def iter_asr(set_name, set_train_mode, set_rr):
assert set_name in ['train', 'dev', 'test']
rr = set_rr
rr = sort_reverse(rr, feat_len[set_name])
rr_key = feat_iterator[set_name].get_key_by_index(rr)
curr_feat_list = feat_iterator[set_name].get_feat_by_key(rr_key)
curr_text_list = text_iterator[set_name].get_text_by_key(rr_key)
curr_feat_mat, curr_feat_len, curr_text_mat, curr_text_len = batch_speech_text(
opts['gpu'], curr_feat_list, curr_text_list)
_loss, _acc = fn_batch_asr(model_asr,
curr_feat_mat,
curr_feat_len,
curr_text_mat,
curr_text_len,
train_step=set_train_mode,
coeff_loss=opts['coeff_pair'])
_loss /= opts['coeff_pair']
assert_nan(_loss)
_count = len(rr)
m_asr_loss[set_name] += _loss * _count
m_asr_acc[set_name] += _acc * _count
m_asr_count[set_name] += _count
if tf_writer is not None:
auto_writer_info_asr(set_name, _loss, _acc)
def iter_tts(set_name, set_train_mode, set_rr):
assert set_name in ['train', 'dev', 'test']
rr = set_rr
rr = sort_reverse(rr, text_len[set_name])
rr_key = text_iterator[set_name].get_key_by_index(rr)
curr_feat_list = feat_iterator[set_name].get_feat_by_key(rr_key)
curr_text_list = text_iterator[set_name].get_text_by_key(rr_key)
if model_tts.TYPE == TacotronType.MULTI_SPEAKER:
curr_spkvec_list = feat_spkvec_iterator.get_feat_by_key(rr_key)
curr_aux_info = {'speaker_vector': curr_spkvec_list}
else:
curr_aux_info = None
curr_feat_mat, curr_feat_len, curr_text_mat, curr_text_len = batch_speech_text(
opts['gpu'],
curr_feat_list,
curr_text_list,
feat_sil=feat_sil,
group=opts['tts_group'],
start_sil=1,
end_sil=opts['tts_pad_sil'])
_loss, _loss_feat, _loss_bce_fend, _loss_spk_emb, _acc_fend = fn_batch_tts(
model_tts,
curr_text_mat,
curr_text_len,
curr_feat_mat,
curr_feat_len,
aux_info=curr_aux_info,
train_step=set_train_mode,
coeff_loss=opts['coeff_pair'])
_loss /= opts['coeff_pair']
assert_nan(_loss)
_count = len(rr)
m_tts_loss[set_name] += _loss * _count
m_tts_loss_feat[set_name] += _loss_feat * _count
m_tts_loss_bce[set_name] += _loss_bce_fend * _count
m_tts_loss_spk_emb[set_name] += _loss_spk_emb * _count
m_tts_acc[set_name] += _acc_fend * _count
m_tts_count[set_name] += _count
if tf_writer is not None:
auto_writer_info_tts(set_name, _loss, _loss_feat,
_loss_bce_fend, _loss_spk_emb, _acc_fend)
curr_key_list)
if opts['strip_sil']:
curr_feat_list = list_feat_sil_strip(curr_feat_list)
curr_label_list = text_iterator[set_name].get_text_by_key(
curr_key_list)
aux_info = None
if model.TYPE == TacotronType.MULTI_SPEAKER:
curr_spkvec_list = feat_spkvec_iterator.get_feat_by_key(
curr_key_list)
aux_info = {'speaker_vector': curr_spkvec_list}
# print(1, timeit.default_timer() - tic); tic = timeit.default_timer()
feat_mat, feat_len, text_mat, text_len = batch_speech_text(
opts['gpu'],
curr_feat_list,
curr_label_list,
feat_sil=feat_sil,
group=opts['group'],
start_sil=1,
end_sil=opts['pad_sil'])
# print(2, timeit.default_timer() - tic); tic = timeit.default_timer()
_tmp_loss, _tmp_loss_feat, _tmp_loss_bernend, _tmp_acc_bernend = fn_batch(
text_mat,
text_len,
feat_mat,
feat_len,
aux_info=aux_info,
train_step=set_train_mode)
# print(3, timeit.default_timer() - tic); tic = timeit.default_timer()
_tmp_count = len(rr)
assert_nan(_tmp_loss)
mloss[set_name] += _tmp_loss * _tmp_count
for set_name, set_rr, set_train_mode in [('train', train_rr, True),
('dev', dev_rr, False),
('test', test_rr, False)]:
for rr in tqdm_wrapper(set_rr):
start = time.time()
curr_key_list = feat_iterator[set_name].get_key_by_index(rr)
curr_feat_list = feat_iterator[set_name].get_feat_by_key(
curr_key_list)
curr_label_list = text_iterator[set_name].get_text_by_key(
curr_key_list)
assert (feat_iterator[set_name].get_key_by_index(
rr) == text_iterator[set_name].get_key_by_index(rr)
), 'key(s) not same'
# print(1, start-time.time()); start = time.time()
feat_mat, feat_len, text_mat, text_len = batch_speech_text(
opts['gpu'], curr_feat_list, curr_label_list)
# print(2, start-time.time()); start = time.time()
_tmp_loss, _tmp_acc = fn_batch(feat_mat,
feat_len,
text_mat,
text_len,
train_step=set_train_mode)
# print(3, start-time.time()); start = time.time()
_tmp_count = len(rr)
assert_nan(_tmp_loss)
mloss[set_name] += _tmp_loss * _tmp_count
macc[set_name] += _tmp_acc * _tmp_count
mcount[set_name] += _tmp_count
pass
info_header = ['set', 'loss', 'acc']
info_table = []
max_target=opts['max_target'])
elif opts['mode'] == 'tf':
curr_key_list = text_iterator.get_key_by_index(rr)
curr_text_list = text_iterator.get_text_by_key(curr_key_list)
curr_feat_list = data_iterator.get_feat_by_key(curr_key_list)
if model.TYPE == TacotronType.MULTI_SPEAKER:
_spk_vec = np.stack(
feat_spkvec_iterator.get_feat_by_key(
curr_key_list)).astype('float32')
_spk_vec = Variable(
tensorauto(opts['gpu'], torch.from_numpy(_spk_vec)))
aux_info = {'speaker_vector': _spk_vec}
feat_mat, feat_len, text_mat, text_len = batch_speech_text(
opts['gpu'],
curr_feat_list,
curr_text_list,
feat_sil=feat_sil,
group=group,
start_sil=1,
end_sil=0)
pred_feat, pred_len, pred_att = generator_speech.decode_greedy_tf_torch(
model,
Variable(text_mat),
text_len,
Variable(feat_mat),
feat_len,
group=group,
feat_sil=feat_sil,
aux_info=aux_info)
pred_len = data_iterator.get_feat_length_by_key(
text_iterator.get_key_by_index(rr))