search

BVLC / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
34.01k stars 18.71k forks source link

Are there any plans to release a prebuilt-binary for Windows? #6874

Closed AlanCPSC closed 4 years ago

AlanCPSC commented 4 years ago

Hello everyone,

I was wondering if you guys have any plans to release official prebuilt-binaries for Windows 7/8/10 (+ Mac OS if possible)? I've been working on helping fellow students at school install caffe onto their laptops - mainly through python's anaconda online package registry (https://anaconda.org/search?q=caffe).

Unfortunately this has been an incredibly challenging task, as finding a compatible distribution for everyone has not been easy.

I believe that releasing official prebuilt-binaries will help beginners/students pick up this library quicker.

willyd commented 4 years ago

Hi @AlanCPSC, I am (maybe was would be more appropriate here) responsible for maintaining the Windows branch. I tried to maintain pre-built binaries but I soon realized that it takes a considerable amount of time to create binaries with support for every python version and every cuda version. I created some packages a year ago but I really don't have time to maintain them as CUDA and Python get updated.

Also, it is very hard to compete with tensorflow and PyTorch in terms of support.

I really don't have much time for this (unpaid) work these days. You may try to build your own packages using the recipes I create here.

AlanCPSC commented 4 years ago

Thank you for your response. I appreciate your contribution and development of this library for the rest of us. Some followup questions that I have related to this matter are:

Question 1) I have looked at the recipes you provided, and I am a bit confused on how to exactly utilize them. Do I just run the bld.bat file on a windows machine? Are there any preparatory steps required before building? Any pointers would be great.

Question 2) Regarding the work required to prepare prebuilt-binaries. Is this due to difficulty with the everchanging CUDA specification. Or is this due to a difficulty building the binaries themselves (e.g. some non-trivial aspect that I am not aware of). Perhaps I can be of assistance.

willyd commented 4 years ago

*1) * You need to install conda build. conda install conda-build then cd to the recipes directory and conda build ..

2) The difficulty is more related to me not having the time to test every combination of Python and CUDA that people want to use.