def main():
args = parser.parse_args()
# Load a config file (.yml)
params = load_config(join(args.model_path, 'config.yml'), is_eval=True)
# Setting for logging
logger = set_logger(args.model_path)
for i, data_type in enumerate(['dev', 'test']):
# Load dataset
dataset = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_freq=params['input_freq'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type=data_type,
label_type=params['label_type'],
batch_size=args.eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False,
tool=params['tool'])
if i == 0:
params['num_classes'] = dataset.num_classes
# Load model
model = load(model_type=params['model_type'],
params=params,
backend=params['backend'])
# Restore the saved parameters
epoch, _, _, _ = model.load_checkpoint(save_path=args.model_path,
epoch=args.epoch)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
logger.info('beam width: %d' % args.beam_width)
logger.info('epoch: %d' % (epoch - 1))
per, df = eval_phone(model=model,
dataset=dataset,
map_file_path='./conf/phones.60-48-39.map',
eval_batch_size=args.eval_batch_size,
beam_width=args.beam_width,
max_decode_len=MAX_DECODE_LEN_PHONE,
min_decode_len=MIN_DECODE_LEN_PHONE,
length_penalty=args.length_penalty,
coverage_penalty=args.coverage_penalty,
progressbar=True)
logger.info(' PER (%s): %.3f %%' % (data_type, (per * 100)))
logger.info(df)
def main():
args = parser.parse_args()
# Load a config file (.yml)
params = load_config(join(args.model_path, 'config.yml'), is_eval=True)
logger = set_logger(args.model_path)
for i, data_type in enumerate(
['dev_clean', 'dev_other', 'test_clean', 'test_other']):
# Load dataset
dataset = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_freq=params['input_freq'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type=data_type,
data_size=params['data_size'],
label_type=params['label_type'],
batch_size=args.eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False,
tool=params['tool'])
if i == 0:
params['num_classes'] = dataset.num_classes
# Load model
model = load(model_type=params['model_type'],
params=params,
backend=params['backend'])
# Restore the saved parameters
epoch, _, _, _ = model.load_checkpoint(save_path=args.model_path,
epoch=args.epoch)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
logger.info('beam width: %d' % args.beam_width)
logger.info('epoch: %d' % (epoch - 1))
if params['label_type'] == 'word':
wer, df = eval_word(models=[model],
dataset=dataset,
eval_batch_size=args.eval_batch_size,
beam_width=args.beam_width,
max_decode_len=MAX_DECODE_LEN_WORD,
min_decode_len=MIN_DECODE_LEN_WORD,
length_penalty=args.length_penalty,
coverage_penalty=args.coverage_penalty,
progressbar=True)
logger.info(' WER (%s): %.3f %%' % (dataset.label_type,
(wer * 100)))
logger.info(df)
else:
wer, cer, df = eval_char(models=[model],
dataset=dataset,
eval_batch_size=args.eval_batch_size,
beam_width=args.beam_width,
max_decode_len=MAX_DECODE_LEN_CHAR,
min_decode_len=MIN_DECODE_LEN_CHAR,
length_penalty=args.length_penalty,
coverage_penalty=args.coverage_penalty,
progressbar=True)
logger.info(' WER / CER (%s): %.3f / %.3f %%' %
(dataset.label_type, (wer * 100), (cer * 100)))
logger.info(df)
def main():
args = parser.parse_args()
# Load a config file (.yml)
params = load_config(join(args.model_path, 'config.yml'), is_eval=True)
# Setting for logging
logger = set_logger(args.model_path)
wer_mean, wer_sub_mean, cer_sub_mean = 0, 0, 0
for i, data_type in enumerate(['eval1', 'eval2', 'eval3']):
# Load dataset
dataset = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_freq=params['input_freq'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type=data_type,
data_size=params['data_size'],
label_type=params['label_type'],
label_type_sub=params['label_type_sub'],
batch_size=args.eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
shuffle=False,
tool=params['tool'])
if i == 0:
params['num_classes'] = dataset.num_classes
params['num_classes_sub'] = dataset.num_classes_sub
# Load model
model = load(model_type=params['model_type'],
params=params,
backend=params['backend'])
# Restore the saved parameters
epoch, _, _, _ = model.load_checkpoint(save_path=args.model_path,
epoch=args.epoch)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
logger.info('beam width (main): %d\n' % args.beam_width)
logger.info('beam width (sub) : %d\n' % args.beam_width_sub)
logger.info('epoch: %d' % (epoch - 1))
logger.info('a2c oracle: %s\n' % str(args.a2c_oracle))
logger.info('resolving_unk: %s\n' % str(args.resolving_unk))
logger.info('joint_decoding: %s\n' % str(args.joint_decoding))
logger.info('score_sub_weight : %f' % args.score_sub_weight)
wer, df = eval_word(models=[model],
dataset=dataset,
eval_batch_size=args.eval_batch_size,
beam_width=args.beam_width,
max_decode_len=MAX_DECODE_LEN_WORD,
min_decode_len=MIN_DECODE_LEN_WORD,
beam_width_sub=args.beam_width_sub,
max_decode_len_sub=MAX_DECODE_LEN_CHAR,
min_decode_len_sub=MIN_DECODE_LEN_CHAR,
length_penalty=args.length_penalty,
coverage_penalty=args.coverage_penalty,
progressbar=True,
resolving_unk=args.resolving_unk,
a2c_oracle=args.a2c_oracle,
joint_decoding=args.joint_decoding,
score_sub_weight=args.score_sub_weight)
wer_mean += wer
logger.info(' WER (%s, main): %.3f %%' % (data_type, (wer * 100)))
logger.info(df)
wer_sub, cer_sub, df_sub = eval_char(
models=[model],
dataset=dataset,
eval_batch_size=args.eval_batch_size,
beam_width=args.beam_width_sub,
max_decode_len=MAX_DECODE_LEN_CHAR,
min_decode_len=MIN_DECODE_LEN_CHAR,
length_penalty=args.length_penalty,
coverage_penalty=args.coverage_penalty,
progressbar=True)
wer_sub_mean += wer_sub
cer_sub_mean += cer_sub
logger.info(' WER / CER (%s, sub): %.3f / %.3f %%' % (data_type,
(wer_sub * 100),
(cer_sub * 100)))
logger.info(df_sub)
logger.info(' WER (mean, main): %.3f %%' % (wer_mean * 100 / 3))
logger.info(' WER / CER (mean, sub): %.3f / %.3f %%' %
((wer_sub_mean * 100 / 3), (cer_sub_mean * 100 / 3)))