def main():
args = parse()
# Load a conf file
dir_name = os.path.dirname(args.recog_model[0])
conf = load_config(os.path.join(dir_name, 'conf.yml'))
# Overwrite conf
for k, v in conf.items():
if 'recog' not in k:
setattr(args, k, v)
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'decode.log')):
os.remove(os.path.join(args.recog_dir, 'decode.log'))
logger = set_logger(os.path.join(args.recog_dir, 'decode.log'),
key='decoding')
ppl_avg = 0
for i, s in enumerate(args.recog_sets):
# Load dataset
dataset = Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=os.path.join(dir_name, 'dict.txt'),
wp_model=os.path.join(dir_name, 'wp.model'),
unit=args.unit,
batch_size=args.recog_batch_size,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize,
is_test=True)
if i == 0:
# Load the LM
model = select_lm(args)
model, checkpoint = load_checkpoint(model, args.recog_model[0])
epoch = checkpoint['epoch']
model.save_path = dir_name
logger.info('epoch: %d' % (epoch - 1))
logger.info('batch size: %d' % args.recog_batch_size)
# logger.info('recog unit: %s' % args.recog_unit)
# logger.info('ensemble: %d' % (len(ensemble_models)))
logger.info('BPTT: %d' % (args.bptt))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache theta: %.3f' % (args.recog_cache_theta))
logger.info('cache lambda: %.3f' % (args.recog_cache_lambda))
model.cache_theta = args.recog_cache_theta
model.cache_lambda = args.recog_cache_lambda
# GPU setting
model.cuda()
start_time = time.time()
# TODO(hirofumi): ensemble
ppl, _ = eval_ppl([model],
dataset,
batch_size=1,
bptt=args.bptt,
n_caches=args.recog_n_caches,
progressbar=True)
ppl_avg += ppl
print('PPL (%s): %.2f' % (dataset.set, ppl))
logger.info('Elasped time: %.2f [sec]:' % (time.time() - start_time))
logger.info('PPL (avg.): %.2f\n' % (ppl_avg / len(args.recog_sets)))
def main():
args = parse()
# Load a conf file
dir_name = os.path.dirname(args.recog_model[0])
conf = load_config(os.path.join(dir_name, 'conf.yml'))
# Overwrite conf
for k, v in conf.items():
if 'recog' not in k:
setattr(args, k, v)
recog_params = vars(args)
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'decode.log')):
os.remove(os.path.join(args.recog_dir, 'decode.log'))
logger = set_logger(os.path.join(args.recog_dir, 'decode.log'),
key='decoding')
skip_thought = 'skip' in args.enc_type
wer_avg, cer_avg, per_avg = 0, 0, 0
ppl_avg, loss_avg = 0, 0
for i, s in enumerate(args.recog_sets):
# Load dataset
dataset = Dataset(
corpus=args.corpus,
tsv_path=s,
dict_path=os.path.join(dir_name, 'dict.txt'),
dict_path_sub1=os.path.join(dir_name, 'dict_sub1.txt') if
os.path.isfile(os.path.join(dir_name, 'dict_sub1.txt')) else False,
dict_path_sub2=os.path.join(dir_name, 'dict_sub2.txt') if
os.path.isfile(os.path.join(dir_name, 'dict_sub2.txt')) else False,
nlsyms=os.path.join(dir_name, 'nlsyms.txt'),
wp_model=os.path.join(dir_name, 'wp.model'),
wp_model_sub1=os.path.join(dir_name, 'wp_sub1.model'),
wp_model_sub2=os.path.join(dir_name, 'wp_sub2.model'),
unit=args.unit,
unit_sub1=args.unit_sub1,
unit_sub2=args.unit_sub2,
batch_size=args.recog_batch_size,
skip_thought=skip_thought,
is_test=True)
if i == 0:
# Load the ASR model
if skip_thought:
model = SkipThought(args, dir_name)
else:
model = Speech2Text(args, dir_name)
model, checkpoint = load_checkpoint(model, args.recog_model[0])
epoch = checkpoint['epoch']
# ensemble (different models)
ensemble_models = [model]
if len(args.recog_model) > 1:
for recog_model_e in args.recog_model[1:]:
conf_e = load_config(
os.path.join(os.path.dirname(recog_model_e),
'conf.yml'))
args_e = copy.deepcopy(args)
for k, v in conf_e.items():
if 'recog' not in k:
setattr(args_e, k, v)
model_e = Speech2Text(args_e)
model_e, _ = load_checkpoint(model_e, recog_model_e)
model_e.cuda()
ensemble_models += [model_e]
# Load the LM for shallow fusion
if not args.lm_fusion:
if args.recog_lm is not None and args.recog_lm_weight > 0:
conf_lm = load_config(
os.path.join(os.path.dirname(args.recog_lm),
'conf.yml'))
args_lm = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm, k, v)
lm = select_lm(args_lm)
lm, _ = load_checkpoint(lm, args.recog_lm)
if args_lm.backward:
model.lm_bwd = lm
else:
model.lm_fwd = lm
if args.recog_lm_bwd is not None and args.recog_lm_weight > 0 \
and (args.recog_fwd_bwd_attention or args.recog_reverse_lm_rescoring):
conf_lm = load_config(
os.path.join(os.path.dirname(args.recog_lm_bwd),
'conf.yml'))
args_lm_bwd = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm_bwd, k, v)
lm_bwd = select_lm(args_lm_bwd)
lm_bwd, _ = load_checkpoint(lm_bwd, args.recog_lm_bwd)
model.lm_bwd = lm_bwd
if not args.recog_unit:
args.recog_unit = args.unit
logger.info('recog unit: %s' % args.recog_unit)
logger.info('recog metric: %s' % args.recog_metric)
logger.info('recog oracle: %s' % args.recog_oracle)
logger.info('epoch: %d' % (epoch - 1))
logger.info('batch size: %d' % args.recog_batch_size)
logger.info('beam width: %d' % args.recog_beam_width)
logger.info('min length ratio: %.3f' % args.recog_min_len_ratio)
logger.info('max length ratio: %.3f' % args.recog_max_len_ratio)
logger.info('length penalty: %.3f' % args.recog_length_penalty)
logger.info('coverage penalty: %.3f' % args.recog_coverage_penalty)
logger.info('coverage threshold: %.3f' %
args.recog_coverage_threshold)
logger.info('CTC weight: %.3f' % args.recog_ctc_weight)
logger.info('LM path: %s' % args.recog_lm)
logger.info('LM path (bwd): %s' % args.recog_lm_bwd)
logger.info('LM weight: %.3f' % args.recog_lm_weight)
logger.info('GNMT: %s' % args.recog_gnmt_decoding)
logger.info('forward-backward attention: %s' %
args.recog_fwd_bwd_attention)
logger.info('reverse LM rescoring: %s' %
args.recog_reverse_lm_rescoring)
logger.info('resolving UNK: %s' % args.recog_resolving_unk)
logger.info('ensemble: %d' % (len(ensemble_models)))
logger.info('ASR decoder state carry over: %s' %
(args.recog_asr_state_carry_over))
logger.info('LM state carry over: %s' %
(args.recog_lm_state_carry_over))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache type: %s' % (args.recog_cache_type))
logger.info('cache word frequency threshold: %s' %
(args.recog_cache_word_freq))
logger.info('cache theta (speech): %.3f' %
(args.recog_cache_theta_speech))
logger.info('cache lambda (speech): %.3f' %
(args.recog_cache_lambda_speech))
logger.info('cache theta (lm): %.3f' % (args.recog_cache_theta_lm))
logger.info('cache lambda (lm): %.3f' %
(args.recog_cache_lambda_lm))
# GPU setting
model.cuda()
start_time = time.time()
if args.recog_metric == 'edit_distance':
if args.recog_unit in ['word', 'word_char']:
wer, cer, _ = eval_word(ensemble_models,
dataset,
recog_params,
epoch=epoch - 1,
recog_dir=args.recog_dir,
progressbar=True)
wer_avg += wer
cer_avg += cer
elif args.recog_unit == 'wp':
wer, cer = eval_wordpiece(ensemble_models,
dataset,
recog_params,
epoch=epoch - 1,
recog_dir=args.recog_dir,
progressbar=True)
wer_avg += wer
cer_avg += cer
elif 'char' in args.recog_unit:
wer, cer = eval_char(ensemble_models,
dataset,
recog_params,
epoch=epoch - 1,
recog_dir=args.recog_dir,
progressbar=True,
task_idx=0)
# task_idx=1 if args.recog_unit and 'char' in args.recog_unit else 0)
wer_avg += wer
cer_avg += cer
elif 'phone' in args.recog_unit:
per = eval_phone(ensemble_models,
dataset,
recog_params,
epoch=epoch - 1,
recog_dir=args.recog_dir,
progressbar=True)
per_avg += per
else:
raise ValueError(args.recog_unit)
elif args.recog_metric == 'acc':
raise NotImplementedError
elif args.recog_metric in ['ppl', 'loss']:
ppl, loss = eval_ppl(ensemble_models,
dataset,
recog_params=recog_params,
progressbar=True)
ppl_avg += ppl
loss_avg += loss
elif args.recog_metric == 'bleu':
raise NotImplementedError
else:
raise NotImplementedError
logger.info('Elasped time: %.2f [sec]:' % (time.time() - start_time))
if args.recog_metric == 'edit_distance':
if 'phone' in args.recog_unit:
logger.info('PER (avg.): %.2f %%\n' %
(per_avg / len(args.recog_sets)))
else:
logger.info('WER / CER (avg.): %.2f / %.2f %%\n' %
(wer_avg / len(args.recog_sets),
cer_avg / len(args.recog_sets)))
elif args.recog_metric in ['ppl', 'loss']:
logger.info('PPL (avg.): %.2f\n' % (ppl_avg / len(args.recog_sets)))
print('PPL (avg.): %.2f' % (ppl_avg / len(args.recog_sets)))
logger.info('Loss (avg.): %.2f\n' % (loss_avg / len(args.recog_sets)))
print('Loss (avg.): %.2f' % (loss_avg / len(args.recog_sets)))
def main():
args = parse()
# Load a conf file
dir_name = os.path.dirname(args.recog_model[0])
conf = load_config(os.path.join(dir_name, 'conf.yml'))
# Overwrite conf
for k, v in conf.items():
if 'recog' not in k:
setattr(args, k, v)
recog_params = vars(args)
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'plot.log')):
os.remove(os.path.join(args.recog_dir, 'plot.log'))
logger = set_logger(os.path.join(args.recog_dir, 'plot.log'),
key='decoding')
for i, s in enumerate(args.recog_sets):
# Load dataset
dataset = Dataset(
corpus=args.corpus,
tsv_path=s,
dict_path=os.path.join(dir_name, 'dict.txt'),
dict_path_sub1=os.path.join(dir_name, 'dict_sub1.txt') if
os.path.isfile(os.path.join(dir_name, 'dict_sub1.txt')) else False,
nlsyms=args.nlsyms,
wp_model=os.path.join(dir_name, 'wp.model'),
unit=args.unit,
unit_sub1=args.unit_sub1,
batch_size=args.recog_batch_size,
is_test=True)
if i == 0:
# Load the ASR model
model = Speech2Text(args, dir_name)
model, checkpoint = load_checkpoint(model, args.recog_model[0])
epoch = checkpoint['epoch']
# ensemble (different models)
ensemble_models = [model]
if len(args.recog_model) > 1:
for recog_model_e in args.recog_model[1:]:
# Load the ASR model
conf_e = load_config(
os.path.join(os.path.dirname(recog_model_e),
'conf.yml'))
args_e = copy.deepcopy(args)
for k, v in conf_e.items():
if 'recog' not in k:
setattr(args_e, k, v)
model_e = Speech2Text(args_e)
model_e, _ = load_checkpoint(model_e, recog_model_e)
model_e.cuda()
ensemble_models += [model_e]
# Load the LM for shallow fusion
if not args.lm_fusion:
if args.recog_lm is not None and args.recog_lm_weight > 0:
conf_lm = load_config(
os.path.join(os.path.dirname(args.recog_lm),
'conf.yml'))
args_lm = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm, k, v)
lm = select_lm(args_lm)
lm, _ = load_checkpoint(lm, args.recog_lm)
if args_lm.backward:
model.lm_bwd = lm
else:
model.lm_fwd = lm
if args.recog_lm_bwd is not None and args.recog_lm_weight > 0 and \
(args.recog_fwd_bwd_attention or args.recog_reverse_lm_rescoring):
conf_lm = load_config(
os.path.join(args.recog_lm_bwd, 'conf.yml'))
args_lm_bwd = argparse.Namespace()
for k, v in conf_lm.items():
setattr(args_lm_bwd, k, v)
lm_bwd = select_lm(args_lm_bwd)
lm_bwd, _ = load_checkpoint(lm_bwd, args.recog_lm_bwd)
model.lm_bwd = lm_bwd
if not args.recog_unit:
args.recog_unit = args.unit
logger.info('recog unit: %s' % args.recog_unit)
logger.info('recog metric: %s' % args.recog_metric)
logger.info('recog oracle: %s' % args.recog_oracle)
logger.info('epoch: %d' % (epoch - 1))
logger.info('batch size: %d' % args.recog_batch_size)
logger.info('beam width: %d' % args.recog_beam_width)
logger.info('min length ratio: %.3f' % args.recog_min_len_ratio)
logger.info('max length ratio: %.3f' % args.recog_max_len_ratio)
logger.info('length penalty: %.3f' % args.recog_length_penalty)
logger.info('coverage penalty: %.3f' % args.recog_coverage_penalty)
logger.info('coverage threshold: %.3f' %
args.recog_coverage_threshold)
logger.info('CTC weight: %.3f' % args.recog_ctc_weight)
logger.info('LM path: %s' % args.recog_lm)
logger.info('LM path (bwd): %s' % args.recog_lm_bwd)
logger.info('LM weight: %.3f' % args.recog_lm_weight)
logger.info('GNMT: %s' % args.recog_gnmt_decoding)
logger.info('forward-backward attention: %s' %
args.recog_fwd_bwd_attention)
logger.info('reverse LM rescoring: %s' %
args.recog_reverse_lm_rescoring)
logger.info('resolving UNK: %s' % args.recog_resolving_unk)
logger.info('ensemble: %d' % (len(ensemble_models)))
logger.info('ASR decoder state carry over: %s' %
(args.recog_asr_state_carry_over))
logger.info('LM state carry over: %s' %
(args.recog_lm_state_carry_over))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache type: %s' % (args.recog_cache_type))
logger.info('cache word frequency threshold: %s' %
(args.recog_cache_word_freq))
logger.info('cache theta (speech): %.3f' %
(args.recog_cache_theta_speech))
logger.info('cache lambda (speech): %.3f' %
(args.recog_cache_lambda_speech))
logger.info('cache theta (lm): %.3f' % (args.recog_cache_theta_lm))
logger.info('cache lambda (lm): %.3f' %
(args.recog_cache_lambda_lm))
# GPU setting
model.cuda()
# TODO(hirofumi): move this
save_path = mkdir_join(args.recog_dir, 'att_weights')
if args.recog_n_caches > 0:
save_path_cache = mkdir_join(args.recog_dir, 'cache')
# Clean directory
if save_path is not None and os.path.isdir(save_path):
shutil.rmtree(save_path)
os.mkdir(save_path)
if args.recog_n_caches > 0:
shutil.rmtree(save_path_cache)
os.mkdir(save_path_cache)
while True:
batch, is_new_epoch = dataset.next(
recog_params['recog_batch_size'])
best_hyps_id, aws, (cache_attn_hist, cache_id_hist) = model.decode(
batch['xs'],
recog_params,
dataset.idx2token[0],
exclude_eos=False,
refs_id=batch['ys'],
ensemble_models=ensemble_models[1:]
if len(ensemble_models) > 1 else [],
speakers=batch['sessions']
if dataset.corpus == 'swbd' else batch['speakers'])
if model.bwd_weight > 0.5:
# Reverse the order
best_hyps_id = [hyp[::-1] for hyp in best_hyps_id]
aws = [aw[::-1] for aw in aws]
for b in range(len(batch['xs'])):
tokens = dataset.idx2token[0](best_hyps_id[b],
return_list=True)
spk = batch['speakers'][b]
plot_attention_weights(
aws[b][:len(tokens)],
tokens,
spectrogram=batch['xs'][b][:, :dataset.input_dim]
if args.input_type == 'speech' else None,
save_path=mkdir_join(save_path, spk,
batch['utt_ids'][b] + '.png'),
figsize=(20, 8))
if args.recog_n_caches > 0 and cache_id_hist is not None and cache_attn_hist is not None:
n_keys, n_queries = cache_attn_hist[0].shape
# mask = np.ones((n_keys, n_queries))
# for i in range(n_queries):
# mask[:n_keys - i, -(i + 1)] = 0
mask = np.zeros((n_keys, n_queries))
plot_cache_weights(
cache_attn_hist[0],
keys=dataset.idx2token[0](cache_id_hist[-1],
return_list=True), # fifo
# keys=dataset.idx2token[0](cache_id_hist, return_list=True), # dict
queries=tokens,
save_path=mkdir_join(save_path_cache, spk,
batch['utt_ids'][b] + '.png'),
figsize=(40, 16),
mask=mask)
if model.bwd_weight > 0.5:
hyp = ' '.join(tokens[::-1])
else:
hyp = ' '.join(tokens)
logger.info('utt-id: %s' % batch['utt_ids'][b])
logger.info('Ref: %s' % batch['text'][b].lower())
logger.info('Hyp: %s' % hyp)
logger.info('-' * 50)
if is_new_epoch:
break
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',
stdout=args.stdout)
# Model setting
model = select_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'])
# 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 = 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)
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:
model = CustomDataParallel(model,
device_ids=list(range(0, args.n_gpus, 1)),
deterministic=False,
benchmark=True)
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, tensorboard=True)
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 /= 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()
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()
reporter.epoch()
# 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): %.2f' % (dev_set.set, ppl_dev))
optimizer.epoch(ppl_dev)
reporter.epoch(ppl_dev, name='perplexity')
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): %.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)))
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 = 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)
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))
if reporter.tensorboard:
reporter.tf_writer.close()
pbar_epoch.close()
return save_path
def main():
args = parse()
# Load a conf file
if args.resume:
conf = load_config(
os.path.join(os.path.dirname(args.resume), 'conf.yml'))
for k, v in conf.items():
if k != 'resume':
setattr(args, k, v)
# Load dataset
train_set = Dataset(corpus=args.corpus,
tsv_path=args.train_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size * args.n_gpus,
n_epochs=args.n_epochs,
min_n_tokens=args.min_n_tokens,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize)
dev_set = Dataset(corpus=args.corpus,
tsv_path=args.dev_set,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=args.batch_size * args.n_gpus,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize)
eval_sets = []
for s in args.eval_sets:
eval_sets += [
Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=args.dict,
nlsyms=args.nlsyms,
unit=args.unit,
wp_model=args.wp_model,
batch_size=1,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize)
]
args.vocab = train_set.vocab
# Set save path
if args.resume:
save_path = os.path.dirname(args.resume)
dir_name = os.path.basename(save_path)
else:
dir_name = make_model_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])
model.set_optimizer(
optimizer='sgd'
if epoch > conf['convert_to_sgd_epoch'] + 1 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)
lr_controller = checkpoint['lr_controller']
epoch = checkpoint['epoch']
step = checkpoint['step']
ppl_dev_best = checkpoint['metric_dev_best']
# Resume between convert_to_sgd_epoch and convert_to_sgd_epoch + 1
if epoch == conf['convert_to_sgd_epoch'] + 1:
model.set_optimizer(optimizer='sgd',
lr=args.learning_rate,
weight_decay=float(conf['weight_decay']))
logger.info('========== Convert to SGD ==========')
else:
# Save the conf file as a yaml file
save_config(vars(args), os.path.join(model.save_path, 'conf.yml'))
# Save the nlsyms, dictionar, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(model.save_path,
'nlsyms.txt'))
shutil.copy(args.dict, os.path.join(model.save_path, 'dict.txt'))
if args.unit == 'wp':
shutil.copy(args.wp_model, os.path.join(model.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)
# Set optimizer
model.set_optimizer(optimizer=args.optimizer,
lr=float(args.learning_rate),
weight_decay=float(args.weight_decay),
transformer=args.lm_type == 'transformer')
epoch, step = 1, 1
ppl_dev_best = 10000
# Set learning rate controller
lr_controller = Controller(
lr=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,
transformer=args.lm_type == 'transformer')
train_set.epoch = epoch - 1 # start from index:0
# 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(model.module.save_path, tensorboard=True)
hidden = None
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
not_improved_epoch = 0
pbar_epoch = tqdm(total=len(train_set))
while True:
# Compute loss in the training set
ys_train, is_new_epoch = train_set.next()
model.module.optimizer.zero_grad()
loss, hidden, reporter = model(ys_train, hidden, reporter)
if len(model.device_ids) > 1:
loss.backward(torch.ones(len(model.device_ids)))
else:
loss.backward()
loss.detach() # Trancate the graph
if args.clip_grad_norm > 0:
torch.nn.utils.clip_grad_norm_(model.module.parameters(),
args.clip_grad_norm)
model.module.optimizer.step()
loss_train = loss.item()
del loss
if 'gated_conv' not in args.lm_type and args.lm_type != 'transformer':
hidden = model.module.repackage_hidden(hidden)
reporter.step(is_eval=False)
# Update learning rate
if step < args.warmup_n_steps or args.lm_type == 'transformer':
model.module.optimizer = lr_controller.warmup(
model.module.optimizer, step=step)
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, train_set.epoch_detail, loss_train, loss_dev,
np.exp(loss_train), np.exp(loss_dev), lr_controller.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:
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.module,
model.module.save_path,
lr_controller,
epoch,
step - 1,
ppl_dev_best,
remove_old_checkpoints=True)
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
model.module.optimizer = lr_controller.decay(
model.module.optimizer, epoch=epoch, value=ppl_dev)
if ppl_dev < ppl_dev_best:
ppl_dev_best = ppl_dev
not_improved_epoch = 0
logger.info('||||| Best Score |||||')
# Save the model
save_checkpoint(model.module,
model.module.save_path,
lr_controller,
epoch,
step - 1,
ppl_dev_best,
remove_old_checkpoints=True)
# 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_epoch += 1
duration_eval = time.time() - start_time_eval
logger.info('Evaluation time: %.2f min' % (duration_eval / 60))
# Early stopping
if not_improved_epoch == args.not_improved_patient_n_epochs:
break
# Convert to fine-tuning stage
if epoch == args.convert_to_sgd_epoch:
model.module.set_optimizer('sgd',
lr=args.learning_rate,
weight_decay=float(
args.weight_decay))
lr_controller = Controller(lr=args.learning_rate,
decay_type='epoch',
decay_start_epoch=epoch,
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()
epoch += 1
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 model.module.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
transformer = 'transformer' in args.enc_type or args.dec_type == 'transformer'
# 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 = make_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])
model.set_optimizer(
optimizer='sgd'
if epoch > conf['convert_to_sgd_epoch'] + 1 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)
lr_controller = checkpoint['lr_controller']
epoch = checkpoint['epoch']
step = checkpoint['step']
metric_dev_best = checkpoint['metric_dev_best']
# Resume between convert_to_sgd_epoch and convert_to_sgd_epoch + 1
if epoch == conf['convert_to_sgd_epoch'] + 1:
model.set_optimizer(optimizer='sgd',
lr=args.learning_rate,
weight_decay=float(conf['weight_decay']))
logger.info('========== Convert to SGD ==========')
else:
# Save the conf file as a yaml file
save_config(vars(args), os.path.join(model.save_path, 'conf.yml'))
if args.lm_fusion:
save_config(args.lm_conf,
os.path.join(model.save_path, 'conf_lm.yml'))
# Save the nlsyms, dictionar, and wp_model
if args.nlsyms:
shutil.copy(args.nlsyms, os.path.join(model.save_path,
'nlsyms.txt'))
for sub in ['', '_sub1', '_sub2']:
if getattr(args, 'dict' + sub):
shutil.copy(
getattr(args, 'dict' + sub),
os.path.join(model.save_path, 'dict' + sub + '.txt'))
if getattr(args, 'unit' + sub) == 'wp':
shutil.copy(
getattr(args, 'wp_model' + sub),
os.path.join(model.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)
# Set optimizer
model.set_optimizer(optimizer=args.optimizer,
lr=float(args.learning_rate),
weight_decay=float(args.weight_decay),
transformer=transformer)
epoch, step = 1, 1
metric_dev_best = 10000
# Set learning rate controller
lr_controller = Controller(
lr=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,
transformer=transformer)
train_set.epoch = epoch - 1 # start from index:0
# 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(model.module.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 tasks 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)
model.module.optimizer.step()
model.module.optimizer.zero_grad()
accum_n_tokens = 0
loss_train = loss.item()
del loss
reporter.step(is_eval=False)
# Update learning rate
if step < args.warmup_n_steps or transformer:
model.module.optimizer = lr_controller.warmup(
model.module.optimizer, step=step)
if step % args.print_step == 0:
# Compute loss in the dev set
batch_dev = dev_set.next()[0]
# Change tasks 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, train_set.epoch_detail, loss_train, loss_dev,
lr_controller.lr, len(batch_train['utt_ids']), xlen, ylen,
duration_step / 60))
start_time_step = time.time()
step += args.n_gpus
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:
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.module,
model.module.save_path,
lr_controller,
epoch,
step - 1,
metric_dev_best,
remove_old_checkpoints=True)
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
model.module.optimizer = lr_controller.decay(
model.module.optimizer, 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.module,
model.module.save_path,
lr_controller,
epoch,
step - 1,
metric_dev_best,
remove_old_checkpoints=True)
# 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:
model.module.set_optimizer('sgd',
lr=args.learning_rate,
weight_decay=float(
args.weight_decay))
lr_controller = Controller(lr=args.learning_rate,
decay_type='epoch',
decay_start_epoch=epoch,
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()
epoch += 1
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 model.module.save_path
def main():
args = parse()
# Load a conf file
dir_name = os.path.dirname(args.recog_model[0])
conf = load_config(os.path.join(dir_name, 'conf.yml'))
# Overwrite conf
for k, v in conf.items():
if 'recog' not in k:
setattr(args, k, v)
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'plot.log')):
os.remove(os.path.join(args.recog_dir, 'plot.log'))
logger = set_logger(os.path.join(args.recog_dir, 'plot.log'),
key='decoding')
for i, s in enumerate(args.recog_sets):
# Load dataset
dataset = Dataset(corpus=args.corpus,
tsv_path=s,
dict_path=os.path.join(dir_name, 'dict.txt'),
wp_model=os.path.join(dir_name, 'wp.model'),
unit=args.unit,
batch_size=args.recog_batch_size,
bptt=args.bptt,
backward=args.backward,
serialize=args.serialize,
is_test=True)
if i == 0:
# Load the LM
model = select_lm(args, dir_name)
model, checkpoint = load_checkpoint(model, args.recog_model[0])
epoch = checkpoint['epoch']
logger.info('epoch: %d' % (epoch - 1))
logger.info('batch size: %d' % args.recog_batch_size)
# logger.info('recog unit: %s' % args.recog_unit)
# logger.info('ensemble: %d' % (len(ensemble_models)))
logger.info('BPTT: %d' % (args.bptt))
logger.info('cache size: %d' % (args.recog_n_caches))
logger.info('cache theta: %.3f' % (args.recog_cache_theta))
logger.info('cache lambda: %.3f' % (args.recog_cache_lambda))
model.cache_theta = args.recog_cache_theta
model.cache_lambda = args.recog_cache_lambda
# GPU setting
model.cuda()
assert args.recog_n_caches > 0
save_path = mkdir_join(args.recog_dir, 'cache')
# Clean directory
if save_path is not None and os.path.isdir(save_path):
shutil.rmtree(save_path)
os.mkdir(save_path)
hidden = None
fig_count = 0
toknen_count = 0
n_tokens = args.recog_n_caches
while True:
ys, is_new_epoch = dataset.next()
for t in range(ys.shape[1] - 1):
loss, hidden = model(ys[:, t:t + 2],
hidden,
is_eval=True,
n_caches=args.recog_n_caches)[:2]
if len(model.cache_attn) > 0:
if toknen_count == n_tokens:
tokens_keys = dataset.idx2token[0](
model.cache_ids[:args.recog_n_caches],
return_list=True)
tokens_query = dataset.idx2token[0](
model.cache_ids[-n_tokens:], return_list=True)
# Slide attention matrix
n_keys = len(tokens_keys)
n_queries = len(tokens_query)
cache_probs = np.zeros(
(n_keys, n_queries)) # `[n_keys, n_queries]`
mask = np.zeros((n_keys, n_queries))
for i, aw in enumerate(model.cache_attn[-n_tokens:]):
cache_probs[:(n_keys - n_queries + i + 1),
i] = aw[0,
-(n_keys - n_queries + i + 1):]
mask[(n_keys - n_queries + i + 1):, i] = 1
plot_cache_weights(cache_probs,
keys=tokens_keys,
queries=tokens_query,
save_path=mkdir_join(
save_path,
str(fig_count) + '.png'),
figsize=(40, 16),
mask=mask)
toknen_count = 0
fig_count += 1
else:
toknen_count += 1
if is_new_epoch:
break