def convert_to_sgd(self, model, lr, weight_decay, decay_type, decay_rate):
self.decay_type = decay_type
self.decay_rate = decay_rate
weight_decay = self.optimizer.weight_decay
self.optimizer = set_optimizer(model, 'sgd', lr, weight_decay)
logger.info('========== Convert to SGD ==========')
def convert_to_sgd(self, model, lr, weight_decay, decay_type, decay_rate):
self.lr = lr
self.decay_type = decay_type
self.decay_rate = decay_rate
self.noam = False
# weight_decay = self.optimizer.defaults['weight_decay']
self.optimizer = set_optimizer(model, 'sgd', lr, weight_decay)
logger.info('========== Convert to SGD ==========')
def main():
args = parse()
args_init = copy.deepcopy(args)
args_teacher = copy.deepcopy(args)
# 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)
recog_params = vars(args)
# Compute subsampling factor
subsample_factor = 1
subsample_factor_sub1 = 1
subsample_factor_sub2 = 1
subsample = [int(s) for s in args.subsample.split('_')]
if args.conv_poolings and 'conv' in args.enc_type:
for p in args.conv_poolings.split('_'):
subsample_factor *= int(p.split(',')[0].replace('(', ''))
else:
subsample_factor = np.prod(subsample)
if args.train_set_sub1:
if args.conv_poolings and 'conv' in args.enc_type:
subsample_factor_sub1 = subsample_factor * np.prod(
subsample[:args.enc_n_layers_sub1 - 1])
else:
subsample_factor_sub1 = subsample_factor
if args.train_set_sub2:
if args.conv_poolings and 'conv' in args.enc_type:
subsample_factor_sub2 = subsample_factor * np.prod(
subsample[:args.enc_n_layers_sub2 - 1])
else:
subsample_factor_sub2 = subsample_factor
# Set save path
if args.resume:
save_path = os.path.dirname(args.resume)
dir_name = os.path.basename(save_path)
else:
dir_name = set_asr_model_name(args, subsample_factor)
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)
# for multi-GPUs
if args.n_gpus > 1:
logger.info("Batch size is automatically reduced from %d to %d" %
(args.batch_size, args.batch_size // 2))
args.batch_size //= 2
skip_thought = 'skip' in args.enc_type
# Load dataset
train_set = Dataset(corpus=args.corpus,
tsv_path=args.train_set,
tsv_path_sub1=args.train_set_sub1,
tsv_path_sub2=args.train_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=args.batch_size * args.n_gpus,
n_epochs=args.n_epochs,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
sort_by='input',
short2long=True,
sort_stop_epoch=args.sort_stop_epoch,
dynamic_batching=args.dynamic_batching,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=subsample_factor,
subsample_factor_sub1=subsample_factor_sub1,
subsample_factor_sub2=subsample_factor_sub2,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
tsv_path_sub1=args.dev_set_sub1,
tsv_path_sub2=args.dev_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=args.batch_size * args.n_gpus,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=subsample_factor,
subsample_factor_sub1=subsample_factor_sub1,
subsample_factor_sub2=subsample_factor_sub2,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought)
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,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought,
is_test=True)
]
args.vocab = train_set.vocab
args.vocab_sub1 = train_set.vocab_sub1
args.vocab_sub2 = train_set.vocab_sub2
args.input_dim = train_set.input_dim
# Load a LM conf file for LM fusion & LM initialization
if not args.resume and (args.lm_fusion or args.lm_init):
if args.lm_fusion:
lm_conf = load_config(
os.path.join(os.path.dirname(args.lm_fusion), 'conf.yml'))
elif args.lm_init:
lm_conf = load_config(
os.path.join(os.path.dirname(args.lm_init), 'conf.yml'))
args.lm_conf = argparse.Namespace()
for k, v in lm_conf.items():
setattr(args.lm_conf, k, v)
assert args.unit == args.lm_conf.unit
assert args.vocab == args.lm_conf.vocab
# Model setting
model = Speech2Text(args, save_path) if not skip_thought else SkipThought(
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
noam = 'transformer' in conf['enc_type'] or conf[
'dec_type'] == 'transformer'
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=noam)
# 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']:
optimizer.convert_to_sgd(model,
'sgd',
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'))
if args.lm_fusion:
save_config(args.lm_conf, os.path.join(save_path, 'conf_lm.yml'))
# Save the nlsyms, dictionar, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(save_path, 'nlsyms.txt'))
for sub in ['', '_sub1', '_sub2']:
if getattr(args, 'dict' + sub):
shutil.copy(getattr(args, 'dict' + sub),
os.path.join(save_path, 'dict' + sub + '.txt'))
if getattr(args, 'unit' + sub) == 'wp':
shutil.copy(getattr(args, 'wp_model' + sub),
os.path.join(save_path, 'wp' + sub + '.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)
# Initialize with pre-trained model's parameters
if args.asr_init and os.path.isfile(args.asr_init):
# Load the ASR model
conf_init = load_config(
os.path.join(os.path.dirname(args.asr_init), 'conf.yml'))
for k, v in conf_init.items():
setattr(args_init, k, v)
model_init = Speech2Text(args_init)
model_init = load_checkpoint(model_init, args.asr_init)[0]
# Overwrite parameters
only_enc = (args.enc_n_layers != args_init.enc_n_layers) or (
args.unit != args_init.unit) or args_init.ctc_weight == 1
param_dict = dict(model_init.named_parameters())
for n, p in model.named_parameters():
if n in param_dict.keys() and p.size() == param_dict[n].size():
if only_enc and 'enc' not in n:
continue
if args.lm_fusion_type == 'cache' and 'output' in n:
continue
p.data = param_dict[n].data
logger.info('Overwrite %s' % n)
# Set optimizer
optimizer = set_optimizer(model, args.optimizer, args.lr,
args.weight_decay)
# Wrap optimizer by learning rate scheduler
noam = 'transformer' in args.enc_type or args.dec_type == 'transformer'
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=noam)
# Load the teacher ASR model
teacher = None
if args.teacher and os.path.isfile(args.teacher):
conf_teacher = load_config(
os.path.join(os.path.dirname(args.teacher), 'conf.yml'))
for k, v in conf_teacher.items():
setattr(args_teacher, k, v)
# Setting for knowledge distillation
args_teacher.ss_prob = 0
args.lsm_prob = 0
teacher = Speech2Text(args_teacher)
teacher = load_checkpoint(teacher, args.teacher)[0]
# Load the teacher LM
teacher_lm = None
if args.teacher_lm and os.path.isfile(args.teacher_lm):
conf_lm = load_config(
os.path.join(os.path.dirname(args.teacher_lm), 'conf.yml'))
args_lm = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm, k, v)
teacher_lm = build_lm(args_lm)
teacher_lm = load_checkpoint(teacher_lm, args.teacher_lm)[0]
# 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()
if teacher is not None:
teacher.cuda()
if teacher_lm is not None:
teacher_lm.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)
if args.mtl_per_batch:
# NOTE: from easier to harder tasks
tasks = []
if 1 - args.bwd_weight - args.ctc_weight - args.sub1_weight - args.sub2_weight > 0:
tasks += ['ys']
if args.bwd_weight > 0:
tasks = ['ys.bwd'] + tasks
if args.ctc_weight > 0:
tasks = ['ys.ctc'] + tasks
for sub in ['sub1', 'sub2']:
if getattr(args, 'train_set_' + sub):
if getattr(args, sub + '_weight') - getattr(
args, 'ctc_weight_' + sub) > 0:
tasks = ['ys_' + sub] + tasks
if getattr(args, 'ctc_weight_' + sub) > 0:
tasks = ['ys_' + sub + '.ctc'] + tasks
else:
tasks = ['all']
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
batch_train, is_new_epoch = train_set.next()
accum_n_tokens += sum([len(y) for y in batch_train['ys']])
# Change mini-batch depending on task
for task in tasks:
if skip_thought:
loss, reporter = model(batch_train['ys'],
ys_prev=batch_train['ys_prev'],
ys_next=batch_train['ys_next'],
reporter=reporter)
else:
loss, reporter = model(batch_train,
reporter,
task,
teacher=teacher,
teacher_lm=teacher_lm)
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
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
batch_dev = dev_set.next()[0]
# Change mini-batch depending on task
for task in tasks:
if skip_thought:
loss, reporter = model(batch_dev['ys'],
ys_prev=batch_dev['ys_prev'],
ys_next=batch_dev['ys_next'],
reporter=reporter,
is_eval=True)
else:
loss, reporter = model(batch_dev,
reporter,
task,
is_eval=True)
loss_dev = loss.item()
del loss
# NOTE: this makes training slow
# Compute WER/CER regardless of the output unit (greedy decoding)
# best_hyps_id, _, _ = model.module.decode(
# batch_dev['xs'], recog_params, dev_set.idx2token[0], exclude_eos=True)
# cer = 0.
# ref_n_words, ref_n_chars = 0, 0
# for b in range(len(batch_dev['xs'])):
# ref = batch_dev['text'][b]
# hyp = dev_set.idx2token[0](best_hyps_id[b])
# cer += editdistance.eval(hyp, ref)
# ref_n_words += len(ref.split())
# ref_n_chars += len(ref)
# wer = cer / ref_n_words
# cer /= ref_n_chars
# reporter.add_tensorboard_scalar('dev/WER', wer)
# reporter.add_tensorboard_scalar('dev/CER', cer)
# logger.info('WER (dev)', wer)
# logger.info('CER (dev)', cer)
reporter.step(is_eval=True)
duration_step = time.time() - start_time_step
if args.input_type == 'speech':
xlen = max(len(x) for x in batch_train['xs'])
ylen = max(len(y) for y in batch_train['ys'])
elif args.input_type == 'text':
xlen = max(len(x) for x in batch_train['ys'])
ylen = max(len(y) for y in batch_train['ys_sub1'])
logger.info(
"step:%d(ep:%.2f) loss:%.3f(%.3f)/lr:%.7f/bs:%d/xlen:%d/ylen:%d (%.2f min)"
%
(optimizer.n_steps, optimizer.n_epochs +
train_set.epoch_detail, loss_train, loss_dev, optimizer.lr,
len(batch_train['utt_ids']), xlen, ylen, duration_step / 60))
start_time_step = time.time()
pbar_epoch.update(len(batch_train['utt_ids']))
# Save fugures of loss and accuracy
if optimizer.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
save_checkpoint(model,
save_path,
optimizer,
optimizer.n_epochs,
remove_old_checkpoints=not noam)
else:
start_time_eval = time.time()
# dev
metric_dev = eval_epoch([model.module], dev_set, recog_params,
args, optimizer.n_epochs + 1, logger)
optimizer.epoch(metric_dev) # lr decay
reporter.epoch(metric_dev) # plot
if optimizer.is_best:
# Save the model
save_checkpoint(model,
save_path,
optimizer,
optimizer.n_epochs,
remove_old_checkpoints=not noam)
# test
for eval_set in eval_sets:
eval_epoch([model.module], eval_set, recog_params,
args, optimizer.n_epochs, logger)
# start scheduled sampling
if args.ss_prob > 0:
model.module.scheduled_sampling_trigger()
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,
'sgd',
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_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):
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
args_init = copy.deepcopy(args)
args_teacher = copy.deepcopy(args)
# 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 not in ['resume', 'local_rank']:
setattr(args, k, v)
args = compute_subsampling_factor(args)
resume_epoch = int(args.resume.split('-')[-1]) if args.resume else 0
# Load dataset
train_set = build_dataloader(args=args,
tsv_path=args.train_set,
tsv_path_sub1=args.train_set_sub1,
tsv_path_sub2=args.train_set_sub2,
batch_size=args.batch_size,
batch_size_type=args.batch_size_type,
max_n_frames=args.max_n_frames,
resume_epoch=resume_epoch,
sort_by=args.sort_by,
short2long=args.sort_short2long,
sort_stop_epoch=args.sort_stop_epoch,
num_workers=args.workers,
pin_memory=args.pin_memory,
distributed=args.distributed,
word_alignment_dir=args.train_word_alignment,
ctc_alignment_dir=args.train_ctc_alignment)
dev_set = build_dataloader(
args=args,
tsv_path=args.dev_set,
tsv_path_sub1=args.dev_set_sub1,
tsv_path_sub2=args.dev_set_sub2,
batch_size=1 if 'transducer' in args.dec_type else args.batch_size,
batch_size_type='seq'
if 'transducer' in args.dec_type else args.batch_size_type,
max_n_frames=1600,
word_alignment_dir=args.dev_word_alignment,
ctc_alignment_dir=args.dev_ctc_alignment)
eval_sets = [
build_dataloader(args=args, tsv_path=s, batch_size=1, is_test=True)
for s in args.eval_sets
]
args.vocab = train_set.vocab
args.vocab_sub1 = train_set.vocab_sub1
args.vocab_sub2 = train_set.vocab_sub2
args.input_dim = train_set.input_dim
# 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_asr_model_name(args)
if args.mbr_training:
assert args.asr_init
args.save_path = mkdir_join(os.path.dirname(args.asr_init),
dir_name)
else:
args.save_path = mkdir_join(
args.model_save_dir,
'_'.join(os.path.basename(args.train_set).split('.')[:-1]),
dir_name)
if args.local_rank > 0:
time.sleep(1)
args.save_path = set_save_path(args.save_path) # avoid overwriting
# Set logger
set_logger(os.path.join(args.save_path, 'train.log'), args.stdout,
args.local_rank)
# Load a LM conf file for LM fusion & LM initialization
if not args.resume and args.external_lm:
lm_conf = load_config(
os.path.join(os.path.dirname(args.external_lm), 'conf.yml'))
args.lm_conf = argparse.Namespace()
for k, v in lm_conf.items():
setattr(args.lm_conf, k, v)
assert args.unit == args.lm_conf.unit
assert args.vocab == args.lm_conf.vocab
# Model setting
model = Speech2Text(args, args.save_path, train_set.idx2token[0])
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'))
for sub in ['', '_sub1', '_sub2']:
if args.get('dict' + sub):
shutil.copy(
args.get('dict' + sub),
os.path.join(args.save_path, 'dict' + sub + '.txt'))
if args.get('unit' + sub) == 'wp':
shutil.copy(
args.get('wp_model' + sub),
os.path.join(args.save_path, 'wp' + sub + '.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)
# Initialize with pre-trained model's parameters
if args.asr_init:
# Load ASR model (full model)
conf_init = load_config(
os.path.join(os.path.dirname(args.asr_init), 'conf.yml'))
for k, v in conf_init.items():
setattr(args_init, k, v)
model_init = Speech2Text(args_init)
load_checkpoint(args.asr_init, model_init)
# Overwrite parameters
param_dict = dict(model_init.named_parameters())
for n, p in model.named_parameters():
if n in param_dict.keys() and p.size() == param_dict[n].size():
if args.asr_init_enc_only and 'enc' not in n:
continue
p.data = param_dict[n].data
logger.info('Overwrite %s' % n)
# Set optimizer
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 = 'former' in args.enc_type or 'former' in args.dec_type or 'former' in args.dec_type_sub1
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,
lower_better=args.metric not in ['accuracy', 'bleu'],
warmup_start_lr=args.warmup_start_lr,
warmup_n_steps=args.warmup_n_steps,
peak_lr=0.05 / (args.get('transformer_enc_d_model', 0)**0.5)
if 'conformer' in args.enc_type else 1e6,
model_size=args.get('transformer_enc_d_model',
args.get('transformer_dec_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)
# Load teacher ASR model
teacher = None
if args.teacher:
assert os.path.isfile(args.teacher), 'There is no checkpoint.'
conf_teacher = load_config(
os.path.join(os.path.dirname(args.teacher), 'conf.yml'))
for k, v in conf_teacher.items():
setattr(args_teacher, k, v)
# Setting for knowledge distillation
args_teacher.ss_prob = 0
args.lsm_prob = 0
teacher = Speech2Text(args_teacher)
load_checkpoint(args.teacher, teacher)
# Load teacher LM
teacher_lm = None
if args.teacher_lm:
assert os.path.isfile(args.teacher_lm), 'There is no checkpoint.'
conf_lm = load_config(
os.path.join(os.path.dirname(args.teacher_lm), 'conf.yml'))
args_lm = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm, k, v)
teacher_lm = build_lm(args_lm)
load_checkpoint(args.teacher_lm, teacher_lm)
# 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) and (not args.cudnn_benchmark))
or args.cudnn_deterministic,
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)
from neural_sp.models.seq2seq.decoders.ctc import CTC
amp.register_float_function(CTC, "loss_fn")
# NOTE: see https://github.com/espnet/espnet/pull/1779
amp.init()
if args.resume:
load_checkpoint(args.resume, amp=amp)
n = torch.cuda.device_count() // args.local_world_size
device_ids = list(range(args.local_rank * n,
(args.local_rank + 1) * n))
torch.cuda.set_device(device_ids[0])
model.cuda(device_ids[0])
scheduler.cuda(device_ids[0])
if args.distributed:
model = DDP(model, device_ids=device_ids)
else:
model = CustomDataParallel(model,
device_ids=list(range(args.n_gpus)))
if teacher is not None:
teacher.cuda()
if teacher_lm is not None:
teacher_lm.cuda()
else:
model = CPUWrapperASR(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, args.local_rank)
args.wandb_id = reporter.wandb_id
if args.resume:
n_steps = scheduler.n_steps * max(
1, args.accum_grad_n_steps // args.local_world_size)
reporter.resume(n_steps, resume_epoch)
# Save conf file as a yaml file
if args.local_rank == 0:
save_config(args, os.path.join(args.save_path, 'conf.yml'))
if args.external_lm:
save_config(args.lm_conf,
os.path.join(args.save_path, 'conf_lm.yml'))
# NOTE: save after reporter for wandb ID
# Define tasks
if args.mtl_per_batch:
# NOTE: from easier to harder tasks
tasks = []
if args.total_weight - args.bwd_weight - args.ctc_weight - args.sub1_weight - args.sub2_weight > 0:
tasks += ['ys']
if args.bwd_weight > 0:
tasks = ['ys.bwd'] + tasks
if args.ctc_weight > 0:
tasks = ['ys.ctc'] + tasks
if args.mbr_ce_weight > 0:
tasks = ['ys.mbr'] + tasks
for sub in ['sub1', 'sub2']:
if args.get('train_set_' + sub) is not None:
if args.get(sub + '_weight', 0) - args.get(
'ctc_weight_' + sub, 0) > 0:
tasks = ['ys_' + sub] + tasks
if args.get('ctc_weight_' + sub, 0) > 0:
tasks = ['ys_' + sub + '.ctc'] + tasks
else:
tasks = ['all']
if args.get('ss_start_epoch', 0) <= resume_epoch:
model.module.trigger_scheduled_sampling()
if args.get('mocha_quantity_loss_start_epoch', 0) <= resume_epoch:
model.module.trigger_quantity_loss()
start_time_train = time.time()
for ep in range(resume_epoch, args.n_epochs):
train_one_epoch(model, train_set, dev_set, eval_sets, scheduler,
reporter, logger, args, amp, scaler, tasks, teacher,
teacher_lm)
# 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
if args.local_rank == 0:
scheduler.save_checkpoint(model,
args.save_path,
amp=amp,
remove_old=(not is_transformer)
and args.remove_old_checkpoints)
else:
start_time_eval = time.time()
# dev
metric_dev = validate([model.module], dev_set, args,
reporter.n_epochs + 1, logger)
scheduler.epoch(metric_dev) # lr decay
reporter.epoch(metric_dev, name=args.metric) # plot
reporter.add_scalar('dev/' + args.metric, metric_dev)
if scheduler.is_topk or is_transformer:
# Save model
if args.local_rank == 0:
scheduler.save_checkpoint(model,
args.save_path,
amp=amp,
remove_old=(not is_transformer)
and args.remove_old_checkpoints)
# test
if scheduler.is_topk:
for eval_set in eval_sets:
validate([model.module], eval_set, args,
reporter.n_epochs, logger)
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
if args.get('ss_start_epoch', 0) == (ep + 1):
model.module.trigger_scheduled_sampling()
if args.get('mocha_quantity_loss_start_epoch', 0) == (ep + 1):
model.module.trigger_quantity_loss()
logger.info('Total time: %.2f hour' %
((time.time() - start_time_train) / 3600))
reporter.close()
return args.save_path
def main():
args = parse_args_train(sys.argv[1:])
args_init = copy.deepcopy(args)
args_teacher = copy.deepcopy(args)
# 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)
recog_params = vars(args)
args = compute_susampling_factor(args)
# 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,
tsv_path_sub1=args.train_set_sub1,
tsv_path_sub2=args.train_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=batch_size,
n_epochs=args.n_epochs,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
shuffle_bucket=args.shuffle_bucket,
sort_by='input',
short2long=args.sort_short2long,
sort_stop_epoch=args.sort_stop_epoch,
dynamic_batching=args.dynamic_batching,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=args.subsample_factor,
subsample_factor_sub1=args.subsample_factor_sub1,
subsample_factor_sub2=args.subsample_factor_sub2,
discourse_aware=args.discourse_aware)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
tsv_path_sub1=args.dev_set_sub1,
tsv_path_sub2=args.dev_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=batch_size,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=args.subsample_factor,
subsample_factor_sub1=args.subsample_factor_sub1,
subsample_factor_sub2=args.subsample_factor_sub2)
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,
is_test=True) for s in args.eval_sets]
args.vocab = train_set.vocab
args.vocab_sub1 = train_set.vocab_sub1
args.vocab_sub2 = train_set.vocab_sub2
args.input_dim = train_set.input_dim
# Set save path
if args.resume:
save_path = os.path.dirname(args.resume)
dir_name = os.path.basename(save_path)
else:
dir_name = set_asr_model_name(args)
if args.mbr_training:
assert args.asr_init
save_path = mkdir_join(os.path.dirname(args.asr_init), dir_name)
else:
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)
# Load a LM conf file for LM fusion & LM initialization
if not args.resume and args.external_lm:
lm_conf = load_config(os.path.join(os.path.dirname(args.external_lm), 'conf.yml'))
args.lm_conf = argparse.Namespace()
for k, v in lm_conf.items():
setattr(args.lm_conf, k, v)
assert args.unit == args.lm_conf.unit
assert args.vocab == args.lm_conf.vocab
# Model setting
model = Speech2Text(args, save_path, train_set.idx2token[0])
if not args.resume:
# Save the conf file as a yaml file
save_config(vars(args), os.path.join(save_path, 'conf.yml'))
if args.external_lm:
save_config(args.lm_conf, os.path.join(save_path, 'conf_lm.yml'))
# Save the nlsyms, dictionary, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(save_path, 'nlsyms.txt'))
for sub in ['', '_sub1', '_sub2']:
if getattr(args, 'dict' + sub):
shutil.copy(getattr(args, 'dict' + sub), os.path.join(save_path, 'dict' + sub + '.txt'))
if getattr(args, 'unit' + sub) == 'wp':
shutil.copy(getattr(args, 'wp_model' + sub), os.path.join(save_path, 'wp' + sub + '.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)
# Initialize with pre-trained model's parameters
if args.asr_init:
# Load the ASR model (full model)
conf_init = load_config(os.path.join(os.path.dirname(args.asr_init), 'conf.yml'))
for k, v in conf_init.items():
setattr(args_init, k, v)
model_init = Speech2Text(args_init)
load_checkpoint(args.asr_init, model_init)
# Overwrite parameters
param_dict = dict(model_init.named_parameters())
for n, p in model.named_parameters():
if n in param_dict.keys() and p.size() == param_dict[n].size():
if args.asr_init_enc_only and 'enc' not in n:
continue
p.data = param_dict[n].data
logger.info('Overwrite %s' % n)
# Set optimizer
resume_epoch = 0
if args.resume:
resume_epoch = int(args.resume.split('-')[-1])
optimizer = set_optimizer(model, 'sgd' if resume_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 = 'former' in args.enc_type or 'former' in args.dec_type
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,
lower_better=args.metric not in ['accuracy', 'bleu'],
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=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, 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)
# Load the teacher ASR model
teacher = None
if args.teacher:
assert os.path.isfile(args.teacher), 'There is no checkpoint.'
conf_teacher = load_config(os.path.join(os.path.dirname(args.teacher), 'conf.yml'))
for k, v in conf_teacher.items():
setattr(args_teacher, k, v)
# Setting for knowledge distillation
args_teacher.ss_prob = 0
args.lsm_prob = 0
teacher = Speech2Text(args_teacher)
load_checkpoint(args.teacher, teacher)
# Load the teacher LM
teacher_lm = None
if args.teacher_lm:
assert os.path.isfile(args.teacher_lm), 'There is no checkpoint.'
conf_lm = load_config(os.path.join(os.path.dirname(args.teacher_lm), 'conf.yml'))
args_lm = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm, k, v)
teacher_lm = build_lm(args_lm)
load_checkpoint(args.teacher_lm, teacher_lm)
# 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=not is_transformer and 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)
from neural_sp.models.seq2seq.decoders.ctc import CTC
amp.register_float_function(CTC, "loss_fn")
# NOTE: see https://github.com/espnet/espnet/pull/1779
amp.init()
if args.resume:
load_checkpoint(args.resume, amp=amp)
model = CustomDataParallel(model, device_ids=list(range(0, args.n_gpus)))
if teacher is not None:
teacher.cuda()
if teacher_lm is not None:
teacher_lm.cuda()
else:
model = CPUWrapperASR(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)
if args.mtl_per_batch:
# NOTE: from easier to harder tasks
tasks = []
if 1 - args.bwd_weight - args.ctc_weight - args.sub1_weight - args.sub2_weight > 0:
tasks += ['ys']
if args.bwd_weight > 0:
tasks = ['ys.bwd'] + tasks
if args.ctc_weight > 0:
tasks = ['ys.ctc'] + tasks
if args.mbr_ce_weight > 0:
tasks = ['ys.mbr'] + tasks
for sub in ['sub1', 'sub2']:
if getattr(args, 'train_set_' + sub):
if getattr(args, sub + '_weight') - getattr(args, 'ctc_weight_' + sub) > 0:
tasks = ['ys_' + sub] + tasks
if getattr(args, 'ctc_weight_' + sub) > 0:
tasks = ['ys_' + sub + '.ctc'] + tasks
else:
tasks = ['all']
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
accum_n_steps = 0
n_steps = optimizer.n_steps * args.accum_grad_n_steps
epoch_detail_prev = 0
for ep in range(resume_epoch, args.n_epochs):
pbar_epoch = tqdm(total=len(train_set))
session_prev = None
for batch_train, is_new_epoch in train_set:
# Compute loss in the training set
if args.discourse_aware and batch_train['sessions'][0] != session_prev:
model.module.reset_session()
session_prev = batch_train['sessions'][0]
accum_n_steps += 1
# Change mini-batch depending on task
if accum_n_steps == 1:
loss_train = 0 # moving average over gradient accumulation
for task in tasks:
loss, observation = model(batch_train, task,
teacher=teacher, teacher_lm=teacher_lm)
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
loss_train = (loss_train * (accum_n_steps - 1) + loss.item()) / accum_n_steps
if accum_n_steps >= args.accum_grad_n_steps or is_new_epoch:
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
# NOTE: parameters are forcibly updated at the end of every epoch
del loss
pbar_epoch.update(len(batch_train['utt_ids']))
reporter.add_tensorboard_scalar('learning_rate', optimizer.lr)
# NOTE: loss/acc/ppl are already added in the model
reporter.step()
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
batch_dev = iter(dev_set).next(batch_size=1 if 'transducer' in args.dec_type else None)[0]
# Change mini-batch depending on task
for task in tasks:
loss, observation = model(batch_dev, task, 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
if args.input_type == 'speech':
xlen = max(len(x) for x in batch_train['xs'])
ylen = max(len(y) for y in batch_train['ys'])
elif args.input_type == 'text':
xlen = max(len(x) for x in batch_train['ys'])
ylen = max(len(y) for y in batch_train['ys_sub1'])
logger.info("step:%d(ep:%.2f) loss:%.3f(%.3f)/lr:%.7f/bs:%d/xlen:%d/ylen:%d (%.2f min)" %
(n_steps, optimizer.n_epochs + train_set.epoch_detail,
loss_train, loss_dev,
optimizer.lr, len(batch_train['utt_ids']),
xlen, ylen, 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()
model.module.plot_ctc()
# Ealuate model every 0.1 epoch during MBR training
if args.mbr_training:
if int(train_set.epoch_detail * 10) != int(epoch_detail_prev * 10):
# dev
evaluate([model.module], dev_set, recog_params, args,
int(train_set.epoch_detail * 10) / 10, logger)
# Save the model
optimizer.save_checkpoint(
model, save_path, remove_old=False, amp=amp,
epoch_detail=train_set.epoch_detail)
epoch_detail_prev = train_set.epoch_detail
if is_new_epoch:
break
# Save checkpoint and evaluate model per 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 and args.remove_old_checkpoints, amp=amp)
else:
start_time_eval = time.time()
# dev
metric_dev = evaluate([model.module], dev_set, recog_params, args,
optimizer.n_epochs + 1, logger)
optimizer.epoch(metric_dev) # lr decay
reporter.epoch(metric_dev, name=args.metric) # plot
if optimizer.is_topk or is_transformer:
# Save the model
optimizer.save_checkpoint(
model, save_path, remove_old=not is_transformer and args.remove_old_checkpoints, amp=amp)
# test
if optimizer.is_topk:
for eval_set in eval_sets:
evaluate([model.module], eval_set, recog_params, args,
optimizer.n_epochs, logger)
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)
if optimizer.n_epochs >= args.n_epochs:
break
# if args.ss_prob > 0:
# model.module.scheduled_sampling_trigger()
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)
# 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():
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)
recog_params = vars(args)
# Compute subsampling factor
subsample_factor = 1
subsample = [int(s) for s in args.subsample.split('_')]
if args.conv_poolings and 'conv' in args.enc_type:
for p in args.conv_poolings.split('_'):
subsample_factor *= int(p.split(',')[0].replace('(', ''))
else:
subsample_factor = np.prod(subsample)
skip_thought = 'skip' in args.enc_type
batch_per_allreduce = args.batch_size
# Load dataset
train_set = Dataset(corpus=args.corpus,
tsv_path=args.train_set,
tsv_path_sub1=args.train_set_sub1,
tsv_path_sub2=args.train_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=args.batch_size,
n_epochs=args.n_epochs,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
sort_by='no_sort',
short2long=True,
sort_stop_epoch=args.sort_stop_epoch,
dynamic_batching=args.dynamic_batching,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=subsample_factor,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
tsv_path_sub1=args.dev_set_sub1,
tsv_path_sub2=args.dev_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=args.batch_size,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=subsample_factor,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought)
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,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought,
is_test=True)
]
args.vocab = train_set.vocab
args.vocab_sub1 = train_set.vocab_sub1
args.vocab_sub2 = train_set.vocab_sub2
args.input_dim = train_set.input_dim
# Horovod: use DistributedSampler to partition data among workers. Manually specify
# `num_replicas=hvd.size()` and `rank=hvd.rank()`.
train_loader = SeqDataloader(train_set,
batch_size=args.batch_size,
num_workers=1,
distributed=True,
num_stacks=args.n_stacks,
num_splices=args.n_splices,
num_skips=args.n_skips,
pin_memory=False,
shuffle=False)
val_loader = SeqDataloader(dev_set,
batch_size=args.batch_size,
num_workers=1,
distributed=True,
num_stacks=args.n_stacks,
num_splices=args.n_splices,
num_skips=args.n_skips,
pin_memory=False,
shuffle=False)
# Load a LM conf file for LM fusion & LM initialization
if not args.resume and (args.lm_fusion or args.lm_init):
if args.lm_fusion:
lm_conf = load_config(
os.path.join(os.path.dirname(args.lm_fusion), 'conf.yml'))
elif args.lm_init:
lm_conf = load_config(
os.path.join(os.path.dirname(args.lm_init), 'conf.yml'))
args.lm_conf = argparse.Namespace()
for k, v in lm_conf.items():
setattr(args.lm_conf, k, v)
assert args.unit == args.lm_conf.unit
assert args.vocab == args.lm_conf.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_asr_model_name(args, subsample_factor)
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 = Speech2Text(args, save_path)
# GPU setting
if args.n_gpus >= 1:
torch.backends.cudnn.benchmark = True
model.cuda()
if args.resume:
# Set optimizer
epochs = int(args.resume.split('-')[-1])
#optimizer = set_optimizer(model, 'sgd' if epochs >= conf['convert_to_sgd_epoch'] else conf['optimizer'],
model, _ = load_checkpoint(model, args.resume, resume=True)
optimizer = set_optimizer(model, 'sgd', conf['lr'],
conf['weight_decay'])
#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
noam = 'transformer' in args.enc_type or args.dec_type == 'transformer'
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=noam)
else:
# Save the conf file as a yaml file
if hvd_rank == 0:
save_config(vars(args), os.path.join(save_path, 'conf.yml'))
if args.lm_fusion:
save_config(args.lm_conf, os.path.join(save_path, 'conf_lm.yml'))
if hvd_rank == 0:
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)
optimizer = hvd.DistributedOptimizer(
optimizer,
named_parameters=model.named_parameters(),
compression=hvd.Compression.none,
backward_passes_per_step=batch_per_allreduce)
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
# Wrap optimizer by learning rate scheduler
noam = 'transformer' in args.enc_type or args.dec_type == 'transformer'
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=noam)
# Set reporter
reporter = Reporter(save_path)
if args.mtl_per_batch:
# NOTE: from easier to harder tasks
tasks = []
if 1 - args.bwd_weight - args.ctc_weight - args.sub1_weight - args.sub2_weight > 0:
tasks += ['ys']
if args.bwd_weight > 0:
tasks = ['ys.bwd'] + tasks
if args.ctc_weight > 0:
tasks = ['ys.ctc'] + tasks
else:
tasks = ['all']
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
accum_n_tokens = 0
verbose = 1 if hvd_rank == 0 else 0
data_size = len(train_set)
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 _, batch_train in enumerate(train_loader):
accum_n_tokens += sum([len(y) for y in batch_train['ys']])
# Change mini-batch depending on task
for task in tasks:
if skip_thought:
loss, reporter = model(batch_train['ys'],
ys_prev=batch_train['ys_prev'],
ys_next=batch_train['ys_next'],
reporter=reporter)
else:
loss, reporter = model(batch_train, reporter, task)
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)
optimizer.step()
optimizer.zero_grad()
accum_n_tokens = 0
loss_train = loss.item()
del loss
if optimizer.n_steps % args.print_step == 0:
# Compute loss in the dev set
model.eval()
batch_dev = dev_set.next()[0]
# Change mini-batch depending on task
for task in tasks:
if skip_thought:
loss, reporter = model(
batch_dev['ys'],
ys_prev=batch_dev['ys_prev'],
ys_next=batch_dev['ys_next'],
reporter=reporter,
is_eval=True)
else:
loss, reporter = model(batch_dev,
reporter,
task,
is_eval=True)
loss_dev = loss.item()
del loss
duration_step = time.time() - start_time_step
if args.input_type == 'speech':
xlen = max(len(x) for x in batch_train['xs'])
ylen = max(len(y) for y in batch_train['ys'])
elif args.input_type == 'text':
xlen = max(len(x) for x in batch_train['ys'])
ylen = max(len(y) for y in batch_train['ys_sub1'])
if hvd_rank == 0:
logger.info(
"step:%d(ep:%.2f) loss:%.3f(%.3f)/lr:%.5f/bs:%d/xlen:%d/ylen:%d (%.2f min)"
% (optimizer.n_steps,
optimizer.n_steps * args.batch_size /
(data_size / hvd.size()), loss_train, loss_dev,
optimizer.lr, len(batch_train['utt_ids']), xlen,
ylen, duration_step / 60))
start_time_step = time.time()
pbar_epoch.update(len(batch_train['utt_ids']))
# Save fugures of loss and accuracy
if optimizer.n_steps % (args.print_step *
10) == 0 and hvd.rank() == 0:
model.plot_attention()
start_time_step = time.time()
# reset dev set
dev_set.reset()
# 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:
optimizer.epoch()
if hvd_rank == 0:
save_checkpoint(model,
save_path,
optimizer,
optimizer.n_epochs,
remove_old_checkpoints=not noam)
else:
start_time_eval = time.time()
# dev
metric_dev = eval_epoch([model], val_loader, recog_params,
args, optimizer.n_epochs + 1)
metric_dev = hvd.allreduce(
np2tensor(np.array([metric_dev], dtype=float),
hvd.local_rank()))
loss_dev = metric_dev.item()
if hvd_rank == 0:
logger.info('Loss : %.2f %%' % (loss_dev))
optimizer.epoch(loss_dev)
if hvd.rank() == 0:
save_checkpoint(model,
save_path,
optimizer,
optimizer.n_epochs,
remove_old_checkpoints=False)
if not optimizer.is_best:
model, _ = load_checkpoint(
model, save_path + '/model.epoch-' +
str(optimizer.best_epochs))
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(),
compression=hvd.Compression.none,
backward_passes_per_step=batch_per_allreduce)
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(optimizer, root_rank=0)
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=noam)
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))
return save_path
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,
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
def main():
args = parse()
args_pt = copy.deepcopy(args)
args_teacher = copy.deepcopy(args)
# 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)
recog_params = vars(args)
# Automatically reduce batch size in multi-GPU setting
if args.n_gpus > 1:
args.batch_size -= 10
args.print_step //= args.n_gpus
# Compute subsampling factor
subsample_factor = 1
subsample_factor_sub1 = 1
subsample_factor_sub2 = 1
subsample = [int(s) for s in args.subsample.split('_')]
if args.conv_poolings and 'conv' in args.enc_type:
for p in args.conv_poolings.split('_'):
subsample_factor *= int(p.split(',')[0].replace('(', ''))
else:
subsample_factor = np.prod(subsample)
if args.train_set_sub1:
if args.conv_poolings and 'conv' in args.enc_type:
subsample_factor_sub1 = subsample_factor * np.prod(subsample[:args.enc_n_layers_sub1 - 1])
else:
subsample_factor_sub1 = subsample_factor
if args.train_set_sub2:
if args.conv_poolings and 'conv' in args.enc_type:
subsample_factor_sub2 = subsample_factor * np.prod(subsample[:args.enc_n_layers_sub2 - 1])
else:
subsample_factor_sub2 = subsample_factor
skip_thought = 'skip' in args.enc_type
# Load dataset
train_set = Dataset(corpus=args.corpus,
tsv_path=args.train_set,
tsv_path_sub1=args.train_set_sub1,
tsv_path_sub2=args.train_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=args.batch_size * args.n_gpus,
n_epochs=args.n_epochs,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
sort_by_input_length=True,
short2long=True,
sort_stop_epoch=args.sort_stop_epoch,
dynamic_batching=args.dynamic_batching,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=subsample_factor,
subsample_factor_sub1=subsample_factor_sub1,
subsample_factor_sub2=subsample_factor_sub2,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
tsv_path_sub1=args.dev_set_sub1,
tsv_path_sub2=args.dev_set_sub2,
dict_path=args.dict,
dict_path_sub1=args.dict_sub1,
dict_path_sub2=args.dict_sub2,
nlsyms=args.nlsyms,
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
wp_model=args.wp_model,
wp_model_sub1=args.wp_model_sub1,
wp_model_sub2=args.wp_model_sub2,
batch_size=args.batch_size * args.n_gpus,
min_n_frames=args.min_n_frames,
max_n_frames=args.max_n_frames,
shuffle=True if args.discourse_aware else False,
ctc=args.ctc_weight > 0,
ctc_sub1=args.ctc_weight_sub1 > 0,
ctc_sub2=args.ctc_weight_sub2 > 0,
subsample_factor=subsample_factor,
subsample_factor_sub1=subsample_factor_sub1,
subsample_factor_sub2=subsample_factor_sub2,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought)
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,
discourse_aware=args.discourse_aware,
skip_thought=skip_thought,
is_test=True)]
args.vocab = train_set.vocab
args.vocab_sub1 = train_set.vocab_sub1
args.vocab_sub2 = train_set.vocab_sub2
args.input_dim = train_set.input_dim
# Load a LM conf file for LM fusion & LM initialization
if not args.resume and (args.lm_fusion or args.lm_init):
if args.lm_fusion:
lm_conf = load_config(os.path.join(os.path.dirname(args.lm_fusion), 'conf.yml'))
elif args.lm_init:
lm_conf = load_config(os.path.join(os.path.dirname(args.lm_init), 'conf.yml'))
args.lm_conf = argparse.Namespace()
for k, v in lm_conf.items():
setattr(args.lm_conf, k, v)
assert args.unit == args.lm_conf.unit
assert args.vocab == args.lm_conf.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_asr_model_name(args, subsample_factor)
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 = SkipThought(args, save_path) if skip_thought else Speech2Text(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']
metric_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'))
if args.lm_fusion:
save_config(args.lm_conf, os.path.join(save_path, 'conf_lm.yml'))
# Save the nlsyms, dictionar, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(save_path, 'nlsyms.txt'))
for sub in ['', '_sub1', '_sub2']:
if getattr(args, 'dict' + sub):
shutil.copy(getattr(args, 'dict' + sub), os.path.join(save_path, 'dict' + sub + '.txt'))
if getattr(args, 'unit' + sub) == 'wp':
shutil.copy(getattr(args, 'wp_model' + sub), os.path.join(save_path, 'wp' + sub + '.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)
# Initialize with pre-trained model's parameters
if args.pretrained_model and os.path.isfile(args.pretrained_model):
# Load the ASR model
conf_pt = load_config(os.path.join(os.path.dirname(args.pretrained_model), 'conf.yml'))
for k, v in conf_pt.items():
setattr(args_pt, k, v)
model_pt = Speech2Text(args_pt)
model_pt, _ = load_checkpoint(model_pt, args.pretrained_model)
# Overwrite parameters
only_enc = (args.enc_n_layers != args_pt.enc_n_layers) or (
args.unit != args_pt.unit) or args_pt.ctc_weight == 1
param_dict = dict(model_pt.named_parameters())
for n, p in model.named_parameters():
if n in param_dict.keys() and p.size() == param_dict[n].size():
if only_enc and 'enc' not in n:
continue
if args.lm_fusion_type == 'cache' and 'output' in n:
continue
p.data = param_dict[n].data
logger.info('Overwrite %s' % n)
epoch, step = 0, 0
metric_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
noam = 'transformer' in args.enc_type or args.dec_type == 'transformer'
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=metric_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=noam)
# Load the teacher ASR model
teacher = None
teacher_lm = None
if args.teacher and os.path.isfile(args.teacher):
conf_teacher = load_config(os.path.join(os.path.dirname(args.teacher), 'conf.yml'))
for k, v in conf_teacher.items():
setattr(args_teacher, k, v)
# Setting for knowledge distillation
args_teacher.ss_prob = 0
args.lsm_prob = 0
teacher = Speech2Text(args_teacher)
teacher, _ = load_checkpoint(teacher, args.teacher)
# Load the teacher LM
if args.teacher_lm and os.path.isfile(args.teacher_lm):
conf_lm = load_config(os.path.join(os.path.dirname(args.teacher_lm), 'conf.yml'))
args_lm = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm, k, v)
teacher_lm = select_lm(args_lm)
teacher_lm, _ = load_checkpoint(teacher_lm, args.teacher_lm)
# 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()
if teacher is not None:
teacher.cuda()
if teacher_lm is not None:
teacher_lm.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)
if args.mtl_per_batch:
# NOTE: from easier to harder tasks
tasks = []
if 1 - args.bwd_weight - args.ctc_weight - args.sub1_weight - args.sub2_weight > 0:
tasks += ['ys']
if args.bwd_weight > 0:
tasks = ['ys.bwd'] + tasks
if args.ctc_weight > 0:
tasks = ['ys.ctc'] + tasks
if args.lmobj_weight > 0:
tasks = ['ys.lmobj'] + tasks
for sub in ['sub1', 'sub2']:
if getattr(args, 'train_set_' + sub):
if getattr(args, sub + '_weight') - getattr(args, 'ctc_weight_' + sub) > 0:
tasks = ['ys_' + sub] + tasks
if getattr(args, 'ctc_weight_' + sub) > 0:
tasks = ['ys_' + sub + '.ctc'] + tasks
else:
tasks = ['all']
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
batch_train, is_new_epoch = train_set.next()
accum_n_tokens += sum([len(y) for y in batch_train['ys']])
# Change mini-batch depending on task
for task in tasks:
if skip_thought:
loss, reporter = model(batch_train['ys'],
ys_prev=batch_train['ys_prev'],
ys_next=batch_train['ys_next'],
reporter=reporter)
else:
loss, reporter = model(batch_train, reporter=reporter, task=task,
teacher=teacher, teacher_lm=teacher_lm)
# 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
reporter.step()
step += args.n_gpus
if step % args.print_step == 0:
# Compute loss in the dev set
batch_dev = dev_set.next()[0]
# Change mini-batch depending on task
for task in tasks:
if skip_thought:
loss, reporter = model(batch_dev['ys'],
ys_prev=batch_dev['ys_prev'],
ys_next=batch_dev['ys_next'],
reporter=reporter,
is_eval=True)
else:
loss, reporter = model(batch_dev, reporter=reporter, task=task,
is_eval=True)
loss_dev = loss.item()
del loss
reporter.step(is_eval=True)
duration_step = time.time() - start_time_step
if args.input_type == 'speech':
xlen = max(len(x) for x in batch_train['xs'])
ylen = max(len(y) for y in batch_train['ys'])
elif args.input_type == 'text':
xlen = max(len(x) for x in batch_train['ys'])
ylen = max(len(y) for y in batch_train['ys_sub1'])
logger.info("step:%d(ep:%.2f) loss:%.3f(%.3f)/lr:%.5f/bs:%d/xlen:%d/ylen:%d (%.2f min)" %
(step, epoch + train_set.epoch_detail,
loss_train, loss_dev,
optimizer.lr, len(batch_train['utt_ids']),
xlen, ylen, duration_step / 60))
start_time_step = time.time()
pbar_epoch.update(len(batch_train['utt_ids']))
# Save fugures of loss and accuracy
if step % (args.print_step * 10) == 0:
reporter.snapshot()
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, metric_dev_best,
remove_old_checkpoints=not noam)
reporter._epoch += 1
# TODO(hirofumi): fix later
else:
start_time_eval = time.time()
# dev
if args.metric == 'edit_distance':
if args.unit in ['word', 'word_char']:
metric_dev = eval_word([model.module], dev_set, recog_params,
epoch=epoch)[0]
logger.info('WER (%s): %.2f %%' % (dev_set.set, metric_dev))
elif args.unit == 'wp':
metric_dev, cer_dev = eval_wordpiece([model.module], dev_set, recog_params,
epoch=epoch)
logger.info('WER (%s): %.2f %%' % (dev_set.set, metric_dev))
logger.info('CER (%s): %.2f %%' % (dev_set.set, cer_dev))
elif 'char' in args.unit:
metric_dev, cer_dev = eval_char([model.module], dev_set, recog_params,
epoch=epoch)
logger.info('WER (%s): %.2f %%' % (dev_set.set, metric_dev))
logger.info('CER (%s): %.2f %%' % (dev_set.set, cer_dev))
elif 'phone' in args.unit:
metric_dev = eval_phone([model.module], dev_set, recog_params,
epoch=epoch)
logger.info('PER (%s): %.2f %%' % (dev_set.set, metric_dev))
elif args.metric == 'ppl':
metric_dev = eval_ppl([model.module], dev_set, batch_size=args.batch_size)[0]
logger.info('PPL (%s): %.2f' % (dev_set.set, metric_dev))
elif args.metric == 'loss':
metric_dev = eval_ppl([model.module], dev_set, batch_size=args.batch_size)[1]
logger.info('Loss (%s): %.2f' % (dev_set.set, metric_dev))
else:
raise NotImplementedError(args.metric)
reporter.epoch(metric_dev)
# Update learning rate
optimizer.decay(epoch=epoch, value=metric_dev)
if metric_dev < metric_dev_best:
metric_dev_best = metric_dev
not_improved_n_epochs = 0
logger.info('||||| Best Score |||||')
# Save the model
save_checkpoint(model, save_path, optimizer,
epoch, step, metric_dev_best,
remove_old_checkpoints=not noam)
# test
for s in eval_sets:
if args.metric == 'edit_distance':
if args.unit in ['word', 'word_char']:
wer_test = eval_word([model.module], s, recog_params,
epoch=epoch)[0]
logger.info('WER (%s): %.2f %%' % (s.set, wer_test))
elif args.unit == 'wp':
wer_test, cer_test = eval_wordpiece([model.module], s, recog_params,
epoch=epoch)
logger.info('WER (%s): %.2f %%' % (s.set, wer_test))
logger.info('CER (%s): %.2f %%' % (s.set, cer_test))
elif 'char' in args.unit:
wer_test, cer_test = eval_char([model.module], s, recog_params,
epoch=epoch)
logger.info('WER (%s): %.2f %%' % (s.set, wer_test))
logger.info('CER (%s): %.2f %%' % (s.set, cer_test))
elif 'phone' in args.unit:
per_test = eval_phone([model.module], s, recog_params,
epoch=epoch)
logger.info('PER (%s): %.2f %%' % (s.set, per_test))
elif args.metric == 'ppl':
ppl_test = eval_ppl([model.module], s, batch_size=args.batch_size)[0]
logger.info('PPL (%s): %.2f' % (s.set, ppl_test))
elif args.metric == 'loss':
loss_test = eval_ppl([model.module], s, batch_size=args.batch_size)[1]
logger.info('Loss (%s): %.2f' % (s.set, loss_test))
else:
raise NotImplementedError(args.metric)
else:
not_improved_n_epochs += 1
# start scheduled sampling
if args.ss_prob > 0:
model.module.scheduled_sampling_trigger()
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