Problem of backward of APN about racnn-pytorch

[YiKeYaTu]

Thanks for sharing the code. It helps me understand the APN, I have been confused by how the author crops the attention region.

In the backward code of APN, I found you used a fixed value of in_size. (If my understanding for the code is right) Did you just backpropagate the gradient to a fixed location? if it is fixed, why did you do that? If not, how you backpropagate the gradient to the attention location?

Thanks in advance

def backward(self, grad_output):
    images, ret_tensor = self.saved_variables[0], self.saved_variables[1]
    in_size = 224
    ret = torch.Tensor(grad_output.size(0), 3).zero_()
    norm = -(grad_output * grad_output).sum(dim=1)
  
    
    x = torch.stack([torch.arange(0, in_size)] * in_size).t()
    y = x.t()
    long_size = (in_size/3*2)
    short_size = (in_size/3)
    mx = (x >= long_size).float() - (x < short_size).float()
    my = (y >= long_size).float() - (y < short_size).float()
    ml = (((x<short_size)+(x>=long_size)+(y<short_size)+(y>=long_size)) > 0).float()*2 - 1
    
    mx_batch = torch.stack([mx.float()] * grad_output.size(0))
    my_batch = torch.stack([my.float()] * grad_output.size(0))
    ml_batch = torch.stack([ml.float()] * grad_output.size(0))
    
    if isinstance(grad_output, torch.cuda.FloatTensor):
        mx_batch = mx_batch.cuda()
        my_batch = my_batch.cuda()
        ml_batch = ml_batch.cuda()
        ret = ret.cuda()
    
    ret[:, 0] = (norm * mx_batch).sum(dim=1).sum(dim=1)
    ret[:, 1] = (norm * my_batch).sum(dim=1).sum(dim=1)
    ret[:, 2] = (norm * ml_batch).sum(dim=1).sum(dim=1)
    return None, ret

I have met the same problem. Have you figured this out yet?