def main():
# Load configuration
args, _, dir_name = parse_args_eval(sys.argv[1:])
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'decode.log')):
os.remove(os.path.join(args.recog_dir, 'decode.log'))
set_logger(os.path.join(args.recog_dir, 'decode.log'),
stdout=args.recog_stdout)
ppl_avg = 0
for i, s in enumerate(args.recog_sets):
# Load dataset
dataset = Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=os.path.join(dir_name, 'dict.txt'),
wp_model=os.path.join(dir_name, 'wp.model'),
unit=args.unit,
batch_size=args.recog_batch_size,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize,
is_test=True)
if i == 0:
# Load the LM
model = build_lm(args)
load_checkpoint(args.recog_model[0], model)
epoch = int(args.recog_model[0].split('-')[-1])
# NOTE: model averaging is not helpful for LM
logger.info('epoch: %d' % epoch)
logger.info('batch size: %d' % args.recog_batch_size)
logger.info('BPTT: %d' % (args.bptt))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache theta: %.3f' % (args.recog_cache_theta))
logger.info('cache lambda: %.3f' % (args.recog_cache_lambda))
logger.info('model average (Transformer): %d' %
(args.recog_n_average))
model.cache_theta = args.recog_cache_theta
model.cache_lambda = args.recog_cache_lambda
# GPU setting
if args.recog_n_gpus > 0:
model.cuda()
start_time = time.time()
ppl, _ = eval_ppl([model],
dataset,
batch_size=1,
bptt=args.bptt,
n_caches=args.recog_n_caches,
progressbar=True)
ppl_avg += ppl
print('PPL (%s): %.2f' % (dataset.set, ppl))
logger.info('Elasped time: %.2f [sec]:' % (time.time() - start_time))
logger.info('PPL (avg.): %.2f\n' % (ppl_avg / len(args.recog_sets)))
def main():
args = parse_args_train(sys.argv[1:])
# 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
batch_size = args.batch_size * args.n_gpus if args.n_gpus >= 1 else args.batch_size
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=batch_size,
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=batch_size,
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 not args.resume:
# Save the conf file as a yaml file
save_config(vars(args), os.path.join(save_path, 'conf.yml'))
# Save the nlsyms, dictionary, 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
resume_epoch = 0
if args.resume:
epoch = int(args.resume.split('-')[-1])
optimizer = set_optimizer(
model,
'sgd' if epoch > args.convert_to_sgd_epoch else args.optimizer,
args.lr, args.weight_decay)
else:
optimizer = set_optimizer(model, args.optimizer, args.lr,
args.weight_decay)
# Wrap optimizer by learning rate scheduler
is_transformer = args.lm_type in ['transformer', 'transformer_xl']
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=getattr(args, 'transformer_d_model', 0),
factor=args.lr_factor,
noam=is_transformer,
save_checkpoints_topk=1)
if args.resume:
# Restore the last saved model
load_checkpoint(args.resume, model, optimizer)
# Resume between convert_to_sgd_epoch -1 and convert_to_sgd_epoch
if resume_epoch == args.convert_to_sgd_epoch:
optimizer.convert_to_sgd(model,
args.lr,
args.weight_decay,
decay_type='always',
decay_rate=0.5)
# GPU setting
use_apex = args.train_dtype in ["O0", "O1", "O2", "O3"]
amp = None
if args.n_gpus >= 1:
model.cudnn_setting(
deterministic=not (is_transformer or args.cudnn_benchmark),
benchmark=args.cudnn_benchmark)
model.cuda()
# Mix precision training setting
if use_apex:
from apex import amp
model, optimizer.optimizer = amp.initialize(
model, optimizer.optimizer, opt_level=args.train_dtype)
amp.init()
if args.resume:
load_checkpoint(args.resume, amp=amp)
model = CustomDataParallel(model,
device_ids=list(range(0, args.n_gpus)))
else:
model = CPUWrapperLM(model)
# Set process name
logger.info('PID: %s' % os.getpid())
logger.info('USERNAME: %s' % os.uname()[1])
logger.info('#GPU: %d' % torch.cuda.device_count())
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)
if use_apex:
with amp.scale_loss(loss, optimizer.optimizer) as scaled_loss:
scaled_loss.backward()
else:
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
# NOTE: n_steps is different from the step counter in Noam Optimizer
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=not is_transformer,
amp=amp)
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 or is_transformer:
# Save the model
optimizer.save_checkpoint(model,
save_path,
remove_old=not is_transformer,
amp=amp)
# 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
dir_name = os.path.dirname(args.recog_model[0])
conf = load_config(os.path.join(dir_name, 'conf.yml'))
# Overwrite conf
for k, v in conf.items():
if 'recog' not in k:
setattr(args, k, v)
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'plot.log')):
os.remove(os.path.join(args.recog_dir, 'plot.log'))
set_logger(os.path.join(args.recog_dir, 'plot.log'),
stdout=args.recog_stdout)
for i, s in enumerate(args.recog_sets):
# Load dataset
dataset = Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=os.path.join(dir_name, 'dict.txt'),
wp_model=os.path.join(dir_name, 'wp.model'),
unit=args.unit,
batch_size=args.recog_batch_size,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize,
is_test=True)
if i == 0:
# Load the LM
model = build_lm(args, dir_name)
topk_list = load_checkpoint(model, args.recog_model[0])
epoch = int(args.recog_model[0].split('-')[-1])
# Model averaging for Transformer
if conf['lm_type'] == 'transformer':
model = average_checkpoints(model,
args.recog_model[0],
n_average=args.recog_n_average,
topk_list=topk_list)
logger.info('epoch: %d' % (epoch - 1))
logger.info('batch size: %d' % args.recog_batch_size)
# logger.info('recog unit: %s' % args.recog_unit)
# logger.info('ensemble: %d' % (len(ensemble_models)))
logger.info('BPTT: %d' % (args.bptt))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache theta: %.3f' % (args.recog_cache_theta))
logger.info('cache lambda: %.3f' % (args.recog_cache_lambda))
model.cache_theta = args.recog_cache_theta
model.cache_lambda = args.recog_cache_lambda
# GPU setting
model.cuda()
assert args.recog_n_caches > 0
save_path = mkdir_join(args.recog_dir, 'cache')
# Clean directory
if save_path is not None and os.path.isdir(save_path):
shutil.rmtree(save_path)
os.mkdir(save_path)
hidden = None
fig_count = 0
toknen_count = 0
n_tokens = args.recog_n_caches
while True:
ys, is_new_epoch = dataset.next()
for t in range(ys.shape[1] - 1):
loss, hidden = model(ys[:, t:t + 2],
hidden,
is_eval=True,
n_caches=args.recog_n_caches)[:2]
if len(model.cache_attn) > 0:
if toknen_count == n_tokens:
tokens_keys = dataset.idx2token[0](
model.cache_ids[:args.recog_n_caches],
return_list=True)
tokens_query = dataset.idx2token[0](
model.cache_ids[-n_tokens:], return_list=True)
# Slide attention matrix
n_keys = len(tokens_keys)
n_queries = len(tokens_query)
cache_probs = np.zeros(
(n_keys, n_queries)) # `[n_keys, n_queries]`
mask = np.zeros((n_keys, n_queries))
for i, aw in enumerate(model.cache_attn[-n_tokens:]):
cache_probs[:(n_keys - n_queries + i + 1),
i] = aw[0,
-(n_keys - n_queries + i + 1):]
mask[(n_keys - n_queries + i + 1):, i] = 1
plot_cache_weights(cache_probs,
keys=tokens_keys,
queries=tokens_query,
save_path=mkdir_join(
save_path,
str(fig_count) + '.png'),
figsize=(40, 16),
mask=mask)
toknen_count = 0
fig_count += 1
else:
toknen_count += 1
if is_new_epoch:
break
def main():
args = parse()
# Load a conf file
dir_name = os.path.dirname(args.recog_model[0])
conf = load_config(os.path.join(dir_name, 'conf.yml'))
# Overwrite conf
for k, v in conf.items():
if 'recog' not in k:
setattr(args, k, v)
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'decode.log')):
os.remove(os.path.join(args.recog_dir, 'decode.log'))
set_logger(os.path.join(args.recog_dir, 'decode.log'), stdout=args.recog_stdout)
ppl_avg = 0
for i, s in enumerate(args.recog_sets):
# Load dataset
dataset = Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=os.path.join(dir_name, 'dict.txt'),
wp_model=os.path.join(dir_name, 'wp.model'),
unit=args.unit,
batch_size=args.recog_batch_size,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize,
is_test=True)
if i == 0:
# Load the LM
model = build_lm(args)
load_checkpoint(model, args.recog_model[0])
epoch = int(args.recog_model[0].split('-')[-1])
# Model averaging for Transformer
if conf['lm_type'] == 'transformer':
model = average_checkpoints(model, args.recog_model[0], epoch,
n_average=args.recog_n_average)
logger.info('epoch: %d' % epoch)
logger.info('batch size: %d' % args.recog_batch_size)
# logger.info('recog unit: %s' % args.recog_unit)
# logger.info('ensemble: %d' % (len(ensemble_models)))
logger.info('BPTT: %d' % (args.bptt))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache theta: %.3f' % (args.recog_cache_theta))
logger.info('cache lambda: %.3f' % (args.recog_cache_lambda))
logger.info('model average (Transformer): %d' % (args.recog_n_average))
model.cache_theta = args.recog_cache_theta
model.cache_lambda = args.recog_cache_lambda
# GPU setting
model.cuda()
start_time = time.time()
# TODO(hirofumi): ensemble
ppl, _ = eval_ppl([model], dataset, batch_size=1, bptt=args.bptt,
n_caches=args.recog_n_caches, progressbar=True)
ppl_avg += ppl
print('PPL (%s): %.2f' % (dataset.set, ppl))
logger.info('Elasped time: %.2f [sec]:' % (time.time() - start_time))
logger.info('PPL (avg.): %.2f\n' % (ppl_avg / len(args.recog_sets)))
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)
# 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',
stdout=args.stdout)
# 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
# Model setting
model = build_lm(args, save_path)
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['d_model'],
factor=conf['lr_factor'],
noam=conf['lm_type'] == 'transformer')
# Restore the last saved model
model, optimizer = load_checkpoint(model,
args.resume,
optimizer,
resume=True)
# 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
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())):
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.d_model,
factor=args.lr_factor,
noam=args.lm_type == 'transformer')
# GPU setting
if args.n_gpus >= 1:
torch.backends.cudnn.benchmark = True
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_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.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.module.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.module.repackage_state(hidden)
reporter.add_tensorboard_scalar('learning_rate', optimizer.lr)
# NOTE: loss/acc/ppl are already added in the model
reporter.step()
if optimizer.n_steps % 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)"
% (optimizer.n_steps,
optimizer.n_epochs + 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()
pbar_epoch.update(ys_train.shape[0] * (ys_train.shape[1] - 1))
# Save fugures of loss and accuracy
if optimizer.n_steps % (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:
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
save_checkpoint(
model,
save_path,
optimizer,
optimizer.n_epochs,
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, epoch:%d): %.2f' %
(dev_set.set, optimizer.n_epochs, ppl_dev))
optimizer.epoch(ppl_dev) # lr decay
reporter.epoch(ppl_dev, name='perplexity') # plot
if optimizer.is_best:
# Save the model
save_checkpoint(
model,
save_path,
optimizer,
optimizer.n_epochs,
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, epoch:%d): %.2f' %
(eval_set.set, optimizer.n_epochs, ppl_test))
ppl_test_avg += ppl_test
if len(eval_sets) > 0:
logger.info('PPL (avg., epoch:%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:
n_epochs = optimizer.n_epochs
n_steps = optimizer.n_steps
optimizer = set_optimizer(model, 'sgd', args.lr,
args.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
logger.info('========== Convert to SGD ==========')
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_args_train(sys.argv[1:])
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# 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)
# for multi-GPUs
if args.n_gpus > 1:
batch_size = args.batch_size * args.n_gpus
accum_grad_n_steps = max(1, args.accum_grad_n_steps // args.n_gpus)
else:
batch_size = args.batch_size
accum_grad_n_steps = args.accum_grad_n_steps
# 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=batch_size,
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=batch_size,
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:
args.save_path = os.path.dirname(args.resume)
dir_name = os.path.basename(args.save_path)
else:
dir_name = set_lm_name(args)
args.save_path = mkdir_join(
args.model_save_dir,
'_'.join(os.path.basename(args.train_set).split('.')[:-1]),
dir_name)
args.save_path = set_save_path(args.save_path) # avoid overwriting
# Set logger
set_logger(os.path.join(args.save_path, 'train.log'), stdout=args.stdout)
# Model setting
model = build_lm(args, args.save_path)
if not args.resume:
# Save nlsyms, dictionary, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(args.save_path,
'nlsyms.txt'))
shutil.copy(args.dict, os.path.join(args.save_path, 'dict.txt'))
if args.unit == 'wp':
shutil.copy(args.wp_model, os.path.join(args.save_path,
'wp.model'))
for k, v in sorted(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('torch version: %s' % str(torch.__version__))
logger.info(model)
# Set optimizer
resume_epoch = int(args.resume.split('-')[-1]) if args.resume else 0
optimizer = set_optimizer(
model,
'sgd' if resume_epoch > args.convert_to_sgd_epoch else args.optimizer,
args.lr, args.weight_decay)
# Wrap optimizer by learning rate scheduler
is_transformer = args.lm_type in ['transformer', 'transformer_xl']
scheduler = 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.get('transformer_d_model', 0),
factor=args.lr_factor,
noam=args.optimizer == 'noam',
save_checkpoints_topk=10 if is_transformer else 1)
if args.resume:
# Restore the last saved model
load_checkpoint(args.resume, model, scheduler)
# Resume between convert_to_sgd_epoch -1 and convert_to_sgd_epoch
if resume_epoch == args.convert_to_sgd_epoch:
scheduler.convert_to_sgd(model,
args.lr,
args.weight_decay,
decay_type='always',
decay_rate=0.5)
# GPU setting
args.use_apex = args.train_dtype in ["O0", "O1", "O2", "O3"]
amp, scaler = None, None
if args.n_gpus >= 1:
model.cudnn_setting(
deterministic=not (is_transformer or args.cudnn_benchmark),
benchmark=not is_transformer and args.cudnn_benchmark)
# Mixed precision training setting
if args.use_apex:
if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
scaler = torch.cuda.amp.GradScaler()
else:
from apex import amp
model, scheduler.optimizer = amp.initialize(
model, scheduler.optimizer, opt_level=args.train_dtype)
amp.init()
if args.resume:
load_checkpoint(args.resume, amp=amp)
model.cuda()
model = CustomDataParallel(model,
device_ids=list(range(0, args.n_gpus)))
else:
model = CPUWrapperLM(model)
# Set process name
logger.info('PID: %s' % os.getpid())
logger.info('USERNAME: %s' % os.uname()[1])
logger.info('#GPU: %d' % torch.cuda.device_count())
setproctitle(args.job_name if args.job_name else dir_name)
# Set reporter
reporter = Reporter(args, model)
if args.resume:
n_steps = scheduler.n_steps * accum_grad_n_steps
reporter.resume(n_steps, resume_epoch)
# Save conf file as a yaml file
if not args.resume:
save_config(args, os.path.join(args.save_path, 'conf.yml'))
# NOTE: save after reporter for wandb ID
hidden = None
start_time_train = time.time()
for ep in range(resume_epoch, args.n_epochs):
for ys_train, is_new_epoch in train_set:
hidden = train(model, train_set, dev_set, scheduler, reporter,
logger, args, accum_grad_n_steps, amp, scaler,
hidden)
# Save checkpoint and validate model per epoch
if reporter.n_epochs + 1 < args.eval_start_epoch:
scheduler.epoch() # lr decay
reporter.epoch() # plot
# Save model
scheduler.save_checkpoint(model,
args.save_path,
remove_old=not is_transformer
and args.remove_old_checkpoints,
amp=amp)
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)
scheduler.epoch(ppl_dev) # lr decay
reporter.epoch(ppl_dev, name='perplexity') # plot
reporter.add_scalar('dev/perplexity', ppl_dev)
logger.info('PPL (%s, ep:%d): %.2f' %
(dev_set.set, reporter.n_epochs, ppl_dev))
if scheduler.is_topk or is_transformer:
# Save model
scheduler.save_checkpoint(model,
args.save_path,
remove_old=not is_transformer
and args.remove_old_checkpoints,
amp=amp)
# 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, reporter.n_epochs, ppl_test))
ppl_test_avg += ppl_test
if len(eval_sets) > 0:
logger.info(
'PPL (avg., ep:%d): %.2f' %
(reporter.n_epochs, ppl_test_avg / len(eval_sets)))
logger.info('Evaluation time: %.2f min' %
((time.time() - start_time_eval) / 60))
# Early stopping
if scheduler.is_early_stop:
break
# Convert to fine-tuning stage
if reporter.n_epochs == args.convert_to_sgd_epoch:
scheduler.convert_to_sgd(model,
args.lr,
args.weight_decay,
decay_type='always',
decay_rate=0.5)
if reporter.n_epochs >= args.n_epochs:
break
logger.info('Total time: %.2f hour' %
((time.time() - start_time_train) / 3600))
reporter.close()
return args.save_path
def main():
# Load configuration
args, _, dir_name = parse_args_eval(sys.argv[1:])
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'plot.log')):
os.remove(os.path.join(args.recog_dir, 'plot.log'))
set_logger(os.path.join(args.recog_dir, 'plot.log'),
stdout=args.recog_stdout)
# Load the LM
model = build_lm(args, dir_name)
load_checkpoint(args.recog_model[0], model)
# NOTE: model averaging is not helpful for LM
logger.info('batch size: %d' % args.recog_batch_size)
logger.info('BPTT: %d' % (args.bptt))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache theta: %.3f' % (args.recog_cache_theta))
logger.info('cache lambda: %.3f' % (args.recog_cache_lambda))
model.cache_theta = args.recog_cache_theta
model.cache_lambda = args.recog_cache_lambda
# GPU setting
if args.recog_n_gpus > 0:
model.cuda()
for s in args.recog_sets:
# Load dataset
dataset = Dataset(corpus=args.corpus,
tsv_path=s,
batch_size=args.recog_batch_size,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize,
is_test=True)
assert args.recog_n_caches > 0
save_path = mkdir_join(args.recog_dir, 'cache')
# Clean directory
if save_path is not None and os.path.isdir(save_path):
shutil.rmtree(save_path)
os.mkdir(save_path)
hidden = None
fig_count = 0
token_count = 0
n_tokens = args.recog_n_caches
while True:
ys, is_new_epoch = dataset.next()
for t in range(ys.shape[1] - 1):
loss, hidden = model(ys[:, t:t + 2],
hidden,
is_eval=True,
n_caches=args.recog_n_caches)[:2]
if len(model.cache_attn) > 0:
if token_count == n_tokens:
tokens_keys = dataset.idx2token[0](
model.cache_ids[:args.recog_n_caches],
return_list=True)
tokens_query = dataset.idx2token[0](
model.cache_ids[-n_tokens:], return_list=True)
# Slide attention matrix
n_keys = len(tokens_keys)
n_queries = len(tokens_query)
cache_probs = np.zeros(
(n_keys, n_queries)) # `[n_keys, n_queries]`
mask = np.zeros((n_keys, n_queries))
for i, aw in enumerate(model.cache_attn[-n_tokens:]):
cache_probs[:(n_keys - n_queries + i + 1),
i] = aw[0,
-(n_keys - n_queries + i + 1):]
mask[(n_keys - n_queries + i + 1):, i] = 1
plot_cache_weights(cache_probs,
keys=tokens_keys,
queries=tokens_query,
save_path=mkdir_join(
save_path,
str(fig_count) + '.png'),
figsize=(40, 16),
mask=mask)
token_count = 0
fig_count += 1
else:
token_count += 1
if is_new_epoch:
break
