def __init__(self, hparams):
super(MixtureGaussianLoss, self).__init__()
self.quantize_channels = hparams.quantize_channels
self.log_scale_min = hparams.log_scale_min
self.mix_gaussian_loss = mix_gaussian_loss(log_scale_min=hparams.log_scale_min, reduce=False)
self.reduce_sum_op = P.ReduceSum()
self.reduce_mean_op = P.ReduceMean()
def test_gaussian_mixture():
np.random.seed(1234)
x, sr = librosa.load(pysptk.util.example_audio_file(), sr=None)
assert sr == 16000
T = len(x)
x = x.reshape(1, T, 1)
y = torch.from_numpy(x).float()
y_hat = torch.rand(1, 30, T).float()
print(y.shape, y_hat.shape)
loss = mix_gaussian_loss(y_hat, y)
print(loss)
loss = mix_gaussian_loss(y_hat, y, reduce=False)
print(loss.size(), y.size())
assert loss.size() == y.size()
y = sample_from_mix_gaussian(y_hat)
print(y.shape)
def forward(self, input, target, lengths=None, mask=None, max_len=None):
if lengths is None and mask is None:
raise RuntimeError("Should provide either lengths or mask")
# (B, T, 1)
if mask is None:
mask = sequence_mask(lengths, max_len).unsqueeze(-1)
# (B, T, 1)
mask_ = mask.expand_as(target)
losses = mix_gaussian_loss(
input, target, log_scale_min=hparams.log_scale_min, reduce=False)
assert losses.size() == target.size()
return ((losses * mask_).sum()) / mask_.sum()