Feature request: group convolutions

[skariel]

as explained, for e.g., here. These are more lightweight than the regular convolutions, and are appropriate for some use-cases, for e.g. this. They are available in tensorflow ad Pytorch

[denizyuret]

Does CuDNN support them? What library does Tensorflow and Pytorch borrow this from?

[skariel]

I changed the name to "grouped convolutions", these are a generalization and allow to trivially build depth-wise convolutions. The functionality is that the channels are divided into groups and are multiplied all-vs-all only withing their group

[skariel]

@denizyuret Grouped convolutions are implemented in libcudnn see the definition of cudnnSetConvolutionGroupCount. Should be simple to add, I'll try this myself later today. Of course, a version for CPU has to be added as well....

[rened]

@skariel Did you have any success? I can add the CPU version.

[skariel]

the significance of this feature: MobileNets, EfficientNet, and basically all SOTA networks on Imagenet use grouped convolutions

[ngphuoc]

Related work done in Flux, NNlib, CuArrays: JuliaGPU/CuArrays.jl#523

This is a nice reading with illustrations:

https://towardsdatascience.com/a-comprehensive-introduction-to-different-types-of-convolutions-in-deep-learning-669281e58215

[denizyuret]

This will be in the next release of Knet for KnetArrays, you can test with pkg"add Knet#bea8d4b8". CuArrays support is needed I don't think anybody is working on this now.