Hi Science and Med Team, I maintain lots of packages related to deep learning and machine learning, but I think I'm reaching an upperbound. So I'm clearly stating my future plans about deep learning packages for Debian, lest potential contributors hesitate to step in and help.
Med team has been added in the recipient list due to their increasing interest in deep/machine learning, and the coronavirus related works. [[[ DL frameworks ]]] My emergy only allows me to cover one modern deep learning framework -- pytorch. I'm very likely unable to help the tensorflow maintenance once bazel gets ready in our archive. I can keep maintaining some tensorflow dependencies since some of them are quite stable. -- I'll only maintain caffe and pytorch. Potential contributors should feel free to deal with any other DL frameworks. -- I can provide suggestions, but I'll not provide code. [[[ CUDA flavor of DL frameworks ]]] Currently I only have the plan to do the cpu version of pytorch (and caffe). The cuda version of caffe had been removed by me because I hate dealing with cuda related stuff in the debian archive. -- I'll not package the most important CUDA library -- cuDNN -- for deep learning frameworks. We need other volunteers for this. -- If someone is willing to maintain cuDNN, I can try to package the cuda version of pytorch [[[ ROCm flavor of DL frameworks ]]] I still don't know how on earth can the ROCm software stack work with the opensource `amdkfd` kernel driver (debian has already enabled that driver for our kernel packages) instead of the proprietary version of `amdkfd`. -- I'm willing to maintain the ROCm software stack as long as it is able to work on a Debian system without non-free components. There is a "ROCm Team" on salsa. -- the answer to the above question is the only blocker for me to make further progress about ROCm. [[[ upper layer applications ]]] -- I'll not maintain any upper layer applications built upon any deep learning framework. -- But I'll maintain ML-Policy to guide the maintenance of these packages. -- I may maintain some toy dataset packages for testing DL framework sanity e.g. src:dataset-fashion-mnist (already in testing and sid) --- These are all things I can do, and I'm not a monoplic maintainer. Volunteers are always welcome, and helps are always appreciated!