def test_train_step(self): input = torch.randint(0, 24, (8, 128)).long().to(device) input_lengths = torch.randint(100, 129, (8, )).long().to(device) input_lengths[-1] = 128 mel_spec = torch.rand(8, 30, c.audio['num_mels']).to(device) linear_spec = torch.rand(8, 30, c.audio['num_freq']).to(device) mel_lengths = torch.randint(20, 30, (8, )).long().to(device) stop_targets = torch.zeros(8, 30, 1).float().to(device) speaker_ids = torch.randint(0, 5, (8, )).long().to(device) for idx in mel_lengths: stop_targets[:, int(idx.item()):, 0] = 1.0 stop_targets = stop_targets.view(input.shape[0], stop_targets.size(1) // c.r, -1) stop_targets = (stop_targets.sum(2) > 0.0).unsqueeze(2).float().squeeze() criterion = L1LossMasked().to(device) criterion_st = nn.BCEWithLogitsLoss().to(device) model = Tacotron( num_chars=32, num_speakers=5, linear_dim=c.audio['num_freq'], mel_dim=c.audio['num_mels'], r=c.r, memory_size=c.memory_size ).to(device) #FIXME: missing num_speakers parameter to Tacotron ctor model.train() print(" > Num parameters for Tacotron model:%s" % (count_parameters(model))) model_ref = copy.deepcopy(model) count = 0 for param, param_ref in zip(model.parameters(), model_ref.parameters()): assert (param - param_ref).sum() == 0, param count += 1 optimizer = optim.Adam(model.parameters(), lr=c.lr) for _ in range(5): mel_out, linear_out, align, stop_tokens = model.forward( input, input_lengths, mel_spec, speaker_ids) optimizer.zero_grad() loss = criterion(mel_out, mel_spec, mel_lengths) stop_loss = criterion_st(stop_tokens, stop_targets) loss = loss + criterion(linear_out, linear_spec, mel_lengths) + stop_loss loss.backward() optimizer.step() # check parameter changes count = 0 for param, param_ref in zip(model.parameters(), model_ref.parameters()): # ignore pre-higway layer since it works conditional # if count not in [145, 59]: assert (param != param_ref).any( ), "param {} with shape {} not updated!! \n{}\n{}".format( count, param.shape, param, param_ref) count += 1
def test_train_step(self): input = torch.randint(0, 24, (8, 128)).long().to(device) mel_spec = torch.rand(8, 30, c.audio['num_mels']).to(device) linear_spec = torch.rand(8, 30, c.audio['num_freq']).to(device) mel_lengths = torch.randint(20, 30, (8, )).long().to(device) stop_targets = torch.zeros(8, 30, 1).float().to(device) for idx in mel_lengths: stop_targets[:, int(idx.item()):, 0] = 1.0 stop_targets = stop_targets.view(input.shape[0], stop_targets.size(1) // c.r, -1) stop_targets = (stop_targets.sum(2) > 0.0).unsqueeze(2).float() criterion = L1LossMasked().to(device) criterion_st = nn.BCELoss().to(device) model = Tacotron(c.embedding_size, c.audio['num_freq'], c.audio['num_mels'], c.r).to(device) model.train() model_ref = copy.deepcopy(model) count = 0 for param, param_ref in zip(model.parameters(), model_ref.parameters()): assert (param - param_ref).sum() == 0, param count += 1 optimizer = optim.Adam(model.parameters(), lr=c.lr) for i in range(5): mel_out, linear_out, align, stop_tokens = model.forward( input, mel_spec) assert stop_tokens.data.max() <= 1.0 assert stop_tokens.data.min() >= 0.0 optimizer.zero_grad() loss = criterion(mel_out, mel_spec, mel_lengths) stop_loss = criterion_st(stop_tokens, stop_targets) loss = loss + criterion(linear_out, linear_spec, mel_lengths) + stop_loss loss.backward() optimizer.step() # check parameter changes count = 0 for param, param_ref in zip(model.parameters(), model_ref.parameters()): # ignore pre-higway layer since it works conditional # if count not in [145, 59]: assert (param != param_ref).any( ), "param {} with shape {} not updated!! \n{}\n{}".format( count, param.shape, param, param_ref) count += 1
def test_in_out(self): layer = L1LossMasked() dummy_input = T.ones(4, 8, 128).float() dummy_target = T.ones(4, 8, 128).float() dummy_length = (T.ones(4) * 8).long() output = layer(dummy_input, dummy_target, dummy_length) assert output.item() == 0.0 dummy_input = T.ones(4, 8, 128).float() dummy_target = T.zeros(4, 8, 128).float() dummy_length = (T.ones(4) * 8).long() output = layer(dummy_input, dummy_target, dummy_length) assert output.item() == 1.0, "1.0 vs {}".format(output.data[0]) dummy_input = T.ones(4, 8, 128).float() dummy_target = T.zeros(4, 8, 128).float() dummy_length = (T.arange(5, 9)).long() mask = ( (sequence_mask(dummy_length).float() - 1.0) * 100.0).unsqueeze(2) output = layer(dummy_input + mask, dummy_target, dummy_length) assert output.item() == 1.0, "1.0 vs {}".format(output.data[0])
def main(args): # DISTRUBUTED if num_gpus > 1: init_distributed(args.rank, num_gpus, args.group_id, c.distributed["backend"], c.distributed["url"]) num_chars = len(phonemes) if c.use_phonemes else len(symbols) model = setup_model(num_chars, c, args.use_half) print(" | > Num output units : {}".format(ap.num_freq), flush=True) if args.use_half: print(' | > Use half mode') optimizer_eps = 1e-08 if not args.use_half else 1e-04 optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0, eps=optimizer_eps) # optimizer = optim.SGD(model.parameters(), lr=c.lr, weight_decay=0) if c.stopnet and c.separate_stopnet: optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr, weight_decay=0, eps=optimizer_eps) # optimizer_st = optim.SGD(model.decoder.stopnet.parameters(), lr=c.lr, weight_decay=0) else: optimizer_st = None if c.loss_masking: criterion = L1LossMasked() if c.model == "Tacotron" else MSELossMasked( ) else: criterion = nn.L1Loss() if c.model == "Tacotron" else nn.MSELoss() criterion_st = nn.BCEWithLogitsLoss() if c.stopnet else None if args.restore_path: checkpoint = torch.load(args.restore_path) try: # TODO: fix optimizer init, model.cuda() needs to be called before # optimizer restore # optimizer.load_state_dict(checkpoint['optimizer']) if len(c.reinit_layers) > 0: raise RuntimeError model.load_state_dict(checkpoint['model']) except: print(" > Partial model initialization.") partial_init_flag = True model_dict = model.state_dict() model_dict = set_init_dict(model_dict, checkpoint, c) model.load_state_dict(model_dict) del model_dict for group in optimizer.param_groups: group['lr'] = c.lr print(" > Model restored from step %d" % checkpoint['step'], flush=True) start_epoch = checkpoint['epoch'] args.restore_step = checkpoint['step'] else: args.restore_step = 0 # use half mode if args.use_half: model.half() for layer in model.modules(): if isinstance(layer, torch.nn.BatchNorm1d): layer.float() if use_cuda: model = model.cuda() criterion.cuda() if criterion_st: criterion_st.cuda() if args.restore_path: # print(checkpoint['optimizer']) # print('---opt', optimizer) optimizer.load_state_dict(checkpoint['optimizer']) # DISTRUBUTED if num_gpus > 1: model = apply_gradient_allreduce(model) # reset lr if args.reset_lr: for group in optimizer.param_groups: group['initial_lr'] = c.lr if c.lr_decay: scheduler = NoamLR( optimizer, warmup_steps=c.warmup_steps, last_epoch=args.restore_step - 1, use_half=args.use_half, ) else: scheduler = None num_params = count_parameters(model) print("\n > Model has {} parameters".format(num_params), flush=True) if 'best_loss' not in locals(): best_loss = float('inf') for epoch in range(0, c.epochs): train_loss, current_step = train(model, criterion, criterion_st, optimizer, optimizer_st, scheduler, ap, epoch, args.use_half) if c.run_eval: val_loss = evaluate(model, criterion, criterion_st, ap, current_step, epoch, args.use_half) print( " | > Training Loss: {:.5f} Validation Loss: {:.5f}".format( train_loss, val_loss), flush=True) target_loss = val_loss else: print(" | > Training Loss: {:.5f}".format(train_loss), flush=True) target_loss = train_loss best_loss = save_best_model(model, optimizer, target_loss, best_loss, OUT_PATH, current_step, epoch)
def main(args): # Setup the dataset train_dataset = LJSpeechDataset(os.path.join(c.data_path, 'train_metadata.csv'), os.path.join(c.data_path, 'wavs'), c.r, c.sample_rate, c.text_cleaner, c.num_mels, c.min_level_db, c.frame_shift_ms, c.frame_length_ms, c.preemphasis, c.ref_level_db, c.num_freq, c.power, min_seq_len=c.min_seq_len ) train_loader = DataLoader(train_dataset, batch_size=c.batch_size, shuffle=False, collate_fn=train_dataset.collate_fn, drop_last=False, num_workers=c.num_loader_workers, pin_memory=True) ''' val_dataset = LJSpeechDataset(os.path.join(c.data_path, 'valid_metadata.csv'), os.path.join(c.data_path, 'wavs'), c.r, c.sample_rate, c.text_cleaner, c.num_mels, c.min_level_db, c.frame_shift_ms, c.frame_length_ms, c.preemphasis, c.ref_level_db, c.num_freq, c.power ) val_loader = DataLoader(val_dataset, batch_size=c.eval_batch_size, shuffle=False, collate_fn=val_dataset.collate_fn, drop_last=False, num_workers=4, pin_memory=True) ''' model = Tacotron(c.embedding_size, c.num_freq, c.num_mels, c.r) optimizer = optim.Adam(model.parameters(), lr=c.lr) optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr) criterion = L1LossMasked() criterion_st = nn.BCELoss() if args.restore_path: checkpoint = torch.load(args.restore_path) model.load_state_dict(checkpoint['model']) optimizer = optim.Adam(model.parameters(), lr=c.lr) optimizer.load_state_dict(checkpoint['optimizer']) for state in optimizer.state.values(): for k, v in state.items(): if torch.is_tensor(v): state[k] = v.cuda() print(" > Model restored from step %d" % checkpoint['step']) start_epoch = checkpoint['step'] // len(train_loader) best_loss = checkpoint['linear_loss'] start_epoch = 0 args.restore_step = checkpoint['step'] optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr) else: args.restore_step = 0 print("\n > Starting a new training") if use_cuda: model = nn.DataParallel(model.cuda()) criterion.cuda() criterion_st.cuda() num_params = count_parameters(model) print(" | > Model has {} parameters".format(num_params)) if not os.path.exists(CHECKPOINT_PATH): os.mkdir(CHECKPOINT_PATH) if 'best_loss' not in locals(): best_loss = float('inf') for epoch in range(0, c.epochs): train_loss, current_step = train( model, criterion, criterion_st, train_loader, optimizer, optimizer_st, epoch) #val_loss = evaluate(model, criterion, criterion_st, val_loader, current_step) best_loss = save_best_model(model, optimizer, train_loss, best_loss, OUT_PATH, current_step, epoch)
def main(args): #pylint: disable=redefined-outer-name # Audio processor ap = AudioProcessor(**c.audio) # DISTRUBUTED if num_gpus > 1: init_distributed(args.rank, num_gpus, args.group_id, c.distributed["backend"], c.distributed["url"]) num_chars = len(phonemes) if c.use_phonemes else len(symbols) if c.use_speaker_embedding: speakers = get_speakers(c.data_path, c.meta_file_train, c.dataset) if args.restore_path: prev_out_path = os.path.dirname(args.restore_path) speaker_mapping = load_speaker_mapping(prev_out_path) assert all([speaker in speaker_mapping for speaker in speakers]), "As of now you, you cannot " \ "introduce new speakers to " \ "a previously trained model." else: speaker_mapping = {name: i for i, name in enumerate(speakers)} save_speaker_mapping(OUT_PATH, speaker_mapping) num_speakers = len(speaker_mapping) print("Training with {} speakers: {}".format(num_speakers, ", ".join(speakers))) else: num_speakers = 0 model = setup_model(num_chars, num_speakers, c) print(" | > Num output units : {}".format(ap.num_freq), flush=True) #optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0) optimizer = Ranger(model.parameters(), lr=c.lr, weight_decay=c.wd) optimizer_gst = Ranger(model.textgst.parameters(), lr=c.lr, weight_decay=c.wd) if c.text_gst else None if c.stopnet and c.separate_stopnet: optimizer_st = Ranger(model.decoder.stopnet.parameters(), lr=c.lr) else: optimizer_st = None if c.loss_masking: criterion = L1LossMasked() if c.model in ["Tacotron", "TacotronGST"] else MSELossMasked() else: criterion = nn.L1Loss() if c.model in ["Tacotron", "TacotronGST"] else nn.MSELoss() criterion_st = nn.BCEWithLogitsLoss() if c.stopnet else None criterion_gst = nn.L1Loss() if c.text_gst else None if args.restore_path: checkpoint = torch.load(args.restore_path) try: # TODO: fix optimizer init, model.cuda() needs to be called before # optimizer restore # optimizer.load_state_dict(checkpoint['optimizer']) if c.reinit_layers: raise RuntimeError model.load_state_dict(checkpoint['model']) except: print(" > Partial model initialization.") model_dict = model.state_dict() model_dict = set_init_dict(model_dict, checkpoint, c) model.load_state_dict(model_dict) del model_dict for group in optimizer.param_groups: group['lr'] = c.lr print( " > Model restored from step %d" % checkpoint['step'], flush=True) args.restore_step = checkpoint['step'] else: args.restore_step = 0 if use_cuda: model = model.cuda() criterion.cuda() if criterion_st: criterion_st.cuda() # DISTRUBUTED if num_gpus > 1: model = apply_gradient_allreduce(model) if c.lr_decay: scheduler = NoamLR( optimizer, warmup_steps=c.warmup_steps, last_epoch=args.restore_step - 1) else: scheduler = None num_params = count_parameters(model) print("\n > Model has {} parameters".format(num_params), flush=True) if 'best_loss' not in locals(): best_loss = float('inf') global_step = args.restore_step for epoch in range(0, c.epochs): # set gradual training if c.gradual_training is not None: r, c.batch_size = gradual_training_scheduler(global_step, c) c.r = r model.decoder.set_r(r) print(" > Number of outputs per iteration:", model.decoder.r) train_loss, global_step = train(model, criterion, criterion_st, optimizer, optimizer_st, scheduler, ap, global_step, epoch, criterion_gst=criterion_gst, optimizer_gst=optimizer_gst) if epoch % 5 == 0: val_loss = evaluate(model, criterion, criterion_st, criterion_gst, ap, global_step, epoch) print( " | > Training Loss: {:.5f} Validation Loss: {:.5f}".format( train_loss, val_loss), flush=True) target_loss = train_loss if c.run_eval: target_loss = val_loss best_loss = save_best_model(model, optimizer, optimizer_st, optimizer_gst, target_loss, best_loss, OUT_PATH, global_step, epoch)
def main(args): # DISTRUBUTED if num_gpus > 1: init_distributed(args.rank, num_gpus, args.group_id, c.distributed["backend"], c.distributed["url"]) num_chars = len(phonemes) if c.use_phonemes else len(symbols) model = MyModel(num_chars=num_chars, r=c.r, attn_norm=c.attention_norm) print(" | > Num output units : {}".format(ap.num_freq), flush=True) optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0) optimizer_st = optim.Adam( model.decoder.stopnet.parameters(), lr=c.lr, weight_decay=0) criterion = L1LossMasked() if c.model == "Tacotron" else MSELossMasked() criterion_st = nn.BCEWithLogitsLoss() if args.restore_path: checkpoint = torch.load(args.restore_path) try: # TODO: fix optimizer init, model.cuda() needs to be called before # optimizer restore # optimizer.load_state_dict(checkpoint['optimizer']) if len(c.reinit_layers) > 0: raise RuntimeError model.load_state_dict(checkpoint['model']) except: print(" > Partial model initialization.") partial_init_flag = True model_dict = model.state_dict() model_dict = set_init_dict(model_dict, checkpoint, c) model.load_state_dict(model_dict) del model_dict if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() for group in optimizer.param_groups: group['lr'] = c.lr print( " > Model restored from step %d" % checkpoint['step'], flush=True) start_epoch = checkpoint['epoch'] # best_loss = checkpoint['postnet_loss'] args.restore_step = checkpoint['step'] else: args.restore_step = 0 if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() # DISTRUBUTED if num_gpus > 1: model = apply_gradient_allreduce(model) if c.lr_decay: scheduler = NoamLR( optimizer, warmup_steps=c.warmup_steps, last_epoch=args.restore_step - 1) else: scheduler = None num_params = count_parameters(model) print("\n > Model has {} parameters".format(num_params), flush=True) if 'best_loss' not in locals(): best_loss = float('inf') for epoch in range(0, c.epochs): train_loss, current_step = train(model, criterion, criterion_st, optimizer, optimizer_st, scheduler, ap, epoch) val_loss = evaluate(model, criterion, criterion_st, ap, current_step, epoch) print( " | > Training Loss: {:.5f} Validation Loss: {:.5f}".format( train_loss, val_loss), flush=True) target_loss = train_loss if c.run_eval: target_loss = val_loss best_loss = save_best_model(model, optimizer, target_loss, best_loss, OUT_PATH, current_step, epoch)
def main(args): model = Tacotron(c.embedding_size, ap.num_freq, ap.num_mels, c.r) print(" | > Num output units : {}".format(ap.num_freq), flush=True) optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0) optimizer_st = optim.Adam( model.decoder.stopnet.parameters(), lr=c.lr, weight_decay=0) criterion = L1LossMasked() criterion_st = nn.BCELoss() if args.restore_path: checkpoint = torch.load(args.restore_path) try: model.load_state_dict(checkpoint['model']) except: model_dict = model.state_dict() # Partial initialization: if there is a mismatch with new and old layer, it is skipped. # 1. filter out unnecessary keys pretrained_dict = { k: v for k, v in checkpoint['model'].items() if k in model_dict } # 2. overwrite entries in the existing state dict model_dict.update(pretrained_dict) # 3. load the new state dict model.load_state_dict(model_dict) if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() optimizer.load_state_dict(checkpoint['optimizer']) print( " > Model restored from step %d" % checkpoint['step'], flush=True) start_epoch = checkpoint['epoch'] best_loss = checkpoint['linear_loss'] args.restore_step = checkpoint['step'] else: args.restore_step = 0 print("\n > Starting a new training", flush=True) if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() if c.lr_decay: scheduler = NoamLR( optimizer, warmup_steps=c.warmup_steps, last_epoch=args.restore_step - 1) else: scheduler = None num_params = count_parameters(model) print(" | > Model has {} parameters".format(num_params), flush=True) if not os.path.exists(CHECKPOINT_PATH): os.mkdir(CHECKPOINT_PATH) if 'best_loss' not in locals(): best_loss = float('inf') for epoch in range(0, c.epochs): train_loss, current_step = train(model, criterion, criterion_st, optimizer, optimizer_st, scheduler, ap, epoch) val_loss = evaluate(model, criterion, criterion_st, ap, current_step) print( " | > Train Loss: {:.5f} Validation Loss: {:.5f}".format( train_loss, val_loss), flush=True) best_loss = save_best_model(model, optimizer, train_loss, best_loss, OUT_PATH, current_step, epoch)
def main(args): # DISTRUBUTED if num_gpus > 1: init_distributed(args.rank, num_gpus, args.group_id, c.distributed["backend"], c.distributed["url"]) num_chars = len(phonemes) if c.use_phonemes else len(symbols) model = Tacotron(num_chars=num_chars, embedding_dim=c.embedding_size, linear_dim=ap.num_freq, mel_dim=ap.num_mels, r=c.r, memory_size=c.memory_size) optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0) optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr, weight_decay=0) criterion = L1LossMasked() criterion_st = nn.BCELoss() if args.restore_path: checkpoint = torch.load(args.restore_path) try: model.load_state_dict(checkpoint['model']) optimizer.load_state_dict(checkpoint['optimizer']) except: print(" > Partial model initialization.") partial_init_flag = True model_dict = model.state_dict() # Partial initialization: if there is a mismatch with new and old layer, it is skipped. # 1. filter out unnecessary keys pretrained_dict = { k: v for k, v in checkpoint['model'].items() if k in model_dict } # 2. filter out different size layers pretrained_dict = { k: v for k, v in pretrained_dict.items() if v.numel() == model_dict[k].numel() } # 3. overwrite entries in the existing state dict model_dict.update(pretrained_dict) # 4. load the new state dict model.load_state_dict(model_dict) print(" | > {} / {} layers are initialized".format( len(pretrained_dict), len(model_dict))) if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() for group in optimizer.param_groups: group['lr'] = c.lr print(" > Model restored from step %d" % checkpoint['step'], flush=True) start_epoch = checkpoint['epoch'] best_loss = checkpoint['linear_loss'] args.restore_step = checkpoint['step'] else: args.restore_step = 0 if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() # DISTRUBUTED if num_gpus > 1: model = apply_gradient_allreduce(model) if c.lr_decay: scheduler = NoamLR(optimizer, warmup_steps=c.warmup_steps, last_epoch=args.restore_step - 1) else: scheduler = None num_params = count_parameters(model) print("\n > Model has {} parameters".format(num_params), flush=True) if 'best_loss' not in locals(): best_loss = float('inf') for epoch in range(0, c.epochs): train_loss, current_step = train(model, criterion, criterion_st, optimizer, optimizer_st, scheduler, ap, epoch) val_loss = evaluate(model, criterion, criterion_st, ap, current_step, epoch) print(" | > Training Loss: {:.5f} Validation Loss: {:.5f}".format( train_loss, val_loss), flush=True) target_loss = train_loss if c.run_eval: target_loss = val_loss best_loss = save_best_model(model, optimizer, target_loss, best_loss, OUT_PATH, current_step, epoch)
def main(args): dataset = importlib.import_module('datasets.' + c.dataset) Dataset = getattr(dataset, 'MyDataset') audio = importlib.import_module('utils.' + c.audio_processor) AudioProcessor = getattr(audio, 'AudioProcessor') ap = AudioProcessor(sample_rate=c.sample_rate, num_mels=c.num_mels, min_level_db=c.min_level_db, frame_shift_ms=c.frame_shift_ms, frame_length_ms=c.frame_length_ms, ref_level_db=c.ref_level_db, num_freq=c.num_freq, power=c.power, preemphasis=c.preemphasis) # Setup the dataset train_dataset = Dataset(c.data_path, c.meta_file_train, c.r, c.text_cleaner, ap=ap, min_seq_len=c.min_seq_len) train_loader = DataLoader(train_dataset, batch_size=c.batch_size, shuffle=False, collate_fn=train_dataset.collate_fn, drop_last=False, num_workers=c.num_loader_workers, pin_memory=True) if c.run_eval: val_dataset = Dataset(c.data_path, c.meta_file_val, c.r, c.text_cleaner, ap=ap) val_loader = DataLoader(val_dataset, batch_size=c.eval_batch_size, shuffle=False, collate_fn=val_dataset.collate_fn, drop_last=False, num_workers=4, pin_memory=True) else: val_loader = None model = Tacotron(c.embedding_size, ap.num_freq, c.num_mels, c.r) print(" | > Num output units : {}".format(ap.num_freq), flush=True) optimizer = optim.Adam(model.parameters(), lr=c.lr) optimizer_st = optim.Adam(model.decoder.stopnet.parameters(), lr=c.lr) criterion = L1LossMasked() criterion_st = nn.BCELoss() if args.restore_path: checkpoint = torch.load(args.restore_path) model.load_state_dict(checkpoint['model']) if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() optimizer.load_state_dict(checkpoint['optimizer']) # optimizer_st.load_state_dict(checkpoint['optimizer_st']) for state in optimizer.state.values(): for k, v in state.items(): if torch.is_tensor(v): state[k] = v.cuda() print(" > Model restored from step %d" % checkpoint['step'], flush=True) start_epoch = checkpoint['step'] // len(train_loader) best_loss = checkpoint['linear_loss'] args.restore_step = checkpoint['step'] else: args.restore_step = 0 print("\n > Starting a new training", flush=True) if use_cuda: model = model.cuda() criterion.cuda() criterion_st.cuda() scheduler = AnnealLR(optimizer, warmup_steps=c.warmup_steps) num_params = count_parameters(model) print(" | > Model has {} parameters".format(num_params), flush=True) if not os.path.exists(CHECKPOINT_PATH): os.mkdir(CHECKPOINT_PATH) if 'best_loss' not in locals(): best_loss = float('inf') for epoch in range(0, c.epochs): train_loss, current_step = train(model, criterion, criterion_st, train_loader, optimizer, optimizer_st, scheduler, ap, epoch) val_loss = evaluate(model, criterion, criterion_st, val_loader, ap, current_step) print(" | > Train Loss: {:.5f} Validation Loss: {:.5f}".format( train_loss, val_loss), flush=True) best_loss = save_best_model(model, optimizer, train_loss, best_loss, OUT_PATH, current_step, epoch)