ReCU
ReCU copied to clipboard
Different shortcut design between cifar and imagenet in ResNet?
For ResNet in cifar experiment <ReCU/cifar/models_cifar/resnet2.py>, the shortcut/downsampling is binarized to [-1,+1] via BinarizeConv2d. However in ImageNet experiment <ReCU/imagenet/models_imagenet/resnet.py>, the shortcut/downsampling remains the fullprecision approach via nn.Conv2d. Is this the mistake in code or some other reason? Does binarizaition of shortcut lead to worse degradation?