def test_batch_mulaw(self):
waveform, sample_rate = torchaudio.load(self.test_filepath) # (2, 278756), 44100
# Single then transform then batch
waveform_encoded = transforms.MuLawEncoding()(waveform)
expected = waveform_encoded.unsqueeze(0).repeat(3, 1, 1)
# Batch then transform
waveform_batched = waveform.unsqueeze(0).repeat(3, 1, 1)
computed = transforms.MuLawEncoding()(waveform_batched)
# shape = (3, 2, 201, 1394)
self.assertTrue(computed.shape == expected.shape, (computed.shape, expected.shape))
self.assertTrue(torch.allclose(computed, expected))
# Single then transform then batch
waveform_decoded = transforms.MuLawDecoding()(waveform_encoded)
expected = waveform_decoded.unsqueeze(0).repeat(3, 1, 1)
# Batch then transform
computed = transforms.MuLawDecoding()(computed)
# shape = (3, 2, 201, 1394)
self.assertTrue(computed.shape == expected.shape, (computed.shape, expected.shape))
self.assertTrue(torch.allclose(computed, expected))
def test_mu_law_companding(self):
sig = self.sig.clone()
quantization_channels = 256
sig = self.sig.numpy()
sig = sig / np.abs(sig).max()
self.assertTrue(sig.min() >= -1. and sig.max() <= 1.)
sig_mu = transforms.MuLawEncoding(quantization_channels)(sig)
self.assertTrue(sig_mu.min() >= 0. and sig.max() <= quantization_channels)
sig_exp = transforms.MuLawExpanding(quantization_channels)(sig_mu)
self.assertTrue(sig_exp.min() >= -1. and sig_exp.max() <= 1.)
sig = self.sig.clone()
sig = sig / torch.abs(sig).max()
self.assertTrue(sig.min() >= -1. and sig.max() <= 1.)
sig_mu = transforms.MuLawEncoding(quantization_channels)(sig)
self.assertTrue(sig_mu.min() >= 0. and sig.max() <= quantization_channels)
sig_exp = transforms.MuLawExpanding(quantization_channels)(sig_mu)
self.assertTrue(sig_exp.min() >= -1. and sig_exp.max() <= 1.)
repr_test = transforms.MuLawEncoding(quantization_channels)
repr_test.__repr__()
repr_test = transforms.MuLawExpanding(quantization_channels)
repr_test.__repr__()
def test_mu_law_companding(self):
sig = self.sig.clone()
quantization_channels = 256
sig = self.sig.numpy()
sig = sig / np.abs(sig).max()
self.assertTrue(sig.min() >= -1. and sig.max() <= 1.)
sig_mu = transforms.MuLawEncoding(quantization_channels)(sig)
self.assertTrue(sig_mu.min() >= 0.
and sig.max() <= quantization_channels)
sig_exp = transforms.MuLawExpanding(quantization_channels)(sig_mu)
self.assertTrue(sig_exp.min() >= -1. and sig_exp.max() <= 1.)
#diff = sig - sig_exp
#mse = np.linalg.norm(diff) / diff.shape[0]
#self.assertTrue(mse, np.isclose(mse, 0., atol=1e-4)) # not always true
sig = self.sig.clone()
sig = sig / torch.abs(sig).max()
self.assertTrue(sig.min() >= -1. and sig.max() <= 1.)
sig_mu = transforms.MuLawEncoding(quantization_channels)(sig)
self.assertTrue(sig_mu.min() >= 0.
and sig.max() <= quantization_channels)
sig_exp = transforms.MuLawExpanding(quantization_channels)(sig_mu)
self.assertTrue(sig_exp.min() >= -1. and sig_exp.max() <= 1.)
def test_mu_law_companding(self):
quantization_channels = 256
waveform = self.waveform.clone()
waveform /= torch.abs(waveform).max()
self.assertTrue(waveform.min() >= -1. and waveform.max() <= 1.)
waveform_mu = transforms.MuLawEncoding(quantization_channels)(waveform)
self.assertTrue(waveform_mu.min() >= 0. and waveform_mu.max() <= quantization_channels)
waveform_exp = transforms.MuLawDecoding(quantization_channels)(waveform_mu)
self.assertTrue(waveform_exp.min() >= -1. and waveform_exp.max() <= 1.)
def test_mu_law_companding(self):
quantization_channels = 256
waveform = self.waveform.clone()
if not waveform.is_floating_point():
waveform = waveform.to(torch.get_default_dtype())
waveform /= torch.abs(waveform).max()
self.assertTrue(waveform.min() >= -1. and waveform.max() <= 1.)
waveform_mu = transforms.MuLawEncoding(quantization_channels)(waveform)
self.assertTrue(waveform_mu.min() >= 0. and waveform_mu.max() <= quantization_channels)
waveform_exp = transforms.MuLawDecoding(quantization_channels)(waveform_mu)
self.assertTrue(waveform_exp.min() >= -1. and waveform_exp.max() <= 1.)
def test_MuLawEncoding(self):
tensor = common_utils.get_whitenoise()
self._assert_consistency(T.MuLawEncoding(), tensor)
def test_MuLawEncoding(self):
tensor = torch.rand((1, 10))
self._assert_consistency(T.MuLawEncoding(), tensor)
inputs = inputs[:x[-1].astype(int)]
targets = targets[:x[-1].astype(int)]
inputs = inputs[:len(inputs) // seq_M * seq_M]
targets = targets[:len(targets) // seq_M * seq_M]
h = f(np.arange(1, len(inputs) + 1))
train_wav.append(inputs)
train_features.append(h)
train_targets.append(targets)
train_wav = np.concatenate(train_wav)
train_features = np.vstack(train_features)
train_targets = np.concatenate(train_targets)
enc = transforms.MuLawEncoding(channels)
dec = transforms.MuLawExpanding(channels)
train_wav = enc(train_wav)
train_targets = enc(train_targets)
scaler = StandardScaler()
train_features = scaler.fit_transform(train_features)
train_wav = train_wav.reshape(-1, seq_M)
train_features = np.rollaxis(
train_features.reshape(-1, seq_M, features_size), 2, 1)
train_targets = train_targets.reshape(-1, seq_M)
train_wav = torch.from_numpy(train_wav).long()
train_features = torch.from_numpy(train_features).float()
depth = args.depth
radixs = [2] * depth
N = np.prod(radixs)
channels = args.channels
lr = args.lr
steps = args.steps
c = args.c
generation_time = args.file_size
filename = args.outfile
maxlen = 50000
print('==> Downloading YesNo Dataset..')
transform = transforms.Compose(
[transforms.Scale(),
transforms.PadTrim(maxlen),
transforms.MuLawEncoding(quantization_channels=channels)])
data = torchaudio.datasets.YESNO('./data', download=True, transform=transform)
data_loader = DataLoader(data, batch_size=batch_size, num_workers=4, shuffle=True)
print('==> Building model..')
net = general_FFTNet(radixs, 128, channels).cuda()
print(sum(p.numel() for p in net.parameters() if p.requires_grad), "of parameters.")
optimizer = optim.Adam(net.parameters(), lr=lr)
criterion = torch.nn.CrossEntropyLoss()
print("Start Training.")
a = datetime.now().replace(microsecond=0)
step = 0