def main():
# Load configuration
args, recog_params, dir_name = parse_args_eval(sys.argv[1:])
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'plot.log')):
os.remove(os.path.join(args.recog_dir, 'plot.log'))
set_logger(os.path.join(args.recog_dir, 'plot.log'),
stdout=args.recog_stdout)
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)
epoch = int(args.recog_model[0].split('-')[-1])
if args.recog_n_average > 1:
# Model averaging for Transformer
model = average_checkpoints(model,
args.recog_model[0],
n_average=args.recog_n_average)
else:
load_checkpoint(args.recog_model[0], model)
if not args.recog_unit:
args.recog_unit = args.unit
logger.info('recog unit: %s' % args.recog_unit)
logger.info('epoch: %d' % epoch)
logger.info('batch size: %d' % args.recog_batch_size)
# GPU setting
if args.recog_n_gpus >= 1:
model.cudnn_setting(deterministic=True, benchmark=False)
model.cuda()
save_path = mkdir_join(args.recog_dir, 'ctc_probs')
# Clean directory
if save_path is not None and os.path.isdir(save_path):
shutil.rmtree(save_path)
os.mkdir(save_path)
while True:
batch, is_new_epoch = dataset.next(
recog_params['recog_batch_size'])
best_hyps_id, _ = model.decode(batch['xs'], recog_params)
# Get CTC probs
ctc_probs, topk_ids, xlens = model.get_ctc_probs(batch['xs'],
temperature=1,
topk=min(
100,
model.vocab))
# NOTE: ctc_probs: '[B, T, topk]'
for b in range(len(batch['xs'])):
tokens = dataset.idx2token[0](best_hyps_id[b],
return_list=True)
spk = batch['speakers'][b]
plot_ctc_probs(
ctc_probs[b, :xlens[b]],
topk_ids[b],
subsample_factor=args.subsample_factor,
spectrogram=batch['xs'][b][:, :dataset.input_dim],
save_path=mkdir_join(save_path, spk,
batch['utt_ids'][b] + '.png'),
figsize=(20, 8))
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():
# Load configuration
args, dir_name = parse_args_eval(sys.argv[1:])
# Setting for logging
if os.path.isfile(os.path.join(args.recog_dir, 'plot.log')):
os.remove(os.path.join(args.recog_dir, 'plot.log'))
set_logger(os.path.join(args.recog_dir, 'plot.log'),
stdout=args.recog_stdout)
for i, s in enumerate(args.recog_sets):
# Load dataloader
dataloader = build_dataloader(
args=args,
tsv_path=s,
batch_size=1,
is_test=True,
first_n_utterances=args.recog_first_n_utt,
longform_max_n_frames=args.recog_longform_max_n_frames)
if i == 0:
# Load ASR model
model = Speech2Text(args, dir_name)
epoch = int(float(args.recog_model[0].split('-')[-1]) * 10) / 10
if args.recog_n_average > 1:
# Model averaging for Transformer
model = average_checkpoints(model,
args.recog_model[0],
n_average=args.recog_n_average)
else:
load_checkpoint(args.recog_model[0], model)
if not args.recog_unit:
args.recog_unit = args.unit
logger.info('recog unit: %s' % args.recog_unit)
logger.info('epoch: %d' % epoch)
logger.info('batch size: %d' % args.recog_batch_size)
# GPU setting
if args.recog_n_gpus >= 1:
model.cudnn_setting(deterministic=True, benchmark=False)
model.cuda()
save_path = mkdir_join(args.recog_dir, 'ctc_probs')
# Clean directory
if save_path is not None and os.path.isdir(save_path):
shutil.rmtree(save_path)
os.mkdir(save_path)
for batch in dataloader:
nbest_hyps_id, _ = model.decode(batch['xs'], args,
dataloader.idx2token[0])
best_hyps_id = [h[0] for h in nbest_hyps_id]
# Get CTC probs
ctc_probs, topk_ids, xlens = model.get_ctc_probs(batch['xs'],
temperature=1,
topk=min(
100,
model.vocab))
# NOTE: ctc_probs: '[B, T, topk]'
for b in range(len(batch['xs'])):
tokens = dataloader.idx2token[0](best_hyps_id[b],
return_list=True)
spk = batch['speakers'][b]
plot_ctc_probs(
ctc_probs[b, :xlens[b]],
topk_ids[b],
factor=args.subsample_factor,
spectrogram=batch['xs'][b][:, :dataloader.input_dim],
save_path=mkdir_join(save_path, spk,
batch['utt_ids'][b] + '.png'),
figsize=(20, 8))
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)
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', stdout=args.recog_stdout)
for i, s in enumerate(args.recog_sets):
subsample_factor = 1
subsample = [int(s) for s in args.subsample.split('_')]
if args.conv_poolings:
for p in args.conv_poolings.split('_'):
p = int(p.split(',')[0].replace('(', ''))
if p > 1:
subsample_factor *= p
subsample_factor *= np.prod(subsample)
# 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 = load_checkpoint(model, args.recog_model[0])[0]
epoch = int(args.recog_model[0].split('-')[-1])
if not args.recog_unit:
args.recog_unit = args.unit
logger.info('recog unit: %s' % args.recog_unit)
logger.info('epoch: %d' % (epoch - 1))
logger.info('batch size: %d' % args.recog_batch_size)
# GPU setting
model.cuda()
save_path = mkdir_join(args.recog_dir, 'ctc_probs')
# Clean directory
if save_path is not None and os.path.isdir(save_path):
shutil.rmtree(save_path)
os.mkdir(save_path)
while True:
batch, is_new_epoch = dataset.next(recog_params['recog_batch_size'])
best_hyps_id, _, _ = model.decode(batch['xs'], recog_params,
exclude_eos=False)
# Get CTC probs
ctc_probs, indices_topk, xlens = model.get_ctc_probs(
batch['xs'], temperature=1, topk=min(100, model.vocab))
# NOTE: ctc_probs: '[B, T, topk]'
for b in range(len(batch['xs'])):
tokens = dataset.idx2token[0](best_hyps_id[b], return_list=True)
spk = batch['speakers'][b]
plot_ctc_probs(
ctc_probs[b, :xlens[b]],
indices_topk[b],
n_frames=xlens[b],
subsample_factor=subsample_factor,
spectrogram=batch['xs'][b][:, :dataset.input_dim],
save_path=mkdir_join(save_path, spk, batch['utt_ids'][b] + '.png'),
figsize=(20, 8))
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
