Autograd safety of Variable Sort

autograd
1 
    def forward(self, input):
       input,_ = input.sort(dim=1, descending=True)
       out = nn.functional.conv2d(input, self.weight, None, self.stride, self.padding, self.dilation, self.groups)

       return out

My concern is if above code of reordering input can affect to the calculation of the autograd gradients or are safe.

Thanks in advance!

2

The returned values from sort() are differentiable if that is what you are asking. For

input2, indices = input1.sort(dim=1, descending=True)

The gradient of input1 will be the re-ordered gradient of input2. For example,

input1 = torch.tensor([1., 3., 2.], requires_grad=True)
input2, _ = input1.sort(descending=True)  # 3, 2, 1
input2.backward(torch.tensor([0.3, -0.2, 0.01]))
print(input1.grad)  # 0.01, 0.3, -0.2
2 Likes