def postprocess(self, data: Dict,
dataset: data_manager.CustomDataset) -> List[Dict]:
samples, outs = [], []
# speech de-normalization, de-emphasis
xs = data['x'].permute(0, 2, 1) # B, T, C
xs = dataset.norm_modules['x'].denormalize_(xs)
xs = de_emphasis(gen.convert(xs[..., 0], astype=np.ndarray))
ys = data['y']
ys = dataset.norm_modules['y'].denormalize_(ys)
ys = de_emphasis(gen.convert(ys, astype=np.ndarray))
outs = data['out']
outs = dataset.norm_modules['y'].denormalize_(outs)
outs = de_emphasis(gen.convert(outs, astype=np.ndarray))
data['x'], data['y'], data['out'] = xs, ys, outs
T_xs = data['T_xs'].int()
T_ys = data['T_ys'].int()
# Dict[List] -> List[Dict]
for idx in range(len(data['x'])):
sample = dict()
T_x, T_y = T_xs[idx], T_ys[idx]
sample['T_xs'], sample['T_ys'] = T_x, T_y
for key, value in data.items():
value = value[idx]
if key == 'x':
value = value[..., :T_x]
value = np.nan_to_num(value, posinf=1., neginf=-1.)
value = np.asfortranarray(np.clip(value, -1, 1))
elif len(key) > 3:
pass
else:
value = value[..., :T_y]
value = np.nan_to_num(value, posinf=1., neginf=-1.)
value = np.asfortranarray(np.clip(value, -1, 1))
sample[key] = value
samples.append(sample)
return samples
def test_pre_emphasis(self):
"""
Tests equality after de-emphasizing pre-emphasized signal.
"""
rand_signal_batch = np.random.randint(low=1,
high=10,
size=(10, 1, 400))
reconst_batch = emph.de_emphasis(emph.pre_emphasis(rand_signal_batch))
# after de-emphasis, the signal must have been restored
self.assertEqual(rand_signal_batch.shape, reconst_batch.shape)
self.assertTrue(np.allclose(rand_signal_batch, reconst_batch))
print(
'Epoch {}, Step {}, d_clean_loss {}, d_noisy_loss {}, g_loss {}, g_loss_cond {}'
.format(epoch + 1, i + 1, clean_loss.data[0],
noisy_loss.data[0], g_loss.data[0],
g_cond_loss.data[0]))
### Functions below print various information about the network. Uncomment to use.
# print('Weight for latent variable z : {}'.format(z))
# print('Generated Outputs : {}'.format(generated_outputs))
# print('Encoding 8th layer weight: {}'.format(generator.module.enc8.weight))
# save sampled audio at the beginning of each epoch
if i == 0:
fake_speech = generator(fixed_test_noise, z)
fake_speech_data = fake_speech.data.cpu().numpy(
) # convert to numpy array
fake_speech_data = emph.de_emphasis(fake_speech_data,
emph_coeff=0.95)
for idx in range(4): # select four samples
generated_sample = fake_speech_data[idx]
filepath = os.path.join(
gen_data_path,
'{}_e{}.wav'.format(test_noise_filenames[idx], epoch + 1))
wavfile.write(filepath, sample_rate, generated_sample.T)
# save the model parameters for each epoch
g_path = os.path.join(models_path, 'generator-{}.pkl'.format(epoch + 1))
d_path = os.path.join(models_path,
'discriminator-{}.pkl'.format(epoch + 1))
torch.save(generator.state_dict(), g_path)
torch.save(discriminator.state_dict(), d_path)
print('Finished Training!')