About preresnet results

[WILL LEE]

Hi, @timgaripov ,@izmailovpavel ,

Thanks for your pretty work & nice paper.

from the issues
bearpaw/pytorch-classification#6
and bearpaw/pytorch-classification#9

preresnet-110 is not really 110 layers,
because we use bottleneck. ((110-2)/6 =18, 18*9+2=164 layers)

        n = (depth - 2) // 6
        block = Bottleneck if depth >= 44 else BasicBlock

the correct way is:

        if depth >= 44:
            assert (depth - 2) % 9 == 0
            n = (depth - 2) // 9
            block = Bottleneck

        else:
            assert (depth - 2) % 6 == 0, 'depth should be 6n+2'
            n = (depth - 2) // 6
            block = BasicBlock

so the results of preresnet seems not correct,
look forward the new experiments in swa!

thanks a lot!

[Pavel Izmailov]

Thanks for bringing this up! We will change the PreResNet110 -> PreResNet164 in the table and also run experiments with the correct PreResNet110.

[Pavel Izmailov]

We fixed the implementation, and added new results for the actual PreResNet-110. Thanks for bringing this issue up! Now closing the issue.