Wrong definition of self.conv_spatial parameter in katago's model_pytorch file?

[onikatanyamaraj]

When I run the standalone synchronized self-play training, the following error is reported：

The format of the padding parameter of conv_spatial should not be a string but rather an integer tuple. So I found this conv_spatial method definition, found out that the default parameter for padding is "same". This is more in the tradition of tensorflow and not in line with pytorch's specification.

I think this is a small loophole in the development of the algorithm, can the default values here be changed?

[lightvector]

The pytorch documentation explicitly says "same" is a valid parameter: https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html

Perhaps you are using a version of pytorch that is too old?

[onikatanyamaraj]

Thank you so much for noticing my problem and helping! I should check my pytorch version.