Closed 8secz-johndpope closed 3 years ago
def INF(B,H,W): return -torch.diag(torch.tensor(float("inf")).cuda().repeat(H),0).unsqueeze(0).repeat(B*W,1,1) relates to the square which is overlaped in criss-cross 0,3,1,2 transposes the tensor for computing similarity score below I have not read 3d ccnet carefully
https://github.com/Serge-weihao/CCNet-Pure-Pytorch/blob/bb502bb32f1d8eadbd7fb06152be570c23e9fbd1/networks/CC.py#L6
def INF(B,H,W): return -torch.diag(torch.tensor(float("inf")).cuda().repeat(H),0).unsqueeze(0).repeat(B*W,1,1)
is this related to the white squares that are not criss crossed?
does the blue dot in question - presumably - it can't go any further left or right? how does algorithm handle this 'edge' case? Are these the 'residual connections'? how does the code handle this?
Were the efforts to change the length of the cross?
If I had to comment the code
what is 0,3,1,2 related to ?
In the paper it mentions a 3d criss cross implementation - with a T / temporarl parameter introduced - does this exist in this code?
Where is H prime? Is that connected to the energy?
Sorry - all these noob questions - thanks for any help you can shed light on.