def _call(self, x):
dim = -self.event_dim
if dim != -1:
x = x.transpose(dim, -1)
y = dct(x)
if dim != -1:
y = y.transpose(dim, -1)
return y
def _call(self, x):
dim = self.dim
if dim != -1:
x = x.transpose(dim, -1)
y = dct(x)
if self.smooth:
y = y * self._weight(y)
if dim != -1:
y = y.transpose(dim, -1)
return y
def _transform_forward(x, dim, duration):
assert not x.requires_grad
assert dim < 0
assert duration == x.size(dim)
time_domain = x
new_size = next_fast_len(duration)
if new_size == duration:
freq_domain = x
else:
freq_domain = pad(x, (0, 0) * (-1 - dim) + (0, new_size - duration))
freq_domain = dct(freq_domain, dim)
return torch.cat([time_domain, freq_domain], dim=dim)
def test_dct(shape):
x = torch.randn(shape)
actual = dct(x)
expected = torch.from_numpy(fftpack.dct(x.numpy(), norm='ortho'))
assert_close(actual, expected)