winycg / hcgnet Goto Github PK

你好！作者的工作很棒！有个问题想请教下，
x_shortcut_weight = self.global_pool(x_shortcut * forget_context_weight) * H * W
全局池化后维度应该是1X1XC，这里为啥要乘以H*W？

Hello, thanks for sharing your work. When I run your code on my machine with cifar10 dataset, I notice that the test accuracy is already as high as 90%, while the train accuracy is still around 50%. In my experience, training accuracy is always higher than the test accuracy, so I have a question "Why test accuracy reaches high as 90% while the train accuracy is about 50%?" Thanks again for sharing your excellent work!