_manifest train_dataset = ManifestDataset.from_config(config) config.data.manifest = config.data.dev_manifest dev_dataset = ManifestDataset.from_config(config) collate_fn_train = SpeechCollator.from_config(config) config.collator.augmentation_config = "" collate_fn_dev = SpeechCollator.from_config(config) if self.parallel: batch_sampler = SortagradDistributedBatchSampler( train_dataset, batch_size=config.collator.batch_size, num_replicas=None, rank=None, shuffle=True, drop_last=True, sortagrad=config.collator.sortagrad, shuffle_method=config.collator.shuffle_method) else: batch_sampler = SortagradBatchSampler( train_dataset, shuffle=True, batch_size=config.collator.batch_size, drop_last=True, sortagrad=config.collator.sortagrad, shuffle_method=config.collator.shuffle_method) self.train_loader = DataLoader( train_dataset, batch_sampler=batch_sampler, collate_fn=collate_fn_train, num_workers=config.collator.num_workers, ) self.valid_loader = DataLoader( dev_dataset, batch_size=config.collator.batch_size, shuffle=False, drop_last=False, collate_fn=collate_fn_dev, num_workers=config.collator.num_workers, ) # test dataset, return raw text config.data.manifest = config.data.test_manifest # filter test examples, will cause less examples, but no mismatch with training # and can use large batch size , save training time, so filter test egs now. config.data.min_input_len = 0.0 # second config.data.max_input_len = float('inf') # second config.data.min_output_len = 0.0 # tokens config.data.max_output_len = float('inf') # tokens config.data.min_output_input_ratio = 0.00 config.data.max_output_input_ratio = float('inf') test_dataset = ManifestDataset.from_config(config) # return text ord id config.collator.keep_transcription_text = True config.collator.augmentation_config = "" self.test_loader = DataLoader( test_dataset, batch_size=config.decoding.batch_size, shuffle=False, drop_last=False, collate_fn=SpeechCollator.from_config(config), num_workers=config.collator.num_workers, ) # return text token id config.collator.keep_transcription_text = False self.align_loader = DataLoader( test_dataset, batch_size=config.decoding.batch_size, shuffle=False, drop_last=False, collate_fn=SpeechCollator.from_config(config), num_workers=config.collator.num_workers, ) logger.info("Setup train/valid/test/align Dataloader!") def setup_model(self): config = self.config model_conf = config.model with UpdateConfig(model_conf): model_conf.input_dim = self.train_loader.collate_fn.feature_size model_conf.output_dim = self.train_loader.collate_fn.vocab_size model = U2Model.from_config(model_conf) if self.parallel: model = paddle.DataParallel(model) logger.info(f"{model}") layer_tools.print_params(model, logger.info) train_config = config.training optim_type = train_config.optim optim_conf = train_config.optim_conf scheduler_type = train_config.scheduler scheduler_conf = train_config.scheduler_conf scheduler_args = { "learning_rate": optim_conf.lr, "verbose": False, "warmup_steps": scheduler_conf.warmup_steps, "gamma": scheduler_conf.lr_decay, "d_model": model_conf.encoder_conf.output_size, } lr_scheduler = LRSchedulerFactory.from_args(scheduler_type, scheduler_args) def optimizer_args( config, parameters, lr_scheduler=None, ): train_config = config.training optim_type = train_config.optim optim_conf = train_config.optim_conf scheduler_type = train_config.scheduler scheduler_conf = train_config.scheduler_conf return { "grad_clip": train_config.global_grad_clip, "weight_decay": optim_conf.weight_decay, "learning_rate": lr_scheduler if lr_scheduler else optim_conf.lr, "parameters": parameters, "epsilon": 1e-9 if optim_type == 'noam' else None, "beta1": 0.9 if optim_type == 'noam' else None, "beat2": 0.98 if optim_type == 'noam' else None, } optimzer_args = optimizer_args(config, model.parameters(), lr_scheduler) optimizer = OptimizerFactory.from_args(optim_type, optimzer_args) self.model = model self.optimizer = optimizer self.lr_scheduler = lr_scheduler logger.info("Setup model/optimizer/lr_scheduler!") class U2Tester(U2Trainer): @classmethod def params(cls, config: Optional[CfgNode]=None) -> CfgNode: # decoding config default = CfgNode( dict( alpha=2.5, # Coef of LM for beam search. beta=0.3, # Coef of WC for beam search. cutoff_prob=1.0, # Cutoff probability for pruning. cutoff_top_n=40, # Cutoff number for pruning. lang_model_path='models/lm/common_crawl_00.prune01111.trie.klm', # Filepath for language model. decoding_method='attention', # Decoding method. Options: 'attention', 'ctc_greedy_search', # 'ctc_prefix_beam_search', 'attention_rescoring' error_rate_type='wer', # Error rate type for evaluation. Options `wer`, 'cer' num_proc_bsearch=8, # # of CPUs for beam search. beam_size=10, # Beam search width. batch_size=16, # decoding batch size ctc_weight=0.0, # ctc weight for attention rescoring decode mode. decoding_chunk_size=-1, # decoding chunk size. Defaults to -1. # <0: for decoding, use full chunk. # >0: for decoding, use fixed chunk size as set. # 0: used for training, it's prohibited here. num_decoding_left_chunks=-1, # number of left chunks for decoding. Defaults to -1. simulate_streaming=False, # simulate streaming inference. Defaults to False. )) if config is not None: config.merge_from_other_cfg(default) return default def __init__(self, config, args): super().__init__(config, args) def ordid2token(self, texts, texts_len): """ ord() id to chr() chr """ trans = [] for text, n in zip(texts, texts_len): n = n.numpy().item() ids = text[:n] trans.append(''.join([chr(i) for i in ids])) return trans def compute_metrics(self, utts, audio, audio_len, texts, texts_len, fout=None): cfg = self.config.decoding errors_sum, len_refs, num_ins = 0.0, 0, 0 errors_func = error_rate.char_errors if cfg.error_rate_type == 'cer' else error_rate.word_errors error_rate_func = error_rate.cer if cfg.error_rate_type == 'cer' else error_rate.wer start_time = time.time() text_feature = self.test_loader.collate_fn.text_feature target_transcripts = self.ordid2token(texts, texts_len) result_transcripts, result_tokenids = self.model.decode( audio, audio_len, text_feature=text_feature, decoding_method=cfg.decoding_method, lang_model_path=cfg.lang_model_path, beam_alpha=cfg.alpha, beam_beta=cfg.beta, beam_size=cfg.beam_size, cutoff_prob=cfg.cutoff_prob, cutoff_top_n=cfg.cutoff_top_n, num_processes=cfg.num_proc_bsearch, ctc_weight=cfg.ctc_weight, decoding_chunk_size=cfg.decoding_chunk_size, num_decoding_left_chunks=cfg.num_decoding_left_chunks, simulate_streaming=cfg.simulate_streaming) decode_time = time.time() - start_time for utt, target, result, rec_tids in zip( utts, target_transcripts, result_transcripts, result_tokenids): errors, len_ref = errors_func(target, result) errors_sum += errors len_refs += len_ref num_ins += 1 if fout: fout.write({ "utt": utt, "refs": [target], "hyps": [result], "hyps_tokenid": [rec_tids], }) logger.info(f"Utt: {utt}") logger.info(f"Ref: {target}") logger.info(f"Hyp: {result}") logger.info("One example error rate [%s] = %f" % (cfg.error_rate_type, error_rate_func(target, result))) return dict( errors_sum=errors_sum, len_refs=len_refs, num_ins=num_ins, # num examples error_rate=errors_sum / len_refs, error_rate_type=cfg.error_rate_type, num_frames=audio_len.sum().numpy().item(), decode_time=decode_time) @mp_tools.rank_zero_only @paddle.no_grad() def test(self): assert self.args.result_file self.model.eval() logger.info(f"Test Total Examples: {len(self.test_loader.dataset)}") stride_ms = self.test_loader.collate_fn.stride_ms error_rate_type = None errors_sum, len_refs, num_ins = 0.0, 0, 0 num_frames = 0.0 num_time = 0.0 with jsonlines.open(self.args.result_file, 'w') as fout: for i, batch in enumerate(self.test_loader): metrics = self.compute_metrics(*batch, fout=fout) num_frames += metrics['num_frames'] num_time += metrics["decode_time"] errors_sum += metrics['errors_sum'] len_refs += metrics['len_refs'] num_ins += metrics['num_ins'] error_rate_type = metrics['error_rate_type'] rtf = num_time / (num_frames * stride_ms) logger.info( "RTF: %f, Error rate [%s] (%d/?) = %f" % (rtf, error_rate_type, num_ins, errors_sum / len_refs)) rtf = num_time / (num_frames * stride_ms) msg = "Test: " msg += "epoch: {}, ".format(self.epoch) msg += "step: {}, ".format(self.iteration) msg += "RTF: {}, ".format(rtf) msg += "Final error rate [%s] (%d/%d) = %f" % ( error_rate_type, num_ins, num_ins, errors_sum / len_refs) logger.info(msg) # test meta results err_meta_path = os.path.splitext(self.args.result_file)[0] + '.err' err_type_str = "{}".format(error_rate_type) with open(err_meta_path, 'w') as f: data = json.dumps({ "epoch": self.epoch, "step": self.iteration, "rtf": rtf, error_rate_type: errors_sum / len_refs, "dataset_hour": (num_frames * stride_ms) / 1000.0 / 3600.0, "process_hour": num_time / 1000.0 / 3600.0, "num_examples": num_ins, "err_sum": errors_sum, "ref_len": len_refs, "decode_method": self.config.decoding.decoding_method, }) f.write(data + '\n') @paddle.no_grad() def align(self): ctc_utils.ctc_align( self.model, self.align_loader, self.config.decoding.batch_size, self.align_loader.collate_fn.stride_ms, self.align_loader.collate_fn.vocab_list, self.args.result_file) def load_inferspec(self): """infer model and input spec. Returns: nn.Layer: inference model List[paddle.static.InputSpec]: input spec. """ from deepspeech.models.u2 import U2InferModel infer_model = U2InferModel.from_pretrained(self.test_loader, self.config.model.clone(), self.args.checkpoint_path) feat_dim = self.test_loader.collate_fn.feature_size input_spec = [ paddle.static.InputSpec(shape=[1, None, feat_dim], dtype='float32'), # audio, [B,T,D] paddle.static.InputSpec(shape=[1], dtype='int64'), # audio_length, [B] ] return infer_model, input_spec @paddle.no_grad() def export(self): infer_model, input_spec = self.load_inferspec() assert isinstance(input_spec, list), type(input_spec) infer_model.eval() static_model = paddle.jit.to_static(infer_model, input_spec=input_spec) logger.info(f"Export code: {static_model.forward.code}") paddle.jit.save(static_model, self.args.export_path)