python 3 builds

[djsutherland]

Is this just to avoid recompiling the library three times? #31, #33, and me earlier today all inadvertently installed caffe from defaults and thought it was the conda-forge version.

[jakirkham]

I think I was unclear at the time as to whether Python 3 was actually supported. There may also have been CI time limit issues as well. Now that we support matrix builds, it should certainly be possible to enable Python 3 builds. Just a question about whether upstream really supports them.

[djsutherland]

Apparently they didn't as of 2014, but maybe it does now if you build with cmake: BVLC/caffe#4015 (comment) and these instructions. The Makefile definitely assumes 2.7 if you don't tell it otherwise: this line.

[djsutherland]

Added python 3 builds to #35.

[jakirkham]

Would add it seems like they are still fixing things on the Python 3 front. PR ( BVLC/caffe#5900 ) went in 5 days ago and is not in a release. Though they do appear to be running a CI build for Python 3, their coverage is much better on Python 2. We can certainly build for Python 3. Am just not going to spend lots of time backporting patches from upstream if Python 3 support turns out to be really buggy.

[djsutherland]

Hmm, okay.

Since python 3 builds are also having boost-python issues for me anyway, I'll split #35 into the version bump and a separate python 3 PR so we can get 1.0 released for py27 sooner.

[jakirkham]

Is there still a PR for Python 3 somewhere? I think I missed it if there is.

[djsutherland]

#35 still contains the amount of work I did on it, which was just 028b921, but skips the python 3 builds.

[jakirkham]

Might be worth trying this again. Seems like the number of open issues for Python 3 is much smaller (only 3 when I last checked) and mostly seem related to GPU or Windows, which we don't support yet for better or worse.

[marcelotrevisani]

Should we close this issue?