Generalization for resneXt

[n01z3]

Hello!
Greaty results! How to generalization this layer for resneXt network with groups != 1 in conv layers inside Bottleneck block?

[liuguilin1225]

If you are just using partial conv for padding, setting groups = 1 in line https://github.com/NVIDIA/partialconv/blob/master/models/partialconv2d.py#L59 will work even the main convolution's groups is not 1. As what we want is just a reweighting map and the mask channel size is 1.
However, if your mask channel size is not 1, you can set the groups to be the main convolution's groups value accordingly, and divide the channel size in https://github.com/NVIDIA/partialconv/blob/master/models/partialconv2d.py#L33 by groups as well.