DGX-1 Caffe Version Problem

I am trying to add a custom c++ layer defining a new loss layer (central loss) for Caffe. The reason going with the c++ approach and not a python layer is for efficiency reasons (since python layers currently are limited to a single GPU). More specifically make use of https://raw.githubusercontent.com/ydwen/caffe-face

However, compared to the docker images with DIGITS or Nividia-Caffe for the DGX-1, they seem to use another version of Caffe. They make use of different functions and different templates etc. for the layers. If we look at Nividia-Caffe 0.15 (on GitHub) it complies with our center loss, while the experimental branch does not.

Is there any docker image using Nividia-Caffe 0.15 (or any other version which complies with the Caffe templates etc) available for the DGX-1?
Also, what is special with the docker images for the DGX-1 compared with the usual docker images publicly available on GitHub?


Are you using the nvcr.io/nvidia/caffe:16.11 docker container for this on DGX-1 ?

Exactly, that’s the one I’m trying to run on, and it has different templates etc.

template <typename Ftype, typename Btype>
instead of
and so on.

I believe that should have nvidia caffe 0.14.5 in it:


can you confirm?

are you able to conform your C+±layer to that version of nvidia caffe?

I don’t believe DGX-1 containers are available at this time with a newer version of nvidia caffe

Actually it has nvidia caffe 0.16.0.

Would be better to have 0.14 or 0.15 I believe, since they seem to comply better with the layer.

Are there any containers available with these versions of nvidia caffe?

We had a 0.15.9 container, and did try to run it. However, in this container I cannot find the nvidia caffe folder with its source code. That’s not the case for 0.16, there the source is present and it is possible to recompile nvidia caffe.

You are correct, the 16.11 container contains 0.16.0 nvidia caffe.

Is there a 0.15 container available which includes the source code of nvidia caffe (making it possible to recompile)?