Johannes Laute

Loss goes to 0

5

Hi, first of all thanks for publishing the source code, I want to try this new method in a segmentation problem I have, to avoid having to weigh the classes...

no performance improvement over nn.Conv2D(... groups=in_channels) in pytorch 1.0

3

Hi, Thanks for open sourcing your work. In my initial tests I don't see improvements over the normal pytorch Conv2D with groups = in_channels, though apparently pytorch is still not...