def main():
args = parse()
hvd.init()
torch.cuda.set_device(hvd.local_rank())
hvd_rank = hvd.rank()
# Load a conf file
if args.resume:
conf = load_config(os.path.join(os.path.dirname(args.resume), 'conf.yml'))
for k, v in conf.items():
if k != 'resume':
setattr(args, k, v)
# Load dataset
train_set = Dataset(corpus=args.corpus,
tsv_path=args.train_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size,
n_epochs=args.n_epochs,
min_n_tokens=args.min_n_tokens,
bptt=args.bptt,
n_customers=hvd.size(),
backward=args.backward,
serialize=args.serialize)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size,
bptt=args.bptt,
n_customers=hvd.size(),
backward=args.backward,
serialize=args.serialize)
eval_set = Dataset(corpus=args.corpus,
tsv_path=args.eval_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size,
bptt=args.bptt,
n_customers=hvd.size(),
backward=args.backward,
serialize=args.serialize)
args.vocab = train_set.vocab
train_loader = ChunkDataloader(train_set,
batch_size=1,
num_workers = 1,
distributed=True,
shuffle=False)
eval_loader = ChunkDataloader(eval_set,
batch_size=1,
num_workers=1,
distributed=True)
# Set save path
if args.resume:
save_path = os.path.dirname(args.resume)
dir_name = os.path.basename(save_path)
else:
dir_name = set_lm_name(args)
save_path = mkdir_join(args.model_save_dir, '_'.join(
os.path.basename(args.train_set).split('.')[:-1]), dir_name)
if hvd.rank() == 0:
save_path = set_save_path(save_path) # avoid overwriting
# Set logger
if hvd_rank == 0:
logger = set_logger(os.path.join(save_path, 'train.log'),
key='training', stdout=args.stdout)
# Set process name
logger.info('PID: %s' % os.getpid())
logger.info('USERNAME: %s' % os.uname()[1])
logger.info('NUMBER_DEVICES: %s' % hvd.size())
setproctitle(args.job_name if args.job_name else dir_name)
# Model setting
model = build_lm(args, save_path)
# GPU setting
if args.n_gpus >= 1:
torch.backends.cudnn.benchmark = True
model.cuda()
if args.resume:
# Set optimizer
epoch = int(args.resume.split('-')[-1])
optimizer = set_optimizer(model, 'sgd' if epoch > conf['convert_to_sgd_epoch'] else conf['optimizer'],
conf['lr'], conf['weight_decay'])
# Restore the last saved model
if hvd_rank == 0:
model, optimizer = load_checkpoint(model, args.resume, optimizer, resume=True)
#broadcast
optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters())
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
# Wrap optimizer by learning rate scheduler
optimizer = LRScheduler(optimizer, conf['lr'],
decay_type=conf['lr_decay_type'],
decay_start_epoch=conf['lr_decay_start_epoch'],
decay_rate=conf['lr_decay_rate'],
decay_patient_n_epochs=conf['lr_decay_patient_n_epochs'],
early_stop_patient_n_epochs=conf['early_stop_patient_n_epochs'],
warmup_start_lr=conf['warmup_start_lr'],
warmup_n_steps=conf['warmup_n_steps'],
model_size=conf['d_model'],
factor=conf['lr_factor'],
noam=conf['lm_type'] == 'transformer')
# Resume between convert_to_sgd_epoch -1 and convert_to_sgd_epoch
if epoch == conf['convert_to_sgd_epoch']:
n_epochs = optimizer.n_epochs
n_steps = optimizer.n_steps
optimizer = set_optimizer(model, 'sgd', args.lr, conf['weight_decay'])
optimizer = LRScheduler(optimizer, args.lr,
decay_type='always',
decay_start_epoch=0,
decay_rate=0.5)
optimizer._epoch = n_epochs
optimizer._step = n_steps
if hvd_rank == 0:
logger.info('========== Convert to SGD ==========')
#broadcast
optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters())
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
else:
# Save the conf file as a yaml file
if hvd_rank == 0:
save_config(vars(args), os.path.join(save_path, 'conf.yml'))
# Save the nlsyms, dictionar, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(save_path, 'nlsyms.txt'))
shutil.copy(args.dict, os.path.join(save_path, 'dict.txt'))
if args.unit == 'wp':
shutil.copy(args.wp_model, os.path.join(save_path, 'wp.model'))
for k, v in sorted(vars(args).items(), key=lambda x: x[0]):
logger.info('%s: %s' % (k, str(v)))
# Count total parameters
for n in sorted(list(model.num_params_dict.keys())):
n_params = model.num_params_dict[n]
if hvd.rank() == 0:
logger.info("%s %d" % (n, n_params))
if hvd_rank == 0:
logger.info("Total %.2f M parameters" % (model.total_parameters / 1000000))
logger.info(model)
# Set optimizer
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
optimizer = set_optimizer(model, args.optimizer, args.lr, args.weight_decay)
optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters())
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
# Wrap optimizer by learning rate scheduler
optimizer = LRScheduler(optimizer, args.lr,
decay_type=args.lr_decay_type,
decay_start_epoch=args.lr_decay_start_epoch,
decay_rate=args.lr_decay_rate,
decay_patient_n_epochs=args.lr_decay_patient_n_epochs,
early_stop_patient_n_epochs=args.early_stop_patient_n_epochs,
warmup_start_lr=args.warmup_start_lr,
warmup_n_steps=args.warmup_n_steps,
model_size=args.d_model,
factor=args.lr_factor,
noam=args.lm_type == 'transformer')
# Set reporter
reporter = Reporter(save_path)
hidden = None
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
data_size = len(train_set)
accum_n_tokens = 0
verbose = 1 if hvd_rank == 0 else 0
while True:
model.train()
with tqdm(total=data_size/hvd.size(),
desc='Train Epoch #{}'.format(optimizer.n_epochs + 1),
disable=not verbose) as pbar_epoch:
# Compute loss in the training set
for _, ys_train in enumerate(train_loader):
accum_n_tokens += sum([len(y) for y in ys_train])
optimizer.zero_grad()
loss, hidden, reporter = model(ys_train, hidden, reporter)
loss.backward()
loss.detach() # Trancate the graph
if args.accum_grad_n_tokens == 0 or accum_n_tokens >= args.accum_grad_n_tokens:
if args.clip_grad_norm > 0:
total_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), args.clip_grad_norm)
#reporter.add_tensorboard_scalar('total_norm', total_norm)
optimizer.step()
optimizer.zero_grad()
accum_n_tokens = 0
loss_train = loss.item()
del loss
hidden = model.repackage_state(hidden)
if optimizer.n_steps % args.print_step == 0:
model.eval()
# Compute loss in the dev set
ys_dev = dev_set.next()[0]
loss, _, reporter = model(ys_dev, None, reporter, is_eval=True)
loss_dev = loss.item()
del loss
duration_step = time.time() - start_time_step
if hvd_rank == 0:
logger.info("step:%d(ep:%.2f) loss:%.3f(%.3f)/ppl:%.3f(%.3f)/lr:%.5f/bs:%d (%.2f min)" %
(optimizer.n_steps, optimizer.n_steps/data_size*hvd.size(),
loss_train, loss_dev,
np.exp(loss_train), np.exp(loss_dev),
optimizer.lr, ys_train.shape[0], duration_step / 60))
start_time_step = time.time()
pbar_epoch.update(1)
# Save checkpoint and evaluate model per epoch
duration_epoch = time.time() - start_time_epoch
if hvd_rank == 0:
logger.info('========== EPOCH:%d (%.2f min) ==========' %(optimizer.n_epochs + 1, duration_epoch / 60))
if optimizer.n_epochs + 1 < args.eval_start_epoch:
# Save the model
if hvd_rank == 0:
optimizer.epoch()
save_checkpoint(model, save_path, optimizer, optimizer.n_epochs,
remove_old_checkpoints=args.lm_type != 'transformer')
else:
start_time_eval = time.time()
# dev
model.eval()
ppl_dev, _ = eval_ppl_parallel([model], eval_loader, optimizer.n_epochs, batch_size=args.batch_size)
ppl_dev = hvd.allreduce(np2tensor(np.array([ppl_dev], dtype=float), hvd.local_rank()))
if hvd_rank == 0:
logger.info('PPL : %.2f' % ppl_dev)
optimizer.epoch(ppl_dev)
if optimizer.is_best and hvd.rank() == 0:
# Save the model
save_checkpoint(model, save_path, optimizer, optimizer.n_epochs,
remove_old_checkpoints=args.lm_type != 'transformer')
duration_eval = time.time() - start_time_eval
if hvd_rank == 0:
logger.info('Evaluation time: %.2f min' % (duration_eval / 60))
# Early stopping
if optimizer.is_early_stop:
break
# Convert to fine-tuning stage
if optimizer.n_epochs == args.convert_to_sgd_epoch:
n_epochs = optimizer.n_epochs
n_steps = optimizer.n_steps
optimizer = set_optimizer(model, 'sgd', args.lr, args.weight_decay)
optimizer = hvd.DistributedOptimizer(
optimizer, named_parameters=model.named_parameters())
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
optimizer = LRScheduler(optimizer, args.lr,
decay_type='always',
decay_start_epoch=0,
decay_rate=0.5)
optimizer._epoch = n_epochs
optimizer._step = n_steps
if hvd_rank == 0:
logger.info('========== Convert to SGD ==========')
if optimizer.n_epochs == args.n_epochs:
break
start_time_step = time.time()
start_time_epoch = time.time()
duration_train = time.time() - start_time_train
if hvd_rank == 0:
logger.info('Total time: %.2f hour' % (duration_train / 3600))
reporter.tf_writer.close()
return save_path
def main():
args = parse()
# Load a conf file
if args.resume:
conf = load_config(
os.path.join(os.path.dirname(args.resume), 'conf.yml'))
for k, v in conf.items():
if k != 'resume':
setattr(args, k, v)
# Load dataset
train_set = Dataset(corpus=args.corpus,
tsv_path=args.train_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size * args.n_gpus,
n_epochs=args.n_epochs,
min_n_tokens=args.min_n_tokens,
bptt=args.bptt,
shuffle=args.shuffle,
backward=args.backward,
serialize=args.serialize)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size * args.n_gpus,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize)
eval_sets = [
Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=1,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize) for s in args.eval_sets
]
args.vocab = train_set.vocab
# Set save path
if args.resume:
save_path = os.path.dirname(args.resume)
dir_name = os.path.basename(save_path)
else:
dir_name = set_lm_name(args)
save_path = mkdir_join(
args.model_save_dir,
'_'.join(os.path.basename(args.train_set).split('.')[:-1]),
dir_name)
save_path = set_save_path(save_path) # avoid overwriting
# Set logger
set_logger(os.path.join(save_path, 'train.log'), stdout=args.stdout)
# Model setting
model = build_lm(args, save_path)
if args.resume:
transformer = conf['lm_type'] in ['transformer', 'transformer_xl']
else:
transformer = args.lm_type in ['transformer', 'transformer_xl']
if args.resume:
# Set optimizer
epoch = int(args.resume.split('-')[-1])
optimizer = set_optimizer(
model, 'sgd' if epoch > conf['convert_to_sgd_epoch'] else
conf['optimizer'], conf['lr'], conf['weight_decay'])
# Wrap optimizer by learning rate scheduler
optimizer = LRScheduler(
optimizer,
conf['lr'],
decay_type=conf['lr_decay_type'],
decay_start_epoch=conf['lr_decay_start_epoch'],
decay_rate=conf['lr_decay_rate'],
decay_patient_n_epochs=conf['lr_decay_patient_n_epochs'],
early_stop_patient_n_epochs=conf['early_stop_patient_n_epochs'],
warmup_start_lr=conf['warmup_start_lr'],
warmup_n_steps=conf['warmup_n_steps'],
model_size=conf['transformer_d_model'],
factor=conf['lr_factor'],
noam=transformer,
save_checkpoints_topk=1)
# Restore the last saved model
load_checkpoint(model, args.resume, optimizer)
# Resume between convert_to_sgd_epoch -1 and convert_to_sgd_epoch
if epoch == conf['convert_to_sgd_epoch']:
optimizer.convert_to_sgd(model,
args.lr,
conf['weight_decay'],
decay_type='always',
decay_rate=0.5)
else:
# Save the conf file as a yaml file
save_config(vars(args), os.path.join(save_path, 'conf.yml'))
# Save the nlsyms, dictionar, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(save_path, 'nlsyms.txt'))
shutil.copy(args.dict, os.path.join(save_path, 'dict.txt'))
if args.unit == 'wp':
shutil.copy(args.wp_model, os.path.join(save_path, 'wp.model'))
for k, v in sorted(vars(args).items(), key=lambda x: x[0]):
logger.info('%s: %s' % (k, str(v)))
# Count total parameters
for n in sorted(list(model.num_params_dict.keys())):
n_params = model.num_params_dict[n]
logger.info("%s %d" % (n, n_params))
logger.info("Total %.2f M parameters" %
(model.total_parameters / 1000000))
logger.info(model)
# Set optimizer
optimizer = set_optimizer(model, args.optimizer, args.lr,
args.weight_decay)
# Wrap optimizer by learning rate scheduler
optimizer = LRScheduler(
optimizer,
args.lr,
decay_type=args.lr_decay_type,
decay_start_epoch=args.lr_decay_start_epoch,
decay_rate=args.lr_decay_rate,
decay_patient_n_epochs=args.lr_decay_patient_n_epochs,
early_stop_patient_n_epochs=args.early_stop_patient_n_epochs,
warmup_start_lr=args.warmup_start_lr,
warmup_n_steps=args.warmup_n_steps,
model_size=args.transformer_d_model,
factor=args.lr_factor,
noam=transformer,
save_checkpoints_topk=1)
# GPU setting
if args.n_gpus >= 1:
model.cudnn_setting(deterministic=False,
benchmark=args.cudnn_benchmark)
model = CustomDataParallel(model,
device_ids=list(range(0, args.n_gpus)))
model.cuda()
# Set process name
logger.info('PID: %s' % os.getpid())
logger.info('USERNAME: %s' % os.uname()[1])
setproctitle(args.job_name if args.job_name else dir_name)
# Set reporter
reporter = Reporter(save_path)
hidden = None
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
pbar_epoch = tqdm(total=len(train_set))
accum_n_steps = 0
n_steps = optimizer.n_steps * args.accum_grad_n_steps
while True:
# Compute loss in the training set
ys_train, is_new_epoch = train_set.next()
accum_n_steps += 1
loss, hidden, observation = model(ys_train, hidden)
reporter.add(observation)
loss.backward()
loss.detach() # Trancate the graph
if args.accum_grad_n_steps == 1 or accum_n_steps >= args.accum_grad_n_steps:
if args.clip_grad_norm > 0:
total_norm = torch.nn.utils.clip_grad_norm_(
model.module.parameters(), args.clip_grad_norm)
reporter.add_tensorboard_scalar('total_norm', total_norm)
optimizer.step()
optimizer.zero_grad()
accum_n_steps = 0
loss_train = loss.item()
del loss
hidden = model.module.repackage_state(hidden)
reporter.add_tensorboard_scalar('learning_rate', optimizer.lr)
# NOTE: loss/acc/ppl are already added in the model
reporter.step()
pbar_epoch.update(ys_train.shape[0] * (ys_train.shape[1] - 1))
n_steps += 1
if n_steps % args.print_step == 0:
# Compute loss in the dev set
ys_dev = dev_set.next(bptt=args.bptt)[0]
loss, _, observation = model(ys_dev, None, is_eval=True)
reporter.add(observation, is_eval=True)
loss_dev = loss.item()
del loss
reporter.step(is_eval=True)
duration_step = time.time() - start_time_step
logger.info(
"step:%d(ep:%.2f) loss:%.3f(%.3f)/lr:%.5f/bs:%d (%.2f min)" %
(n_steps, optimizer.n_epochs + train_set.epoch_detail,
loss_train, loss_dev, optimizer.lr, ys_train.shape[0],
duration_step / 60))
start_time_step = time.time()
# Save fugures of loss and accuracy
if n_steps % (args.print_step * 10) == 0:
reporter.snapshot()
model.module.plot_attention()
# Save checkpoint and evaluate model per epoch
if is_new_epoch:
duration_epoch = time.time() - start_time_epoch
logger.info('========== EPOCH:%d (%.2f min) ==========' %
(optimizer.n_epochs + 1, duration_epoch / 60))
if optimizer.n_epochs + 1 < args.eval_start_epoch:
optimizer.epoch() # lr decay
reporter.epoch() # plot
# Save the model
optimizer.save_checkpoint(model, save_path, remove_old=True)
else:
start_time_eval = time.time()
# dev
model.module.reset_length(args.bptt)
ppl_dev, _ = eval_ppl([model.module],
dev_set,
batch_size=1,
bptt=args.bptt)
model.module.reset_length(args.bptt)
optimizer.epoch(ppl_dev) # lr decay
reporter.epoch(ppl_dev, name='perplexity') # plot
logger.info('PPL (%s, ep:%d): %.2f' %
(dev_set.set, optimizer.n_epochs, ppl_dev))
if optimizer.is_topk:
# Save the model
optimizer.save_checkpoint(model,
save_path,
remove_old=True)
# test
ppl_test_avg = 0.
for eval_set in eval_sets:
model.module.reset_length(args.bptt)
ppl_test, _ = eval_ppl([model.module],
eval_set,
batch_size=1,
bptt=args.bptt)
model.module.reset_length(args.bptt)
logger.info(
'PPL (%s, ep:%d): %.2f' %
(eval_set.set, optimizer.n_epochs, ppl_test))
ppl_test_avg += ppl_test
if len(eval_sets) > 0:
logger.info('PPL (avg., ep:%d): %.2f' %
(optimizer.n_epochs,
ppl_test_avg / len(eval_sets)))
duration_eval = time.time() - start_time_eval
logger.info('Evaluation time: %.2f min' % (duration_eval / 60))
# Early stopping
if optimizer.is_early_stop:
break
# Convert to fine-tuning stage
if optimizer.n_epochs == args.convert_to_sgd_epoch:
optimizer.convert_to_sgd(model,
args.lr,
args.weight_decay,
decay_type='always',
decay_rate=0.5)
pbar_epoch = tqdm(total=len(train_set))
if optimizer.n_epochs == args.n_epochs:
break
start_time_step = time.time()
start_time_epoch = time.time()
duration_train = time.time() - start_time_train
logger.info('Total time: %.2f hour' % (duration_train / 3600))
reporter.tf_writer.close()
pbar_epoch.close()
return save_path
def main():
args = parse()
# Load a conf file
if args.resume:
conf = load_config(
os.path.join(os.path.dirname(args.resume), 'conf.yml'))
for k, v in conf.items():
if k != 'resume':
setattr(args, k, v)
# Load dataset
train_set = Dataset(corpus=args.corpus,
tsv_path=args.train_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size * args.n_gpus,
n_epochs=args.n_epochs,
min_n_tokens=args.min_n_tokens,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size * args.n_gpus,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize)
eval_sets = []
for s in args.eval_sets:
eval_sets += [
Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=1,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize)
]
args.vocab = train_set.vocab
# Set save path
if args.resume:
save_path = os.path.dirname(args.resume)
dir_name = os.path.basename(save_path)
else:
dir_name = set_lm_name(args)
save_path = mkdir_join(
args.model_save_dir,
'_'.join(os.path.basename(args.train_set).split('.')[:-1]),
dir_name)
save_path = set_save_path(save_path) # avoid overwriting
# Set logger
logger = set_logger(os.path.join(save_path, 'train.log'), key='training')
# Model setting
model = select_lm(args, save_path)
if args.resume:
# Set optimizer
epoch = int(args.resume.split('-')[-1])
optimizer = set_optimizer(
model,
optimizer='sgd'
if epoch > conf['convert_to_sgd_epoch'] else conf['optimizer'],
lr=float(conf['learning_rate']), # on-the-fly
weight_decay=float(conf['weight_decay']))
# Restore the last saved model
model, checkpoint = load_checkpoint(model, args.resume, resume=True)
optimizer = checkpoint['optimizer']
epoch = checkpoint['epoch']
step = checkpoint['step']
ppl_dev_best = checkpoint['metric_dev_best']
# Resume between convert_to_sgd_epoch -1 and convert_to_sgd_epoch
if epoch == conf['convert_to_sgd_epoch']:
optimizer = set_optimizer(model,
optimizer='sgd',
lr=float(args.learning_rate),
weight_decay=float(conf['weight_decay']))
optimizer = LRScheduler(optimizer,
lr_max=args.learning_rate,
decay_type='epoch',
decay_start_epoch=0,
decay_rate=0.5,
lower_better=True)
logger.info('========== Convert to SGD ==========')
else:
# Save the conf file as a yaml file
save_config(vars(args), os.path.join(save_path, 'conf.yml'))
# Save the nlsyms, dictionar, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(save_path, 'nlsyms.txt'))
shutil.copy(args.dict, os.path.join(save_path, 'dict.txt'))
if args.unit == 'wp':
shutil.copy(args.wp_model, os.path.join(save_path, 'wp.model'))
for k, v in sorted(vars(args).items(), key=lambda x: x[0]):
logger.info('%s: %s' % (k, str(v)))
# Count total parameters
for n in sorted(list(model.num_params_dict.keys())):
nparams = model.num_params_dict[n]
logger.info("%s %d" % (n, nparams))
logger.info("Total %.2f M parameters" %
(model.total_parameters / 1000000))
logger.info(model)
epoch, step = 0, 0
ppl_dev_best = 10000
# Set optimizer
optimizer = set_optimizer(model,
optimizer=args.optimizer,
lr=float(args.learning_rate),
weight_decay=float(args.weight_decay))
# Wrap optimizer by learning rate scheduler
optimizer = LRScheduler(
optimizer,
lr_max=float(args.learning_rate),
decay_type=args.decay_type,
decay_start_epoch=args.decay_start_epoch,
decay_rate=args.decay_rate,
decay_patient_n_epochs=args.decay_patient_n_epochs,
lower_better=True,
best_value=ppl_dev_best,
model_size=args.d_model,
warmup_start_lr=args.warmup_start_learning_rate,
warmup_n_steps=args.warmup_n_steps,
lr_factor=args.learning_rate_factor,
noam=args.lm_type == 'transformer')
# GPU setting
if args.n_gpus >= 1:
model = CustomDataParallel(model,
device_ids=list(range(0, args.n_gpus, 1)),
deterministic=False,
benchmark=True)
model.cuda()
logger.info('PID: %s' % os.getpid())
logger.info('USERNAME: %s' % os.uname()[1])
# Set process name
if args.job_name:
setproctitle(args.job_name)
else:
setproctitle(dir_name)
# Set reporter
reporter = Reporter(save_path, tensorboard=True)
hidden = None
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
not_improved_n_epochs = 0
pbar_epoch = tqdm(total=len(train_set))
accum_n_tokens = 0
while True:
# Compute loss in the training set
ys_train, is_new_epoch = train_set.next()
accum_n_tokens += sum([len(y) for y in ys_train])
optimizer.zero_grad()
loss, hidden, reporter = model(ys_train, hidden, reporter)
# loss /= args.accum_grad_n_steps
if len(model.device_ids) > 1:
loss.backward(torch.ones(len(model.device_ids)))
else:
loss.backward()
loss.detach() # Trancate the graph
if args.accum_grad_n_tokens == 0 or accum_n_tokens >= args.accum_grad_n_tokens:
if args.clip_grad_norm > 0:
torch.nn.utils.clip_grad_norm_(model.module.parameters(),
args.clip_grad_norm)
optimizer.step()
optimizer.zero_grad()
accum_n_tokens = 0
loss_train = loss.item()
del loss
hidden = model.module.repackage_state(hidden)
reporter.step()
# step += args.n_gpus
if step % args.print_step == 0:
# Compute loss in the dev set
ys_dev = dev_set.next()[0]
loss, _, reporter = model(ys_dev, None, reporter, is_eval=True)
loss_dev = loss.item()
del loss
reporter.step(is_eval=True)
duration_step = time.time() - start_time_step
logger.info(
"step:%d(ep:%.2f) loss:%.3f(%.3f)/ppl:%.3f(%.3f)/lr:%.5f/bs:%d (%.2f min)"
% (step, epoch + train_set.epoch_detail, loss_train, loss_dev,
np.exp(loss_train), np.exp(loss_dev), optimizer.lr,
ys_train.shape[0], duration_step / 60))
start_time_step = time.time()
step += args.n_gpus
pbar_epoch.update(ys_train.shape[0] * (ys_train.shape[1] - 1))
# Save fugures of loss and accuracy
if step % (args.print_step * 10) == 0:
reporter.snapshot()
if args.lm_type == 'transformer':
model.module.plot_attention()
# Save checkpoint and evaluate model per epoch
if is_new_epoch:
epoch += 1
duration_epoch = time.time() - start_time_epoch
logger.info('========== EPOCH:%d (%.2f min) ==========' %
(epoch, duration_epoch / 60))
if epoch < args.eval_start_epoch:
# Save the model
save_checkpoint(
model,
save_path,
optimizer,
epoch,
step,
ppl_dev_best,
remove_old_checkpoints=args.lm_type != 'transformer')
else:
start_time_eval = time.time()
# dev
ppl_dev, _ = eval_ppl([model.module],
dev_set,
batch_size=1,
bptt=args.bptt)
logger.info('PPL (%s): %.2f' % (dev_set.set, ppl_dev))
# Update learning rate
optimizer.decay(epoch=epoch, value=ppl_dev)
if ppl_dev < ppl_dev_best:
ppl_dev_best = ppl_dev
not_improved_n_epochs = 0
logger.info('||||| Best Score |||||')
# Save the model
save_checkpoint(
model,
save_path,
optimizer,
epoch,
step,
ppl_dev_best,
remove_old_checkpoints=args.lm_type != 'transformer')
# test
ppl_test_avg = 0.
for eval_set in eval_sets:
ppl_test, _ = eval_ppl([model.module],
eval_set,
batch_size=1,
bptt=args.bptt)
logger.info('PPL (%s): %.2f' %
(eval_set.set, ppl_test))
ppl_test_avg += ppl_test
if len(eval_sets) > 0:
logger.info('PPL (avg.): %.2f' %
(ppl_test_avg / len(eval_sets)))
else:
not_improved_n_epochs += 1
duration_eval = time.time() - start_time_eval
logger.info('Evaluation time: %.2f min' % (duration_eval / 60))
# Early stopping
if not_improved_n_epochs == args.not_improved_patient_n_epochs:
break
# Convert to fine-tuning stage
if epoch == args.convert_to_sgd_epoch:
optimizer = set_optimizer(model,
optimizer='sgd',
lr=args.learning_rate,
weight_decay=float(
args.weight_decay))
optimizer = LRScheduler(optimizer,
lr_max=args.learning_rate,
decay_type='epoch',
decay_start_epoch=0,
decay_rate=0.5,
lower_better=True)
logger.info('========== Convert to SGD ==========')
pbar_epoch = tqdm(total=len(train_set))
if epoch == args.n_epochs:
break
start_time_step = time.time()
start_time_epoch = time.time()
duration_train = time.time() - start_time_train
logger.info('Total time: %.2f hour' % (duration_train / 3600))
if reporter.tensorboard:
reporter.tf_writer.close()
pbar_epoch.close()
return save_path