Initial image size about pspnet-pytorch

[Lextal]

It's fully convolutional, the only requirement for an image is that it must be divisible by 8.

[achaiah]

Hi, I thought so too, but it does not seem to work. If I pass in a 512x512 image I get: RuntimeError: Need input.size[1] == 2048 but got 512 instead. So something isn't completely right. Would you have any ideas?

Thanks

[Tshzzz]

Useing resnet50 is ok.

[Lextal]

It was due to the incorrect initialization of extractors. Fixed.

[achaiah]

Hey thanks for fixing it! I can confirm that it now loads with resnet, however, it does not train. My loss never changes. I have many other models that train just fine with the same code so it looks like the issue is somewhere within the model. Any ideas?

Thanks