I am trying to implement a paper that uses a 1/2 stride CONV layer as in-network upsampling. However, when I enter 0.5 as value for stride in nn.Conv2d, it (obviously) throws an error. Any suggestions on this? This is a really popular paper, so I’m sure it has been done in PyTorch before!
Link to paper’s network architecture: http://cs.stanford.edu/people/jcjohns/papers/eccv16/JohnsonECCV16Supplementary.pdf