def inference():
hparams = create_hparams()
_type = 'test'
name = 'pff'
# name = 'position-encoding-train'
vocoder = 'griffin_lim'
# vocoder = 'waveglow'
# name = 'multi-gpu'
step = 320 * 1000
waveglow_path = '/data1/hfzeng/work/tacotron/original_tacotron/waveglow/waveglow_old.pt'
waveglow = torch.load(waveglow_path)['model']
waveglow.remove_weightnorm()
waveglow.cuda().eval()
# texts = get_texts('./sentence.txt')
h_path = './presets/{}.json'.format(name)
hparams = load_hparams(h_path, hparams)
model = load_inference_model(hparams, name, step)
for _type in ['val']: #, 'val', 'test']:
texts = get_texts(_type=_type)
inference_texts(model,
hparams,
texts,
step,
name,
vocoder=vocoder,
waveglow=waveglow,
_type=_type,
postnet=False)
def main():
config = load_hparams()
data_handler = DataHandler(config)
batcher = Batcher(data_handler)
with tf.Session() as session:
model = Model(session=session, config=config, batcher=batcher)
model.build_graph(session)
model.load()
model.train()
def test_inference():
hparams = create_hparams()
name = 'position-encoding-train'
h_path = './presets/{}.json'.format(name)
hparams = load_hparams(h_path, hparams)
step = 770 * 1000
# model = load_inference_model(hparams, name, step)
texts = get_texts('./sentence.txt')
seq, pos = prepare_inputs(hparams, text=texts[0])
model = load_inference_model(hparams, name, step)
encoder_outputs = model.encoder.inference(seq, pos)
print(encoder_outputs)
import os
from unittest import TestCase
from data.data_manager import DataManager, LabeledPairDataManager
from hparams import load_hparams
abspath = os.path.abspath(os.path.dirname(__file__))
hparams = load_hparams(
os.path.join(abspath, '../data/test_files/datasets/unlabeled_pair/'))
class TestDataManager(TestCase):
def setUp(self):
self.sess = None
self.data_manager = DataManager(hparams=hparams)
def tearDown(self):
del self.data_manager
def test_indexing(self):
sentence = '1 4 6 8 2 3'
answer = [2, 4, 0, 0, 0, 5, 6, 1, 1, 3]
result = self.data_manager.indexing(sentence)
self.assertListEqual(answer, result)
class TestLabeledPairDataManager(TestCase):
def setUp(self):
self.sess = None
self.data_manager = LabeledPairDataManager(hparams=hparams,
fmax=8000)
print('melspectrogram.shape', melspectrogram.shape)
print(melspectrogram)
plt.figure(figsize=(10, 4))
plt.subplot(2, 1, 1)
librosa.display.specshow(np.log(melspectrogram),
y_axis='mel',
x_axis='time',
hop_length=256,
fmin=0,
fmax=8000)
plt.title('log Mel spectrogram')
plt.subplot(2, 1, 2)
hparams = load_hparams()
log_mel = get_mel("LJ001-0002.wav", hparams)
librosa.display.specshow(log_mel.data.numpy(),
y_axis='mel',
x_axis='time',
hop_length=256,
fmin=0,
fmax=8000)
plt.title('original Mel spectrogram')
plt.tight_layout()
plt.show()
audio_signal = librosa.feature.inverse.mel_to_audio(melspectrogram,
sr=22050,
n_fft=1024,
hop_length=256,