Closed bominn closed 2 years ago
Hi Bominn, Good observation. Thanks for your feedback. Yes, it's true that I changed the source different with original paper on this part. The problem that the convolution layer in my case will change the original image size, so c_next is different shape with c_cur. Then I used c_cur as peephole input in my implement. If your c_next keep same size as c_cur. There is no issue to change to c_next in your case.
https://github.com/KimUyen/ConvLSTM-Pytorch/blob/c7b4bd108335a4d6c7d99c00c263346026186b0b/convlstm.py#L114
According to the formula in README, it should be c_next in this line instead of c_cur, right?