def _wrapper(self, bit=10, mulaw=True):
model = _create_model(local_size=0)
dataset = SignWaveDataset(
sampling_rate=sampling_rate,
sampling_length=sampling_length,
bit=bit,
mulaw=mulaw,
)
updater, reporter = setup_support(batch_size, gpu, model, dataset)
trained_nll = _get_trained_nll()
def _first_hook(o):
self.assertTrue(o["main/nll_coarse"] > trained_nll)
def _last_hook(o):
self.assertTrue(o["main/nll_coarse"] < trained_nll)
train_support(iteration, reporter, updater, _first_hook, _last_hook)
# save model
torch.save(
model.predictor.state_dict(),
"/tmp/"
f"test_training_wavernn"
f"-bit={bit}"
f"-mulaw={mulaw}"
f"-speaker_size=0"
f"-iteration={iteration}.pth",
)
def _wrapper(self, bit=10, mulaw=True):
scale = 4
model = _create_model(
local_size=2 * scale,
local_scale=scale,
)
dataset = DownLocalRandomDataset(
sampling_rate=sampling_rate,
sampling_length=sampling_length,
scale=scale,
bit=bit,
mulaw=mulaw,
)
updater, reporter = setup_support(batch_size, gpu, model, dataset)
trained_nll = _get_trained_nll()
def _first_hook(o):
self.assertTrue(o["main/nll_coarse"] > trained_nll)
def _last_hook(o):
self.assertTrue(o["main/nll_coarse"] < trained_nll)
train_support(iteration, reporter, updater, _first_hook, _last_hook)
def _wrapper(self, to_double=False, bit=10, mulaw=True):
scale = 4
model = _create_model(
local_size=2 * scale,
local_scale=scale,
)
dataset = DownLocalRandomDataset(
sampling_length=sampling_length,
scale=scale,
to_double=to_double,
bit=bit,
mulaw=mulaw,
local_padding_size=0,
)
updater, reporter = setup_support(batch_size, gpu, model, dataset)
trained_nll = _get_trained_nll()
def _first_hook(o):
self.assertTrue(o['main/nll_coarse'].data > trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data > trained_nll)
def _last_hook(o):
self.assertTrue(o['main/nll_coarse'].data < trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data < trained_nll)
train_support(iteration, reporter, updater, _first_hook, _last_hook)
def _wrapper(self, to_double=False, bit=10, mulaw=True):
speaker_size = 4
model = _create_model(
local_size=0,
speaker_size=speaker_size,
)
datasets = [
SignWaveDataset(
sampling_rate=sampling_rate,
sampling_length=sampling_length,
to_double=to_double,
bit=bit,
mulaw=mulaw,
frequency=(i + 1) * 110,
) for i in range(speaker_size)
]
dataset = SpeakerWavesDataset(
wave_dataset=ConcatenatedDataset(*datasets),
speaker_nums=list(
chain.from_iterable([i] * len(d)
for i, d in enumerate(datasets))),
)
updater, reporter = setup_support(batch_size, gpu, model, dataset)
trained_nll = _get_trained_nll()
def _first_hook(o):
self.assertTrue(o['main/nll_coarse'].data > trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data > trained_nll)
def _last_hook(o):
self.assertTrue(o['main/nll_coarse'].data < trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data < trained_nll)
train_support(iteration, reporter, updater, _first_hook, _last_hook)
# save model
serializers.save_npz(
'/tmp/'
f'test_training_wavernn'
f'-to_double={to_double}'
f'-bit={bit}'
f'-mulaw={mulaw}'
f'-speaker_size={speaker_size}'
f'-iteration={iteration}.npz',
model.predictor,
)
def _wrapper(self, bit=10, mulaw=True):
speaker_size = 4
model = _create_model(
local_size=0,
speaker_size=speaker_size,
)
datasets = [
SignWaveDataset(
sampling_rate=sampling_rate,
sampling_length=sampling_length,
bit=bit,
mulaw=mulaw,
frequency=(i + 1) * 110,
) for i in range(speaker_size)
]
dataset = SpeakerWavesDataset(
wave_dataset=ConcatDataset(datasets),
speaker_nums=list(
chain.from_iterable([i] * len(d)
for i, d in enumerate(datasets))),
)
updater, reporter = setup_support(batch_size, gpu, model, dataset)
trained_nll = _get_trained_nll()
def _first_hook(o):
self.assertTrue(o["main/nll_coarse"] > trained_nll)
def _last_hook(o):
self.assertTrue(o["main/nll_coarse"] < trained_nll)
train_support(iteration, reporter, updater, _first_hook, _last_hook)
# save model
torch.save(
model.predictor.state_dict(),
"/tmp/"
f"test_training_wavernn"
f"-bit={bit}"
f"-mulaw={mulaw}"
f"-speaker_size={speaker_size}"
f"-iteration={iteration}.pth",
)
def _wrapper(self, to_double=False, bit=10, mulaw=True):
model = _create_model(local_size=0)
dataset = SignWaveDataset(
sampling_rate=sampling_rate,
sampling_length=sampling_length,
to_double=to_double,
bit=bit,
mulaw=mulaw,
)
updater, reporter = setup_support(batch_size, gpu, model, dataset)
trained_nll = _get_trained_nll()
def _first_hook(o):
self.assertTrue(o['main/nll_coarse'].data > trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data > trained_nll)
def _last_hook(o):
self.assertTrue(o['main/nll_coarse'].data < trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data < trained_nll)
train_support(iteration, reporter, updater, _first_hook, _last_hook)
# save model
serializers.save_npz(
'/tmp/'
f'test_training_wavernn'
f'-to_double={to_double}'
f'-bit={bit}'
f'-mulaw={mulaw}'
f'-speaker_size=0'
f'-iteration={iteration}.npz',
model.predictor,
)