Hard-coded paadings lead to bugs when changing input image size in pytorch implementation

[JohnTargaryen]

Greetings! I change the input image size from 256,256,3 to 128,128,3, and mask size from 128,128 to 64,64 in the pytorch implementation. However, there seem to be hard-coded padding in the network that causes bugs when changing input size.

Could you please explain what these padding operation do? And how should I modify them to adapt to my desired input sizes?

[JohnTargaryen]

And I also wonder is multi-gpu training implemented in the pytorch implementation?

[shepnerd]

The padding size calculation using dilation conv is a bit different from the common conv, and you highlighted part just hard-codes these parameters. For the detailed rules, you can refer to this post.

If you intend to process 128x128 images, you can remove the fourth dilation conv layer (with dilation=16) and its corresponding padding operation, or ensure the dilation ratio will not exceed than a specific number (refer to the padding computation in that post) and alter its corresponding padding.

For the multi-gpu training, you can use data parallel (torch.nn.DataParallel). As far as I know, current some implementations (maybe losses) are not supported in that and I will check that soon.