How about ShotcutA ？

[I-Doctor]

Hi，thank you for your code firstly but I'm confused why you using ShortcutB(1*1 conv) when the size of feature map is halved. Because as far as I know, ShortcutA(insert 0) is used in the net of cifar10 dataset in the original paper. I have tried to train resnet-20 with ShotcutA with Pytorch but I couldn't get such a good result as you reported. But I don't think it's the problem of shotcut since I couldn't get the result as good as the original paper either. Have you tried ShortcutA, How about the result?

[ethanhe42]

Haven't tried this, I didn't notice this detail.