def train(config: DictConfig) -> nn.DataParallel:
random.seed(config.train.seed)
torch.manual_seed(config.train.seed)
torch.cuda.manual_seed_all(config.train.seed)
device = check_envirionment(config.train.use_cuda)
if hasattr(config.train,
"num_threads") and int(config.train.num_threads) > 0:
torch.set_num_threads(config.train.num_threads)
vocab = KsponSpeechVocabulary(
f'/home/seungmin/dmount/KoSpeech/data/vocab/aihub_{config.train.output_unit}_vocabs.csv',
output_unit=config.train.output_unit,
)
if not config.train.resume:
epoch_time_step, trainset_list, validset = split_dataset(
config, config.train.transcripts_path, vocab)
model = build_model(config, vocab, device)
optimizer = get_optimizer(model, config)
lr_scheduler = get_lr_scheduler(config, optimizer, epoch_time_step)
optimizer = Optimizer(optimizer, lr_scheduler,
config.train.total_steps,
config.train.max_grad_norm)
criterion = get_criterion(config, vocab)
else:
trainset_list = None
validset = None
model = None
optimizer = None
epoch_time_step = None
criterion = get_criterion(config, vocab)
trainer = SupervisedTrainer(
optimizer=optimizer,
criterion=criterion,
trainset_list=trainset_list,
validset=validset,
num_workers=config.train.num_workers,
device=device,
teacher_forcing_step=config.model.teacher_forcing_step,
min_teacher_forcing_ratio=config.model.min_teacher_forcing_ratio,
print_every=config.train.print_every,
save_result_every=config.train.save_result_every,
checkpoint_every=config.train.checkpoint_every,
architecture=config.model.architecture,
vocab=vocab,
joint_ctc_attention=config.model.joint_ctc_attention,
)
model = trainer.train(
model=model,
batch_size=config.train.batch_size,
epoch_time_step=epoch_time_step,
num_epochs=config.train.num_epochs,
teacher_forcing_ratio=config.model.teacher_forcing_ratio,
resume=config.train.resume,
)
return model
def train(opt):
random.seed(opt.seed)
torch.manual_seed(opt.seed)
torch.cuda.manual_seed_all(opt.seed)
device = check_envirionment(opt.use_cuda)
if not opt.resume:
audio_paths, script_paths = load_data_list(opt.data_list_path, opt.dataset_path)
epoch_time_step, trainset_list, validset = split_dataset(opt, audio_paths, script_paths)
model = build_model(opt, device)
optimizer = optim.Adam(model.module.parameters(), lr=opt.init_lr, weight_decay=opt.weight_decay)
if opt.rampup_period > 0:
scheduler = RampUpLR(optimizer, opt.init_lr, opt.high_plateau_lr, opt.rampup_period)
optimizer = Optimizer(optimizer, scheduler, opt.rampup_period, opt.max_grad_norm)
else:
optimizer = Optimizer(optimizer, None, 0, opt.max_grad_norm)
criterion = LabelSmoothedCrossEntropyLoss(
num_classes=len(char2id), ignore_index=PAD_token,
smoothing=opt.label_smoothing, dim=-1,
reduction=opt.reduction, architecture=opt.architecture
).to(device)
else:
trainset_list = None
validset = None
model = None
optimizer = None
criterion = LabelSmoothedCrossEntropyLoss(
num_classes=len(char2id), ignore_index=PAD_token,
smoothing=opt.label_smoothing, dim=-1,
reduction=opt.reduction, architecture=opt.architecture
).to(device)
epoch_time_step = None
trainer = SupervisedTrainer(
optimizer=optimizer, criterion=criterion, trainset_list=trainset_list,
validset=validset, num_workers=opt.num_workers,
high_plateau_lr=opt.high_plateau_lr, low_plateau_lr=opt.low_plateau_lr,
decay_threshold=opt.decay_threshold, exp_decay_period=opt.exp_decay_period,
device=device, teacher_forcing_step=opt.teacher_forcing_step,
min_teacher_forcing_ratio=opt.min_teacher_forcing_ratio, print_every=opt.print_every,
save_result_every=opt.save_result_every, checkpoint_every=opt.checkpoint_every,
architecture=opt.architecture
)
model = trainer.train(
model=model,
batch_size=opt.batch_size,
epoch_time_step=epoch_time_step,
num_epochs=opt.num_epochs,
teacher_forcing_ratio=opt.teacher_forcing_ratio,
resume=opt.resume
)
return model
def train(config: DictConfig) -> nn.DataParallel:
random.seed(config.train.seed)
torch.manual_seed(config.train.seed)
torch.cuda.manual_seed_all(config.train.seed)
device = check_envirionment(config.train.use_cuda)
if config.train.dataset == 'kspon':
if config.train.output_unit == 'subword':
vocab = KsponSpeechVocabulary(
vocab_path=KSPONSPEECH_VOCAB_PATH,
output_unit=config.train.output_unit,
sp_model_path=KSPONSPEECH_SP_MODEL_PATH,
)
else:
vocab = KsponSpeechVocabulary(
f'../../../data/vocab/aihub_{config.train.output_unit}_vocabs.csv',
output_unit=config.train.output_unit,
)
elif config.train.dataset == 'libri':
vocab = LibriSpeechVocabulary(LIBRISPEECH_VOCAB_PATH, LIBRISPEECH_TOKENIZER_PATH)
else:
raise ValueError("Unsupported Dataset : {0}".format(config.train.dataset))
if not config.train.resume:
epoch_time_step, trainset_list, validset = split_dataset(config, config.train.transcripts_path, vocab)
model = build_model(config, vocab, device)
optimizer = get_optimizer(model, config)
lr_scheduler = get_lr_scheduler(config, optimizer, epoch_time_step)
optimizer = Optimizer(optimizer, lr_scheduler, config.train.warmup_steps, config.train.max_grad_norm)
criterion = get_criterion(config, vocab)
else:
trainset_list = None
validset = None
model = None
optimizer = None
epoch_time_step = None
criterion = get_criterion(config, vocab)
trainer = SupervisedTrainer(
optimizer=optimizer,
criterion=criterion,
trainset_list=trainset_list,
validset=validset,
num_workers=config.train.num_workers,
device=device,
teacher_forcing_step=config.model.teacher_forcing_step,
min_teacher_forcing_ratio=config.model.min_teacher_forcing_ratio,
print_every=config.train.print_every,
save_result_every=config.train.save_result_every,
checkpoint_every=config.train.checkpoint_every,
architecture=config.model.architecture,
vocab=vocab,
joint_ctc_attention=config.model.joint_ctc_attention,
)
model = trainer.train(
model=model,
batch_size=config.train.batch_size,
epoch_time_step=epoch_time_step,
num_epochs=config.train.num_epochs,
teacher_forcing_ratio=config.model.teacher_forcing_ratio,
resume=config.train.resume,
)
return model
def train(config: DictConfig):
random.seed(config.train.seed)
torch.manual_seed(config.train.seed)
torch.cuda.manual_seed_all(config.train.seed)
device = check_envirionment(config.train.use_cuda)
if config.train.dataset == 'kspon':
if config.train.output_unit == 'subword':
vocab = KsponSpeechVocabulary(
vocab_path='../../../data/vocab/kspon_sentencepiece.vocab',
output_unit=config.train.output_unit,
sp_model_path='../../../data/vocab/kspon_sentencepiece.model',
)
else:
vocab = KsponSpeechVocabulary(
f'../../../data/vocab/aihub_{config.train.output_unit}_vocabs.csv',
output_unit=config.train.output_unit)
elif config.train.dataset == 'libri':
vocab = LibriSpeechVocabulary('../../../data/vocab/tokenizer.vocab',
'../../../data/vocab/tokenizer.model')
else:
raise ValueError("Unsupported Dataset : {0}".format(
config.train.dataset))
if not config.train.resume:
epoch_time_step, trainset_list, validset = split_dataset(
config, config.train.transcripts_path, vocab)
model = build_model(config, vocab, device)
optimizer = get_optimizer(model, config)
lr_scheduler = TriStageLRScheduler(
optimizer=optimizer,
init_lr=config.train.init_lr,
peak_lr=config.train.peak_lr,
final_lr=config.train.final_lr,
init_lr_scale=config.train.init_lr_scale,
final_lr_scale=config.train.final_lr_scale,
warmup_steps=config.train.warmup_steps,
total_steps=int(config.train.num_epochs * epoch_time_step))
optimizer = Optimizer(optimizer, lr_scheduler,
config.train.warmup_steps,
config.train.max_grad_norm)
criterion = get_criterion(config, vocab)
else:
trainset_list = None
validset = None
model = None
optimizer = None
epoch_time_step = None
criterion = get_criterion(config, vocab)
trainer = SupervisedTrainer(
optimizer=optimizer,
criterion=criterion,
trainset_list=trainset_list,
validset=validset,
num_workers=config.train.num_workers,
device=device,
teacher_forcing_step=config.model.teacher_forcing_step,
min_teacher_forcing_ratio=config.model.min_teacher_forcing_ratio,
print_every=config.train.print_every,
save_result_every=config.train.save_result_every,
checkpoint_every=config.train.checkpoint_every,
architecture=config.model.architecture,
vocab=vocab,
joint_ctc_attention=config.model.joint_ctc_attention,
)
model = trainer.train(
model=model,
batch_size=config.train.batch_size,
epoch_time_step=epoch_time_step,
num_epochs=config.train.num_epochs,
teacher_forcing_ratio=config.model.teacher_forcing_ratio,
resume=config.train.resume,
)
return model
def train(opt):
random.seed(opt.seed)
torch.manual_seed(opt.seed)
torch.cuda.manual_seed_all(opt.seed)
device = check_envirionment(opt.use_cuda)
if opt.dataset == 'kspon':
if opt.output_unit == 'subword':
vocab = KsponSpeechVocabulary(
vocab_path='../data/vocab/kspon_sentencepiece.vocab',
output_unit=opt.output_unit,
sp_model_path='../data/vocab/kspon_sentencepiece.model')
else:
vocab = KsponSpeechVocabulary(
f'../data/vocab/aihub_{opt.output_unit}_vocabs.csv',
output_unit=opt.output_unit)
elif opt.dataset == 'libri':
vocab = LibriSpeechVocabulary('../data/vocab/tokenizer.vocab',
'../data/vocab/tokenizer.model')
else:
raise ValueError("Unsupported Dataset : {0}".format(opt.dataset))
if not opt.resume:
epoch_time_step, trainset_list, validset = split_dataset(
opt, opt.transcripts_path, vocab)
model = build_model(opt, vocab, device)
if opt.optimizer.lower() == 'adam':
optimizer = optim.Adam(model.module.parameters(),
lr=opt.init_lr,
weight_decay=opt.weight_decay)
elif opt.optimizer.lower() == 'radam':
optimizer = RAdam(model.module.parameters(),
lr=opt.init_lr,
weight_decay=opt.weight_decay)
elif opt.optimizer.lower() == 'adamp':
optimizer = AdamP(model.module.parameters(),
lr=opt.init_lr,
weight_decay=opt.weight_decay)
elif opt.optimizer.lower() == 'adadelta':
optimizer = optim.Adadelta(model.module.parameters(),
lr=opt.init_lr,
weight_decay=opt.weight_decay)
elif opt.optimizer.lower() == 'adagrad':
optimizer = optim.Adagrad(model.module.parameters(),
lr=opt.init_lr,
weight_decay=opt.weight_decay)
else:
raise ValueError(
f"Unsupported Optimizer, Supported Optimizer : Adam, RAdam, Adadelta, Adagrad"
)
lr_scheduler = TriStageLRScheduler(optimizer=optimizer,
init_lr=opt.init_lr,
peak_lr=opt.peak_lr,
final_lr=opt.final_lr,
init_lr_scale=opt.init_lr_scale,
final_lr_scale=opt.final_lr_scale,
warmup_steps=opt.warmup_steps,
total_steps=int(opt.num_epochs *
epoch_time_step))
optimizer = Optimizer(optimizer, lr_scheduler, opt.warmup_steps,
opt.max_grad_norm)
if opt.architecture == 'deepspeech2':
criterion = nn.CTCLoss(blank=vocab.blank_id,
reduction=opt.reduction).to(device)
else:
criterion = LabelSmoothedCrossEntropyLoss(
num_classes=len(vocab),
ignore_index=vocab.pad_id,
smoothing=opt.label_smoothing,
dim=-1,
reduction=opt.reduction,
architecture=opt.architecture).to(device)
else:
trainset_list = None
validset = None
model = None
optimizer = None
criterion = LabelSmoothedCrossEntropyLoss(
num_classes=len(vocab),
ignore_index=vocab.pad_id,
smoothing=opt.label_smoothing,
dim=-1,
reduction=opt.reduction,
architecture=opt.architecture).to(device)
epoch_time_step = None
trainer = SupervisedTrainer(
optimizer=optimizer,
criterion=criterion,
trainset_list=trainset_list,
validset=validset,
num_workers=opt.num_workers,
device=device,
teacher_forcing_step=opt.teacher_forcing_step,
min_teacher_forcing_ratio=opt.min_teacher_forcing_ratio,
print_every=opt.print_every,
save_result_every=opt.save_result_every,
checkpoint_every=opt.checkpoint_every,
architecture=opt.architecture,
vocab=vocab)
model = trainer.train(model=model,
batch_size=opt.batch_size,
epoch_time_step=epoch_time_step,
num_epochs=opt.num_epochs,
teacher_forcing_ratio=opt.teacher_forcing_ratio,
resume=opt.resume)
return model