def _transform_inverse(x: torch.Tensor, dim: int, duration: int):
assert not x.requires_grad
assert dim < 0
dots = (slice(None), ) * (x.dim() + dim)
left = dots + (slice(None, duration), )
right = dots + (slice(duration, None), )
return idct(x[right], dim)[left].add_(x[left])
def _inverse(self, y):
dim = -self.event_dim
if dim != -1:
y = y.transpose(dim, -1)
x = idct(y)
if dim != -1:
x = x.transpose(dim, -1)
return x
def _inverse(self, y):
dim = self.dim
if dim != -1:
y = y.transpose(dim, -1)
if self.smooth:
y = y / self._weight(y)
x = idct(y)
if dim != -1:
x = x.transpose(dim, -1)
return x
def test_idct(shape):
x = torch.randn(shape)
actual = idct(x)
expected = torch.from_numpy(fftpack.idct(x.numpy(), norm='ortho'))
assert_close(actual, expected)