Yingbo Zhou

I was not benchmarking, I simply substitute nn.Conv2d with Conv2dDepthwise, my conv structure looks something like the following ``` layers = [ nn.BatchNorm2d(in_feat), nn.Conv2d(in_feat, in_feat, kernel_size=kernel_size, stride=stride, padding=(f_h // 2,...

the forward is faster but the backward is a lot slower, following is the code for reproducing I got the result on K80 ``` import torch import time from torch...