the activation fn of depthwiseconv in res connectionis different from that of tensorflow version?? about ghostnet.pytorch

[iamhankai]

Thanks for your remind. I'll fix it soon. The tensorflow version is the standard version (https://github.com/iamhankai/ghostnet).

[iamhankai]

In addition, there is another bug: the kernel size in self.shortcut = nn.Sequential( depthwise_conv(inp, inp, 3, stride, relu=True), nn.Conv2d(inp, oup, 1, 1, 0, bias=False), nn.BatchNorm2d(oup), should be set as kernel_size.

[glenn-jocher]

@iamhankai has this bug #21 (comment) been fixed in the code? Last update to ghost_net.py says two months ago but your comment is more recent.

If not can you point to where to make the change?

[glenn-jocher]

@hurri2000 @iamhankai I've submitted #23 to fix #21 (comment),

For the kernel size replacement, do you mean

ghostnet.pytorch/ghost_net.py

Line 101 in 2c90e67

depthwise_conv(inp, inp, 3, stride, relu=True),

should be this?
depthwise_conv(inp, inp, kernel_size, stride, relu=False)

[glenn-jocher]

@hurri2000 @iamhankai I've updated both changes in #23. @iamhankai can you review please?

Also @iamhankai are there any other bug fixes or changes needed that you know of? Thank you!!

[iamhankai]

Thanks @glenn-jocher I merged your pull request.