def __init__( self, lang_dir: Pathlike, scripted_model_path: Optional[Pathlike] = None, model_dir: Optional[Pathlike] = None, average_epochs: Sequence[int] = (7, 8, 9), device: torch.device = 'cpu', sampling_rate: int = 16000, ): if isinstance(device, str): self.device = torch.device(device) self.sampling_rate = sampling_rate self.extractor = Fbank(FbankConfig(num_mel_bins=80)) self.lexicon = Lexicon(lang_dir) phone_ids = self.lexicon.phone_symbols() self.P = create_bigram_phone_lm(phone_ids) if model_dir is not None: # Read model from regular checkpoints, assume it's a Conformer self.model = Conformer(num_features=80, num_classes=len(phone_ids) + 1, num_decoder_layers=0) self.P.scores = torch.zeros_like(self.P.scores) self.model.P_scores = torch.nn.Parameter(self.P.scores.clone(), requires_grad=False) average_checkpoint(filenames=[ model_dir / f'epoch-{n}.pt' for n in average_epochs ], model=self.model) elif scripted_model_path is not None: # Read model from a serialized TorchScript module, no assumptions needed self.model = torch.jit.load(scripted_model_path) else: raise ValueError( "One of scripted_model_path or model_dir needs to be provided." ) # Freeze the params by default. for p in self.model.parameters(): p.requires_grad_(False) self.compiler = MmiTrainingGraphCompiler(lexicon=self.lexicon, device=self.device) self.HLG = k2.Fsa.from_dict(torch.load(lang_dir / 'HLG.pt')).to( self.device)
def main(): exp_dir = Path('exp-lstm-adam-mmi-mbr-musan') setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug') # load L, G, symbol_table lang_dir = Path('data/lang_nosp') symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt') phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt') phone_ids = get_phone_symbols(phone_symbol_table) P = create_bigram_phone_lm(phone_ids) phone_ids_with_blank = [0] + phone_ids ctc_topo = k2.arc_sort(build_ctc_topo(phone_ids_with_blank)) logging.debug("About to load model") # Note: Use "export CUDA_VISIBLE_DEVICES=N" to setup device id to N # device = torch.device('cuda', 1) device = torch.device('cuda') model = TdnnLstm1b( num_features=40, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=3) model.P_scores = torch.nn.Parameter(P.scores.clone(), requires_grad=False) checkpoint = os.path.join(exp_dir, 'epoch-9.pt') load_checkpoint(checkpoint, model) model.to(device) model.eval() assert P.requires_grad is False P.scores = model.P_scores.cpu() print_transition_probabilities(P, phone_symbol_table, phone_ids, filename='model_P_scores.txt') P.set_scores_stochastic_(model.P_scores) print_transition_probabilities(P, phone_symbol_table, phone_ids, filename='P_scores.txt') if not os.path.exists(lang_dir / 'HLG.pt'): logging.debug("Loading L_disambig.fst.txt") with open(lang_dir / 'L_disambig.fst.txt') as f: L = k2.Fsa.from_openfst(f.read(), acceptor=False) logging.debug("Loading G.fst.txt") with open(lang_dir / 'G.fst.txt') as f: G = k2.Fsa.from_openfst(f.read(), acceptor=False) first_phone_disambig_id = find_first_disambig_symbol( phone_symbol_table) first_word_disambig_id = find_first_disambig_symbol(symbol_table) HLG = compile_HLG(L=L, G=G, H=ctc_topo, labels_disambig_id_start=first_phone_disambig_id, aux_labels_disambig_id_start=first_word_disambig_id) torch.save(HLG.as_dict(), lang_dir / 'HLG.pt') else: logging.debug("Loading pre-compiled HLG") d = torch.load(lang_dir / 'HLG.pt') HLG = k2.Fsa.from_dict(d) # load dataset feature_dir = Path('exp/data') logging.debug("About to get test cuts") cuts_test = CutSet.from_json(feature_dir / 'cuts_test-clean.json.gz') logging.info("About to create test dataset") test = K2SpeechRecognitionDataset(cuts_test) sampler = SingleCutSampler(cuts_test, max_frames=100000) logging.info("About to create test dataloader") test_dl = torch.utils.data.DataLoader(test, batch_size=None, sampler=sampler, num_workers=1) # if not torch.cuda.is_available(): # logging.error('No GPU detected!') # sys.exit(-1) logging.debug("convert HLG to device") HLG = HLG.to(device) HLG.aux_labels = k2.ragged.remove_values_eq(HLG.aux_labels, 0) HLG.requires_grad_(False) logging.debug("About to decode") results = decode(dataloader=test_dl, model=model, device=device, HLG=HLG, symbols=symbol_table) s = '' for ref, hyp in results: s += f'ref={ref}\n' s += f'hyp={hyp}\n' logging.info(s) # compute WER dists = [edit_distance(r, h) for r, h in results] errors = { key: sum(dist[key] for dist in dists) for key in ['sub', 'ins', 'del', 'total'] } total_words = sum(len(ref) for ref, _ in results) # Print Kaldi-like message: # %WER 8.20 [ 4459 / 54402, 695 ins, 427 del, 3337 sub ] logging.info( f'%WER {errors["total"] / total_words:.2%} ' f'[{errors["total"]} / {total_words}, {errors["ins"]} ins, {errors["del"]} del, {errors["sub"]} sub ]' )
def main(): fix_random_seed(42) exp_dir = f'exp-lstm-adam-mmi-mbr-musan' setup_logger('{}/log/log-train'.format(exp_dir)) tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') if not torch.cuda.is_available(): logging.warn('No GPU detected!') logging.warn('USE CPU (very slow)!') device = torch.device('cpu') else: logging.info('Use GPU') device_id = 0 device = torch.device('cuda', device_id) # load L, G, symbol_table lang_dir = Path('data/lang_nosp') phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt') word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt') logging.info("Loading L.fst") if (lang_dir / 'Linv.pt').exists(): logging.info('Loading precompiled L') L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt')) else: logging.info('Compiling L') with open(lang_dir / 'L.fst.txt') as f: L = k2.Fsa.from_openfst(f.read(), acceptor=False) L_inv = k2.arc_sort(L.invert_()) torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt') logging.info("Loading L_disambig.fst") if (lang_dir / 'L_disambig.pt').exists(): logging.info('Loading precompiled L_disambig') L_disambig = k2.Fsa.from_dict(torch.load(lang_dir / 'L_disambig.pt')) else: logging.info('Compiling L_disambig') with open(lang_dir / 'L_disambig.fst.txt') as f: L_disambig = k2.Fsa.from_openfst(f.read(), acceptor=False) L_disambig = k2.arc_sort(L_disambig) torch.save(L_disambig.as_dict(), lang_dir / 'L_disambig.pt') logging.info("Loading G.fst") if (lang_dir / 'G_uni.pt').exists(): logging.info('Loading precompiled G') G = k2.Fsa.from_dict(torch.load(lang_dir / 'G_uni.pt')) else: logging.info('Compiling G') with open(lang_dir / 'G_uni.fst.txt') as f: G = k2.Fsa.from_openfst(f.read(), acceptor=False) G = k2.arc_sort(G) torch.save(G.as_dict(), lang_dir / 'G_uni.pt') graph_compiler = MmiMbrTrainingGraphCompiler(L_inv=L_inv, L_disambig=L_disambig, G=G, device=device, phones=phone_symbol_table, words=word_symbol_table) phone_ids = get_phone_symbols(phone_symbol_table) P = create_bigram_phone_lm(phone_ids) P.scores = torch.zeros_like(P.scores) # load dataset feature_dir = Path('exp/data') logging.info("About to get train cuts") cuts_train = CutSet.from_json(feature_dir / 'cuts_train-clean-100.json.gz') logging.info("About to get dev cuts") cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev-clean.json.gz') logging.info("About to get Musan cuts") cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz') logging.info("About to create train dataset") train = K2SpeechRecognitionIterableDataset(cuts_train, max_frames=30000, shuffle=True, aug_cuts=cuts_musan, aug_prob=0.5, aug_snr=(10, 20)) logging.info("About to create dev dataset") validate = K2SpeechRecognitionIterableDataset(cuts_dev, max_frames=60000, shuffle=False, concat_cuts=False) logging.info("About to create train dataloader") train_dl = torch.utils.data.DataLoader(train, batch_size=None, num_workers=4) logging.info("About to create dev dataloader") valid_dl = torch.utils.data.DataLoader(validate, batch_size=None, num_workers=1) logging.info("About to create model") model = TdnnLstm1b( num_features=40, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=3) model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True) start_epoch = 0 num_epochs = 10 best_objf = np.inf best_valid_objf = np.inf best_epoch = start_epoch best_model_path = os.path.join(exp_dir, 'best_model.pt') best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info') global_batch_idx_train = 0 # for logging only use_adam = True if start_epoch > 0: model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(start_epoch - 1)) ckpt = load_checkpoint(filename=model_path, model=model) best_objf = ckpt['objf'] best_valid_objf = ckpt['valid_objf'] global_batch_idx_train = ckpt['global_batch_idx_train'] logging.info( f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}" ) model.to(device) describe(model) P = P.to(device) if use_adam: learning_rate = 1e-3 weight_decay = 5e-4 optimizer = optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay) # Equivalent to the following in the epoch loop: # if epoch > 6: # curr_learning_rate *= 0.8 lr_scheduler = optim.lr_scheduler.LambdaLR( optimizer, lambda ep: 1.0 if ep < 7 else 0.8**(ep - 6)) else: learning_rate = 5e-5 weight_decay = 1e-5 momentum = 0.9 lr_schedule_gamma = 0.7 optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=momentum, weight_decay=weight_decay) lr_scheduler = optim.lr_scheduler.ExponentialLR( optimizer=optimizer, gamma=lr_schedule_gamma, last_epoch=start_epoch - 1) for epoch in range(start_epoch, num_epochs): # LR scheduler can hold multiple learning rates for multiple parameter groups; # For now we report just the first LR which we assume concerns most of the parameters. curr_learning_rate = lr_scheduler.get_last_lr()[0] tb_writer.add_scalar('train/learning_rate', curr_learning_rate, global_batch_idx_train) tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train) logging.info('epoch {}, learning rate {}'.format( epoch, curr_learning_rate)) objf, valid_objf, global_batch_idx_train = train_one_epoch( dataloader=train_dl, valid_dataloader=valid_dl, model=model, P=P, device=device, graph_compiler=graph_compiler, optimizer=optimizer, current_epoch=epoch, tb_writer=tb_writer, num_epochs=num_epochs, global_batch_idx_train=global_batch_idx_train, ) # the lower, the better if valid_objf < best_valid_objf: best_valid_objf = valid_objf best_objf = objf best_epoch = epoch save_checkpoint(filename=best_model_path, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train) save_training_info(filename=best_epoch_info_filename, model_path=best_model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch) # we always save the model for every epoch model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch)) save_checkpoint(filename=model_path, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train) epoch_info_filename = os.path.join(exp_dir, 'epoch-{}-info'.format(epoch)) save_training_info(filename=epoch_info_filename, model_path=model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch) lr_scheduler.step() logging.warning('Done')
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() ''' model_type = args.model_type start_epoch = args.start_epoch num_epochs = args.num_epochs accum_grad = args.accum_grad den_scale = args.den_scale att_rate = args.att_rate fix_random_seed(42) setup_dist(rank, world_size, args.master_port) exp_dir = Path('exp-' + model_type + '-noam-mmi-att-musan-sa-vgg') setup_logger(f'{exp_dir}/log/log-train-{rank}') if args.tensorboard and rank == 0: tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') else: tb_writer = None # tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') if args.tensorboard and rank == 0 else None logging.info("Loading lexicon and symbol tables") lang_dir = Path('data/lang_nosp') lexicon = Lexicon(lang_dir) device_id = rank device = torch.device('cuda', device_id) graph_compiler = MmiTrainingGraphCompiler( lexicon=lexicon, device=device, ) phone_ids = lexicon.phone_symbols() P = create_bigram_phone_lm(phone_ids) P.scores = torch.zeros_like(P.scores) P = P.to(device) mls = MLSAsrDataModule(args) train_dl = mls.train_dataloaders() valid_dl = mls.valid_dataloaders() if not torch.cuda.is_available(): logging.error('No GPU detected!') sys.exit(-1) logging.info("About to create model") if att_rate != 0.0: num_decoder_layers = 6 else: num_decoder_layers = 0 if model_type == "transformer": model = Transformer( num_features=80, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers, vgg_frontend=True) elif model_type == "conformer": model = Conformer( num_features=80, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers, vgg_frontend=True) elif model_type == "contextnet": model = ContextNet(num_features=80, num_classes=len(phone_ids) + 1) # +1 for the blank symbol else: raise NotImplementedError("Model of type " + str(model_type) + " is not implemented") model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True) model.to(device) describe(model) model = DDP(model, device_ids=[rank]) # Now for the aligment model, if any if args.use_ali_model: ali_model = TdnnLstm1b( num_features=80, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4) ali_model_fname = Path( f'exp-lstm-adam-ctc-musan/epoch-{args.ali_model_epoch}.pt') assert ali_model_fname.is_file(), \ f'ali model filename {ali_model_fname} does not exist!' ali_model.load_state_dict( torch.load(ali_model_fname, map_location='cpu')['state_dict']) ali_model.to(device) ali_model.eval() ali_model.requires_grad_(False) logging.info(f'Use ali_model: {ali_model_fname}') else: ali_model = None logging.info('No ali_model') optimizer = Noam(model.parameters(), model_size=args.attention_dim, factor=args.lr_factor, warm_step=args.warm_step, weight_decay=args.weight_decay) scaler = GradScaler(enabled=args.amp) best_objf = np.inf best_valid_objf = np.inf best_epoch = start_epoch best_model_path = os.path.join(exp_dir, 'best_model.pt') best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info') global_batch_idx_train = 0 # for logging only if start_epoch > 0: model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(start_epoch - 1)) ckpt = load_checkpoint(filename=model_path, model=model, optimizer=optimizer, scaler=scaler) best_objf = ckpt['objf'] best_valid_objf = ckpt['valid_objf'] global_batch_idx_train = ckpt['global_batch_idx_train'] logging.info( f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}" ) for epoch in range(start_epoch, num_epochs): train_dl.sampler.set_epoch(epoch) curr_learning_rate = optimizer._rate if tb_writer is not None: tb_writer.add_scalar('train/learning_rate', curr_learning_rate, global_batch_idx_train) tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train) logging.info('epoch {}, learning rate {}'.format( epoch, curr_learning_rate)) objf, valid_objf, global_batch_idx_train = train_one_epoch( dataloader=train_dl, valid_dataloader=valid_dl, model=model, ali_model=ali_model, P=P, device=device, graph_compiler=graph_compiler, optimizer=optimizer, accum_grad=accum_grad, den_scale=den_scale, att_rate=att_rate, current_epoch=epoch, tb_writer=tb_writer, num_epochs=num_epochs, global_batch_idx_train=global_batch_idx_train, world_size=world_size, scaler=scaler) # the lower, the better if valid_objf < best_valid_objf: best_valid_objf = valid_objf best_objf = objf best_epoch = epoch save_checkpoint(filename=best_model_path, optimizer=None, scheduler=None, scaler=None, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train, local_rank=rank) save_training_info(filename=best_epoch_info_filename, model_path=best_model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch, local_rank=rank) # we always save the model for every epoch model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch)) save_checkpoint(filename=model_path, optimizer=optimizer, scheduler=None, scaler=scaler, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train, local_rank=rank) epoch_info_filename = os.path.join(exp_dir, 'epoch-{}-info'.format(epoch)) save_training_info(filename=epoch_info_filename, model_path=model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch, local_rank=rank) logging.warning('Done') torch.distributed.barrier() cleanup_dist()
def main(): parser = get_parser() LibriSpeechAsrDataModule.add_arguments(parser) args = parser.parse_args() model_type = args.model_type epoch = args.epoch avg = args.avg att_rate = args.att_rate num_paths = args.num_paths use_lm_rescoring = args.use_lm_rescoring use_whole_lattice = False if use_lm_rescoring and num_paths < 1: # It doesn't make sense to use n-best list for rescoring # when n is less than 1 use_whole_lattice = True output_beam_size = args.output_beam_size exp_dir = Path('exp-' + model_type + '-noam-mmi-att-musan-sa-vgg') setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug') logging.info(f'output_beam_size: {output_beam_size}') # load L, G, symbol_table lang_dir = Path('data/lang_nosp') symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt') phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt') phone_ids = get_phone_symbols(phone_symbol_table) P = create_bigram_phone_lm(phone_ids) phone_ids_with_blank = [0] + phone_ids ctc_topo = k2.arc_sort(build_ctc_topo(phone_ids_with_blank)) logging.debug("About to load model") # Note: Use "export CUDA_VISIBLE_DEVICES=N" to setup device id to N # device = torch.device('cuda', 1) device = torch.device('cuda') if att_rate != 0.0: num_decoder_layers = 6 else: num_decoder_layers = 0 if model_type == "transformer": model = Transformer( num_features=80, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers, vgg_frontend=True) elif model_type == "conformer": model = Conformer( num_features=80, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers, vgg_frontend=True) elif model_type == "contextnet": model = ContextNet(num_features=80, num_classes=len(phone_ids) + 1) # +1 for the blank symbol else: raise NotImplementedError("Model of type " + str(model_type) + " is not implemented") model.P_scores = torch.nn.Parameter(P.scores.clone(), requires_grad=False) if avg == 1: checkpoint = os.path.join(exp_dir, 'epoch-' + str(epoch - 1) + '.pt') load_checkpoint(checkpoint, model) else: checkpoints = [ os.path.join(exp_dir, 'epoch-' + str(avg_epoch) + '.pt') for avg_epoch in range(epoch - avg, epoch) ] average_checkpoint(checkpoints, model) model.to(device) model.eval() assert P.requires_grad is False P.scores = model.P_scores.cpu() print_transition_probabilities(P, phone_symbol_table, phone_ids, filename='model_P_scores.txt') P.set_scores_stochastic_(model.P_scores) print_transition_probabilities(P, phone_symbol_table, phone_ids, filename='P_scores.txt') if not os.path.exists(lang_dir / 'HLG.pt'): logging.debug("Loading L_disambig.fst.txt") with open(lang_dir / 'L_disambig.fst.txt') as f: L = k2.Fsa.from_openfst(f.read(), acceptor=False) logging.debug("Loading G.fst.txt") with open(lang_dir / 'G.fst.txt') as f: G = k2.Fsa.from_openfst(f.read(), acceptor=False) first_phone_disambig_id = find_first_disambig_symbol( phone_symbol_table) first_word_disambig_id = find_first_disambig_symbol(symbol_table) HLG = compile_HLG(L=L, G=G, H=ctc_topo, labels_disambig_id_start=first_phone_disambig_id, aux_labels_disambig_id_start=first_word_disambig_id) torch.save(HLG.as_dict(), lang_dir / 'HLG.pt') else: logging.debug("Loading pre-compiled HLG") d = torch.load(lang_dir / 'HLG.pt') HLG = k2.Fsa.from_dict(d) if use_lm_rescoring: if use_whole_lattice: logging.info('Rescoring with the whole lattice') else: logging.info(f'Rescoring with n-best list, n is {num_paths}') first_word_disambig_id = find_first_disambig_symbol(symbol_table) if not os.path.exists(lang_dir / 'G_4_gram.pt'): logging.debug('Loading G_4_gram.fst.txt') with open(lang_dir / 'G_4_gram.fst.txt') as f: G = k2.Fsa.from_openfst(f.read(), acceptor=False) # G.aux_labels is not needed in later computations, so # remove it here. del G.aux_labels # CAUTION(fangjun): The following line is crucial. # Arcs entering the back-off state have label equal to #0. # We have to change it to 0 here. G.labels[G.labels >= first_word_disambig_id] = 0 G = k2.create_fsa_vec([G]).to(device) G = k2.arc_sort(G) torch.save(G.as_dict(), lang_dir / 'G_4_gram.pt') else: logging.debug('Loading pre-compiled G_4_gram.pt') d = torch.load(lang_dir / 'G_4_gram.pt') G = k2.Fsa.from_dict(d).to(device) if use_whole_lattice: # Add epsilon self-loops to G as we will compose # it with the whole lattice later G = k2.add_epsilon_self_loops(G) G = k2.arc_sort(G) G = G.to(device) else: logging.debug('Decoding without LM rescoring') G = None logging.debug("convert HLG to device") HLG = HLG.to(device) HLG.aux_labels = k2.ragged.remove_values_eq(HLG.aux_labels, 0) HLG.requires_grad_(False) if not hasattr(HLG, 'lm_scores'): HLG.lm_scores = HLG.scores.clone() # load dataset librispeech = LibriSpeechAsrDataModule(args) test_sets = ['test-clean', 'test-other'] # test_sets = ['test-other'] for test_set, test_dl in zip(test_sets, librispeech.test_dataloaders()): logging.info(f'* DECODING: {test_set}') results = decode(dataloader=test_dl, model=model, device=device, HLG=HLG, symbols=symbol_table, num_paths=num_paths, G=G, use_whole_lattice=use_whole_lattice, output_beam_size=output_beam_size) recog_path = exp_dir / f'recogs-{test_set}.txt' store_transcripts(path=recog_path, texts=results) logging.info(f'The transcripts are stored in {recog_path}') # The following prints out WERs, per-word error statistics and aligned # ref/hyp pairs. errs_filename = exp_dir / f'errs-{test_set}.txt' with open(errs_filename, 'w') as f: write_error_stats(f, test_set, results) logging.info('Wrote detailed error stats to {}'.format(errs_filename))
def main(): parser = get_parser() AishellAsrDataModule.add_arguments(parser) args = parser.parse_args() model_type = args.model_type epoch = args.epoch avg = args.avg att_rate = args.att_rate exp_dir = Path('exp-' + model_type + '-noam-mmi-att-musan') setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug') # load L, G, symbol_table lang_dir = Path('data/lang_nosp') symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt') phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt') phone_ids = get_phone_symbols(phone_symbol_table) P = create_bigram_phone_lm(phone_ids) phone_ids_with_blank = [0] + phone_ids ctc_topo = k2.arc_sort(build_ctc_topo(phone_ids_with_blank)) logging.debug("About to load model") # Note: Use "export CUDA_VISIBLE_DEVICES=N" to setup device id to N # device = torch.device('cuda', 1) device = torch.device('cuda') if att_rate != 0.0: num_decoder_layers = 6 else: num_decoder_layers = 0 if model_type == "transformer": model = Transformer( num_features=40, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers) else: model = Conformer( num_features=40, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers) model.P_scores = torch.nn.Parameter(P.scores.clone(), requires_grad=False) if avg == 1: checkpoint = os.path.join(exp_dir, 'epoch-' + str(epoch - 1) + '.pt') load_checkpoint(checkpoint, model) else: checkpoints = [os.path.join(exp_dir, 'epoch-' + str(avg_epoch) + '.pt') for avg_epoch in range(epoch - avg, epoch)] average_checkpoint(checkpoints, model) model.to(device) model.eval() assert P.requires_grad is False P.scores = model.P_scores.cpu() print_transition_probabilities(P, phone_symbol_table, phone_ids, filename='model_P_scores.txt') P.set_scores_stochastic_(model.P_scores) print_transition_probabilities(P, phone_symbol_table, phone_ids, filename='P_scores.txt') if not os.path.exists(lang_dir / 'LG.pt'): logging.debug("Loading L_disambig.fst.txt") with open(lang_dir / 'L_disambig.fst.txt') as f: L = k2.Fsa.from_openfst(f.read(), acceptor=False) logging.debug("Loading G.fst.txt") with open(lang_dir / 'G.fst.txt') as f: G = k2.Fsa.from_openfst(f.read(), acceptor=False) first_phone_disambig_id = find_first_disambig_symbol(phone_symbol_table) first_word_disambig_id = find_first_disambig_symbol(symbol_table) LG = compile_LG(L=L, G=G, ctc_topo=ctc_topo, labels_disambig_id_start=first_phone_disambig_id, aux_labels_disambig_id_start=first_word_disambig_id) torch.save(LG.as_dict(), lang_dir / 'LG.pt') else: logging.debug("Loading pre-compiled LG") d = torch.load(lang_dir / 'LG.pt') LG = k2.Fsa.from_dict(d) # load dataset aishell = AishellAsrDataModule(args) test_dl = aishell.test_dataloaders() # if not torch.cuda.is_available(): # logging.error('No GPU detected!') # sys.exit(-1) logging.debug("convert LG to device") LG = LG.to(device) LG.aux_labels = k2.ragged.remove_values_eq(LG.aux_labels, 0) LG.requires_grad_(False) logging.debug("About to decode") results = decode(dataloader=test_dl, model=model, device=device, LG=LG, symbols=symbol_table) s = '' results2 = [] for ref, hyp in results: s += f'ref={ref}\n' s += f'hyp={hyp}\n' results2.append((list(''.join(ref)), list(''.join(hyp)))) logging.info(s) # compute WER dists = [edit_distance(r, h) for r, h in results] dists2 = [edit_distance(r, h) for r, h in results2] errors = { key: sum(dist[key] for dist in dists) for key in ['sub', 'ins', 'del', 'total'] } errors2 = { key: sum(dist[key] for dist in dists2) for key in ['sub', 'ins', 'del', 'total'] } total_words = sum(len(ref) for ref, _ in results) total_chars = sum(len(ref) for ref, _ in results2) # Print Kaldi-like message: # %WER 8.20 [ 4459 / 54402, 695 ins, 427 del, 3337 sub ] logging.info( f'%WER {errors["total"] / total_words:.2%} ' f'[{errors["total"]} / {total_words}, {errors["ins"]} ins, {errors["del"]} del, {errors["sub"]} sub ]' ) logging.info( f'%WER {errors2["total"] / total_chars:.2%} ' f'[{errors2["total"]} / {total_chars}, {errors2["ins"]} ins, {errors2["del"]} del, {errors2["sub"]} sub ]' )
def main(): args = get_parser().parse_args() print('World size:', args.world_size, 'Rank:', args.local_rank) setup_dist(rank=args.local_rank, world_size=args.world_size) fix_random_seed(42) start_epoch = 0 num_epochs = 10 use_adam = True exp_dir = f'exp-lstm-adam-mmi-bigram-musan-dist' setup_logger('{}/log/log-train'.format(exp_dir), use_console=args.local_rank == 0) tb_writer = SummaryWriter( log_dir=f'{exp_dir}/tensorboard') if args.local_rank == 0 else None # load L, G, symbol_table lang_dir = Path('data/lang_nosp') phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt') word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt') logging.info("Loading L.fst") if (lang_dir / 'Linv.pt').exists(): L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt')) else: with open(lang_dir / 'L.fst.txt') as f: L = k2.Fsa.from_openfst(f.read(), acceptor=False) L_inv = k2.arc_sort(L.invert_()) torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt') graph_compiler = MmiTrainingGraphCompiler(L_inv=L_inv, phones=phone_symbol_table, words=word_symbol_table) phone_ids = get_phone_symbols(phone_symbol_table) P = create_bigram_phone_lm(phone_ids) P.scores = torch.zeros_like(P.scores) # load dataset feature_dir = Path('exp/data') logging.info("About to get train cuts") cuts_train = CutSet.from_json(feature_dir / 'cuts_train-clean-100.json.gz') logging.info("About to get dev cuts") cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev-clean.json.gz') logging.info("About to get Musan cuts") cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz') logging.info("About to create train dataset") transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))] if not args.bucketing_sampler: # We don't mix concatenating the cuts and bucketing # Here we insert concatenation before mixing so that the # noises from Musan are mixed onto almost-zero-energy # padding frames. transforms = [CutConcatenate()] + transforms train = K2SpeechRecognitionDataset(cuts_train, cut_transforms=transforms) if args.bucketing_sampler: logging.info('Using BucketingSampler.') train_sampler = BucketingSampler(cuts_train, max_frames=40000, shuffle=True, num_buckets=30) else: logging.info('Using regular sampler with cut concatenation.') train_sampler = SingleCutSampler( cuts_train, max_frames=30000, shuffle=True, ) logging.info("About to create train dataloader") train_dl = torch.utils.data.DataLoader(train, sampler=train_sampler, batch_size=None, num_workers=4) logging.info("About to create dev dataset") validate = K2SpeechRecognitionDataset(cuts_dev) # Note: we explicitly set world_size to 1 to disable the auto-detection of # distributed training inside the sampler. This way, every GPU will # perform the computation on the full dev set. It is a bit wasteful, # but unfortunately loss aggregation between multiple processes with # torch.distributed.all_reduce() tends to hang indefinitely inside # NCCL after ~3000 steps. With the current approach, we can still report # the loss on the full validation set. valid_sampler = SingleCutSampler(cuts_dev, max_frames=90000, world_size=1, rank=0) logging.info("About to create dev dataloader") valid_dl = torch.utils.data.DataLoader(validate, sampler=valid_sampler, batch_size=None, num_workers=1) if not torch.cuda.is_available(): logging.error('No GPU detected!') sys.exit(-1) logging.info("About to create model") device_id = args.local_rank device = torch.device('cuda', device_id) model = TdnnLstm1b( num_features=40, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=3) model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True) model.to(device) describe(model) if use_adam: learning_rate = 1e-3 weight_decay = 5e-4 optimizer = optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay) # Equivalent to the following in the epoch loop: # if epoch > 6: # curr_learning_rate *= 0.8 lr_scheduler = optim.lr_scheduler.LambdaLR( optimizer, lambda ep: 1.0 if ep < 7 else 0.8**(ep - 6)) else: learning_rate = 5e-5 weight_decay = 1e-5 momentum = 0.9 lr_schedule_gamma = 0.7 optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=momentum, weight_decay=weight_decay) lr_scheduler = optim.lr_scheduler.ExponentialLR( optimizer=optimizer, gamma=lr_schedule_gamma) best_objf = np.inf best_valid_objf = np.inf best_epoch = start_epoch best_model_path = os.path.join(exp_dir, 'best_model.pt') best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info') global_batch_idx_train = 0 # for logging only if start_epoch > 0: model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(start_epoch - 1)) ckpt = load_checkpoint(filename=model_path, model=model, optimizer=optimizer, scheduler=lr_scheduler) best_objf = ckpt['objf'] best_valid_objf = ckpt['valid_objf'] global_batch_idx_train = ckpt['global_batch_idx_train'] logging.info( f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}" ) if args.world_size > 1: logging.info( 'Using DistributedDataParallel in training. ' 'The reported loss, num_frames, etc. for training steps include ' 'only the batches seen in the master process (the actual loss ' 'includes batches from all GPUs, and the actual num_frames is ' f'approx. {args.world_size}x larger.') # For now do not sync BatchNorm across GPUs due to NCCL hanging in all_gather... # model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model) model = DDP(model, device_ids=[args.local_rank], output_device=args.local_rank) for epoch in range(start_epoch, num_epochs): train_sampler.set_epoch(epoch) # LR scheduler can hold multiple learning rates for multiple parameter groups; # For now we report just the first LR which we assume concerns most of the parameters. curr_learning_rate = lr_scheduler.get_last_lr()[0] if tb_writer is not None: tb_writer.add_scalar('train/learning_rate', curr_learning_rate, global_batch_idx_train) tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train) logging.info('epoch {}, learning rate {}'.format( epoch, curr_learning_rate)) objf, valid_objf, global_batch_idx_train = train_one_epoch( dataloader=train_dl, valid_dataloader=valid_dl, model=model, P=P, device=device, graph_compiler=graph_compiler, optimizer=optimizer, current_epoch=epoch, tb_writer=tb_writer, num_epochs=num_epochs, global_batch_idx_train=global_batch_idx_train, ) lr_scheduler.step() # the lower, the better if valid_objf < best_valid_objf: best_valid_objf = valid_objf best_objf = objf best_epoch = epoch save_checkpoint(filename=best_model_path, model=model, optimizer=None, scheduler=None, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, local_rank=args.local_rank, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train) save_training_info(filename=best_epoch_info_filename, model_path=best_model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch) # we always save the model for every epoch model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch)) save_checkpoint(filename=model_path, model=model, optimizer=optimizer, scheduler=lr_scheduler, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, local_rank=args.local_rank, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train) epoch_info_filename = os.path.join(exp_dir, 'epoch-{}-info'.format(epoch)) save_training_info(filename=epoch_info_filename, model_path=model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch) logging.warning('Done') cleanup_dist()
def main(): fix_random_seed(42) exp_dir = f'exp-lstm-adam-mmi-bigram-musan' setup_logger('{}/log/log-train'.format(exp_dir)) tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') # load L, G, symbol_table lang_dir = Path('data/lang_nosp') phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt') word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt') logging.info("Loading L.fst") if (lang_dir / 'Linv.pt').exists(): L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt')) else: with open(lang_dir / 'L.fst.txt') as f: L = k2.Fsa.from_openfst(f.read(), acceptor=False) L_inv = k2.arc_sort(L.invert_()) torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt') graph_compiler = MmiTrainingGraphCompiler(L_inv=L_inv, phones=phone_symbol_table, words=word_symbol_table) phone_ids = get_phone_symbols(phone_symbol_table) P = create_bigram_phone_lm(phone_ids) P.scores = torch.zeros_like(P.scores) # load dataset feature_dir = Path('exp/data') logging.info("About to get train cuts") cuts_train = CutSet.from_json(feature_dir / 'cuts_train-clean-100.json.gz') logging.info("About to get dev cuts") cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev-clean.json.gz') logging.info("About to get Musan cuts") cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz') logging.info("About to create train dataset") train = K2SpeechRecognitionIterableDataset(cuts_train, max_frames=30000, shuffle=True, aug_cuts=cuts_musan, aug_prob=0.5, aug_snr=(10, 20)) logging.info("About to create dev dataset") validate = K2SpeechRecognitionIterableDataset(cuts_dev, max_frames=30000, shuffle=False, concat_cuts=False) logging.info("About to create train dataloader") train_dl = torch.utils.data.DataLoader(train, batch_size=None, num_workers=2) logging.info("About to create dev dataloader") valid_dl = torch.utils.data.DataLoader(validate, batch_size=None, num_workers=1) if not torch.cuda.is_available(): logging.error('No GPU detected!') sys.exit(-1) logging.info("About to create model") device_id = 0 device = torch.device('cuda', device_id) model = TdnnLstm1b( num_features=40, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=3) model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True) learning_rate = 1e-3 start_epoch = 0 num_epochs = 10 best_objf = np.inf best_epoch = start_epoch best_model_path = os.path.join(exp_dir, 'best_model.pt') best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info') global_batch_idx_train = 0 # for logging only global_batch_idx_valid = 0 # for logging only if start_epoch > 0: model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(start_epoch - 1)) (epoch, learning_rate, objf) = load_checkpoint(filename=model_path, model=model) best_objf = objf logging.info("epoch = {}, objf = {}".format(epoch, objf)) model.to(device) describe(model) # optimizer = optim.SGD(model.parameters(), # lr=learning_rate, # momentum=0.9, # weight_decay=5e-4) optimizer = optim.AdamW( model.parameters(), # lr=learning_rate, weight_decay=5e-4) curr_learning_rate = learning_rate for epoch in range(start_epoch, num_epochs): # curr_learning_rate = learning_rate * pow(0.4, epoch) if epoch > 6: curr_learning_rate *= 0.8 for param_group in optimizer.param_groups: param_group['lr'] = curr_learning_rate tb_writer.add_scalar('learning_rate', curr_learning_rate, epoch) logging.info('epoch {}, learning rate {}'.format( epoch, curr_learning_rate)) objf = train_one_epoch(dataloader=train_dl, valid_dataloader=valid_dl, model=model, P=P, device=device, graph_compiler=graph_compiler, optimizer=optimizer, current_epoch=epoch, tb_writer=tb_writer, num_epochs=num_epochs, global_batch_idx_train=global_batch_idx_train, global_batch_idx_valid=global_batch_idx_valid) # the lower, the better if objf < best_objf: best_objf = objf best_epoch = epoch save_checkpoint(filename=best_model_path, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf) save_training_info(filename=best_epoch_info_filename, model_path=best_model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=best_objf, best_objf=best_objf, best_epoch=best_epoch) # we always save the model for every epoch model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch)) save_checkpoint(filename=model_path, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf) epoch_info_filename = os.path.join(exp_dir, 'epoch-{}-info'.format(epoch)) save_training_info(filename=epoch_info_filename, model_path=model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, best_epoch=best_epoch) logging.warning('Done')
def main(): args = get_parser().parse_args() start_epoch = args.start_epoch num_epochs = args.num_epochs max_frames = args.max_frames accum_grad = args.accum_grad den_scale = args.den_scale att_rate = args.att_rate fix_random_seed(42) exp_dir = Path('exp-transformer-noam-mmi-att-musan') setup_logger('{}/log/log-train'.format(exp_dir)) tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') # load L, G, symbol_table lang_dir = Path('data/lang_nosp') phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt') word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt') logging.info("Loading L.fst") if (lang_dir / 'Linv.pt').exists(): L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt')) else: with open(lang_dir / 'L.fst.txt') as f: L = k2.Fsa.from_openfst(f.read(), acceptor=False) L_inv = k2.arc_sort(L.invert_()) torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt') graph_compiler = MmiTrainingGraphCompiler(L_inv=L_inv, phones=phone_symbol_table, words=word_symbol_table) phone_ids = get_phone_symbols(phone_symbol_table) P = create_bigram_phone_lm(phone_ids) P.scores = torch.zeros_like(P.scores) # load dataset feature_dir = Path('exp/data') logging.info("About to get train cuts") cuts_train = CutSet.from_json(feature_dir / 'cuts_train-clean-100.json.gz') logging.info("About to get dev cuts") cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev-clean.json.gz') logging.info("About to get Musan cuts") cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz') logging.info("About to create train dataset") transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))] if not args.bucketing_sampler: # We don't mix concatenating the cuts and bucketing # Here we insert concatenation before mixing so that the # noises from Musan are mixed onto almost-zero-energy # padding frames. transforms = [CutConcatenate()] + transforms train = K2SpeechRecognitionDataset(cuts_train, cut_transforms=transforms) if args.bucketing_sampler: logging.info('Using BucketingSampler.') train_sampler = BucketingSampler(cuts_train, max_frames=max_frames, shuffle=True, num_buckets=args.num_buckets) else: logging.info('Using regular sampler with cut concatenation.') train_sampler = SingleCutSampler( cuts_train, max_frames=max_frames, shuffle=True, ) logging.info("About to create train dataloader") train_dl = torch.utils.data.DataLoader(train, sampler=train_sampler, batch_size=None, num_workers=4) logging.info("About to create dev dataset") validate = K2SpeechRecognitionDataset(cuts_dev) valid_sampler = SingleCutSampler(cuts_dev, max_frames=max_frames) logging.info("About to create dev dataloader") valid_dl = torch.utils.data.DataLoader(validate, sampler=valid_sampler, batch_size=None, num_workers=1) if not torch.cuda.is_available(): logging.error('No GPU detected!') sys.exit(-1) logging.info("About to create model") device_id = 0 device = torch.device('cuda', device_id) if att_rate != 0.0: num_decoder_layers = 6 else: num_decoder_layers = 0 model = Transformer( num_features=40, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers) model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True) model.to(device) describe(model) optimizer = Noam(model.parameters(), model_size=256, factor=1.0, warm_step=args.warm_step) best_objf = np.inf best_valid_objf = np.inf best_epoch = start_epoch best_model_path = os.path.join(exp_dir, 'best_model.pt') best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info') global_batch_idx_train = 0 # for logging only if start_epoch > 0: model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(start_epoch - 1)) ckpt = load_checkpoint(filename=model_path, model=model, optimizer=optimizer) best_objf = ckpt['objf'] best_valid_objf = ckpt['valid_objf'] global_batch_idx_train = ckpt['global_batch_idx_train'] logging.info( f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}" ) for epoch in range(start_epoch, num_epochs): train_sampler.set_epoch(epoch) curr_learning_rate = optimizer._rate tb_writer.add_scalar('train/learning_rate', curr_learning_rate, global_batch_idx_train) tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train) logging.info('epoch {}, learning rate {}'.format( epoch, curr_learning_rate)) objf, valid_objf, global_batch_idx_train = train_one_epoch( dataloader=train_dl, valid_dataloader=valid_dl, model=model, P=P, device=device, graph_compiler=graph_compiler, optimizer=optimizer, accum_grad=accum_grad, den_scale=den_scale, att_rate=att_rate, current_epoch=epoch, tb_writer=tb_writer, num_epochs=num_epochs, global_batch_idx_train=global_batch_idx_train, ) # the lower, the better if valid_objf < best_valid_objf: best_valid_objf = valid_objf best_objf = objf best_epoch = epoch save_checkpoint(filename=best_model_path, optimizer=None, scheduler=None, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train) save_training_info(filename=best_epoch_info_filename, model_path=best_model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch) # we always save the model for every epoch model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch)) save_checkpoint(filename=model_path, optimizer=optimizer, scheduler=None, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train) epoch_info_filename = os.path.join(exp_dir, 'epoch-{}-info'.format(epoch)) save_training_info(filename=epoch_info_filename, model_path=model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch) logging.warning('Done')
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() ''' model_type = args.model_type start_epoch = args.start_epoch num_epochs = args.num_epochs accum_grad = args.accum_grad den_scale = args.den_scale att_rate = args.att_rate fix_random_seed(42) setup_dist(rank, world_size, args.master_port) exp_dir = Path('exp-' + model_type + '-noam-mmi-att-musan-sa') setup_logger(f'{exp_dir}/log/log-train-{rank}') if args.tensorboard and rank == 0: tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') else: tb_writer = None # tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') if args.tensorboard and rank == 0 else None logging.info("Loading lexicon and symbol tables") lang_dir = Path('data/lang_nosp') lexicon = Lexicon(lang_dir) device_id = rank device = torch.device('cuda', device_id) graph_compiler = MmiTrainingGraphCompiler( lexicon=lexicon, device=device, ) phone_ids = lexicon.phone_symbols() P = create_bigram_phone_lm(phone_ids) P.scores = torch.zeros_like(P.scores) P = P.to(device) librispeech = LibriSpeechAsrDataModule(args) train_dl = librispeech.train_dataloaders() valid_dl = librispeech.valid_dataloaders() if not torch.cuda.is_available(): logging.error('No GPU detected!') sys.exit(-1) logging.info("About to create model") if att_rate != 0.0: num_decoder_layers = 6 else: num_decoder_layers = 0 if model_type == "transformer": model = Transformer( num_features=80, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers) else: model = Conformer( num_features=80, nhead=args.nhead, d_model=args.attention_dim, num_classes=len(phone_ids) + 1, # +1 for the blank symbol subsampling_factor=4, num_decoder_layers=num_decoder_layers) model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True) model.to(device) describe(model) model = DDP(model, device_ids=[rank]) optimizer = Noam(model.parameters(), model_size=args.attention_dim, factor=1.0, warm_step=args.warm_step) best_objf = np.inf best_valid_objf = np.inf best_epoch = start_epoch best_model_path = os.path.join(exp_dir, 'best_model.pt') best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info') global_batch_idx_train = 0 # for logging only if start_epoch > 0: model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(start_epoch - 1)) ckpt = load_checkpoint(filename=model_path, model=model, optimizer=optimizer) best_objf = ckpt['objf'] best_valid_objf = ckpt['valid_objf'] global_batch_idx_train = ckpt['global_batch_idx_train'] logging.info( f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}" ) for epoch in range(start_epoch, num_epochs): train_dl.sampler.set_epoch(epoch) curr_learning_rate = optimizer._rate if tb_writer is not None: tb_writer.add_scalar('train/learning_rate', curr_learning_rate, global_batch_idx_train) tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train) logging.info('epoch {}, learning rate {}'.format( epoch, curr_learning_rate)) objf, valid_objf, global_batch_idx_train = train_one_epoch( dataloader=train_dl, valid_dataloader=valid_dl, model=model, P=P, device=device, graph_compiler=graph_compiler, optimizer=optimizer, accum_grad=accum_grad, den_scale=den_scale, att_rate=att_rate, current_epoch=epoch, tb_writer=tb_writer, num_epochs=num_epochs, global_batch_idx_train=global_batch_idx_train, world_size=world_size, ) # the lower, the better if valid_objf < best_valid_objf: best_valid_objf = valid_objf best_objf = objf best_epoch = epoch save_checkpoint(filename=best_model_path, optimizer=None, scheduler=None, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train, local_rank=rank) save_training_info(filename=best_epoch_info_filename, model_path=best_model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch, local_rank=rank) # we always save the model for every epoch model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch)) save_checkpoint(filename=model_path, optimizer=optimizer, scheduler=None, model=model, epoch=epoch, learning_rate=curr_learning_rate, objf=objf, valid_objf=valid_objf, global_batch_idx_train=global_batch_idx_train, local_rank=rank) epoch_info_filename = os.path.join(exp_dir, 'epoch-{}-info'.format(epoch)) save_training_info(filename=epoch_info_filename, model_path=model_path, current_epoch=epoch, learning_rate=curr_learning_rate, objf=objf, best_objf=best_objf, valid_objf=valid_objf, best_valid_objf=best_valid_objf, best_epoch=best_epoch, local_rank=rank) logging.warning('Done') torch.distributed.barrier() # NOTE: The training process is very likely to hang at this point. # If you press ctrl + c, your GPU memory will not be freed. # To free you GPU memory, you can run: # # $ ps aux | grep multi # # And it will print something like below: # # kuangfa+ 430518 98.9 0.6 57074236 3425732 pts/21 Rl Apr02 639:01 /root/fangjun/py38/bin/python3 -c from multiprocessing.spawn # # You can kill the process manually by: # # $ kill -9 430518 # # And you will see that your GPU is now not occupied anymore. cleanup_dist()