Hello all,
Does anyone know if there is any module or function for doing the inverse of nn.PixelShuffle()? That is reduce spatial dimension and add to the channel.
Thanks,
Saeed
Hello all,
Does anyone know if there is any module or function for doing the inverse of nn.PixelShuffle()? That is reduce spatial dimension and add to the channel.
Thanks,
Saeed
This code should work. For some reasons I don’t know, the community don’t like to add this function into the core implementation.
def squeeze(x, r):
[B, C, H, W] = list(x.size())
x = x.reshape(B, C, H//r, r, W//r, r)
x = x.permute(0, 1, 3, 5, 2, 4)
x = x.reshape(B, C*(r**2), H//r, W//r)
return x