Bug#822791: ITP: minerva -- deep learning framework

[Paul Wise]

On Wed, Apr 27, 2016 at 10:49 PM, Daniel Stender wrote:

> GPU processing via CUDA (therefore it is going to be a contrib candidate).

If upstream could be convinced to use OpenCL that would be nice.

Otherwise, open source CUDA support is being worked on by the LLVM folks:

http://llvm.org/docs/CompileCudaWithLLVM.html
http://lists.llvm.org/pipermail/llvm-dev/2016-March/096576.html

I guess cuDNN is non-free too though.

-- 
bye,
pabs

https://wiki.debian.org/PaulWise

Bug#822791: ITP: minerva -- deep learning framework

[Daniel Stender]

Package: wnpp
Severity: wishlist
Owner: Daniel Stender 
X-Debbugs-Cc: debian-de...@lists.debian.org

* Package name: minerva
  Version : 0.0~git20150804.7d21ed2
  Upstream Author : Yuptan Li 
* URL : https://github.com/dmlc/minerva
* License : Apache-2.0
  Programming Lang: C++/Python
  Description : deep learning framework

Minerva is a deep learning library with a Numpy like NDarray programming
interface, and Python bindings implemented in Cython. It supports
CPU and (also multi) GPU processing via CUDA (therefore it is going to be
a contrib candidate).

However there are two questions concerning packaging this, there hasn't been
much development in the past months which is a little bit noticeable [1],
and instead of other deep learning frameworks already included into Debian
Minerva *must* have cuDNN [2] available for GPU processing, which isn't yet.

So, I would like to watch the project if it's active and when cuDNN would
become available (which would be highly appreciated for the Caffe and Theano
which could employ this to speed up) this could be a packaging candidate.

Moreinfo when this questions have been solved.

Thanks,
DS

[1] there hasn't been no official stable release (no tags?) but Minerva is 
commonly
mentioned among the "big" deep leaning frameworks (Caffe, Tensorflow, 
Theano etc.), like here:
https://developer.nvidia.com/deep-learning-software.

[2] special Nvidia GPU library for primitives of deep neural networks:
https://developer.nvidia.com/cudnn

-- 
4096R/DF5182C8
http://www.danielstender.com/blog/