How do you implement depthwise convolution in the paper? I know that right now tensorflow doesn't support group convolution so the naive way to achieve it is to write serial for loop between different groups. Obviously it will slow down the speed, I see your results that the forward speed is pretty fast so I'm wondering how you did it.
How do you implement depthwise convolution in the paper? I know that right now tensorflow doesn't support group convolution so the naive way to achieve it is to write serial for loop between different groups. Obviously it will slow down the speed, I see your results that the forward speed is pretty fast so I'm wondering how you did it.
Thanks!