Questions about Channel-wise Separable Convolutions

[wangyue7777]

Hi! Thanks for the great work! I have a question related to the Channel-wise Separable Convolutions part in the RNL module.

In your paper, you mention that Fθ should not fuse together information across channels. But in your code, the parameter of groups in DepthwiseConv3d is not equal to the in_channels, it is "groups = max(in_channels // 32, 1)".

Are there any specific reasons for you to choose this group?

Thank you!

[russellllaputa]

Hi,

Thank you for interesting my work.

For the implementation of Fθ, I did use channel-wise(depth-wise) convolution, which mean the numbe of groups is equal to the channel number. However, in Pytorch, setting the groups parameter as the channel number for a Conv would cuase low efficency. I guess Pytorch1.3 implemented it with a loop operation. So I code the DepthwiseConv3d myself by setting masks for some channels.

I have tested that in Pytorch-1.6 or higher versions, they have make the Convolution with multiple goups more efficient. But considering that the trained models still use Pytorch1.3, so I did not change it.

[wangyue7777]

Hi,

Thanks for the reply! It solved all my problems.

[wangyue7777]

Hi,

Thanks for the reply! It solved all my problems.