Francisco Massa notifications@github.com kirjoitti 22.10.2015 kello 1.05:

Is there any interest in adding groups support to SpationConvolutionMM ?

Groups behaviour can be simulated by combining a few modules (like Concat + Narrow + SpatialConvolution, for example), but is it worth it adding dedicated C/Cuda paths for this option ?

The only reason why I would see some interest is to allow some Caffe models to work in vanilla nn (like AlexNet).

I am interested in using AlexNet in vanilla nn (without GPU/cuda).

But I am just a newcomer in deep learning so you are free to ignore my wishes.

Hannu Töyrylä
http://liipetti.net/erratic/

If there is interest in it, I can do it.

—
Reply to this email directly or view it on GitHub.

@fmassa the group feature in Caffe and the need to convert caffe models to torch models is what brought me here. If you can add in this feature, it would be amazing!

@varunagrawal I think the easiest for now would be to add the groups support in the lua level, as done in PyTorch. I unfortunately won't have time in the near future to do it, as I've moved my stack to PyTorch.

Alright. I wanted to convert some caffe models to torch and then to PyTorch, but the lack of the groups feature in torch is preventing me from doing that. Any ideas on what can I do alternatively?

Nevermind, found a solution. loadcaffe can take cudnn as an argument for the backend and the torch cudnn bindings support groups. Cheers!