While trying to implement WGan with gradient penalty for 3D volumes, I realized that MaxPool3d cannot be differentiated twice, while it works for MaxPool2d or even AvgPool3d.
Here is the code I am using :
class Network(nn.Module):
def __init__(self):
super().__init__()
self.main = nn.AvgPool3d(2) # Works
# self.main = nn.MaxPool3d(2) # Does not work
# self.main = nn.MaxPool2d(2) # Works for 2d
def forward(self, x):
return self.main(x)
net = Network()
net.cuda()
in_var = Variable(torch.Tensor(2, 2, 2, 2, 2).cuda(), requires_grad=True)
grad_out = Variable(torch.ones(2, 2, 1, 1, 1).cuda())
# in_var = Variable(torch.Tensor(2, 2, 2, 2).cuda(), requires_grad=True) # Works for 2d
# grad_out = Variable(torch.ones(2, 2, 1, 1).cuda()) # Works for 2d
gradient = autograd.grad(outputs=net(in_var), inputs=in_var,
grad_outputs=grad_out,
create_graph=True, retain_graph=True,
only_inputs=True)[0]
gradient.mean().backward()Thanks
While trying to implement WGan with gradient penalty for 3D volumes, I realized that MaxPool3d cannot be differentiated twice, while it works for MaxPool2d or even AvgPool3d.
Here is the code I am using :
Thanks