Theano on Windows with Python 3.5 probably don't work

[nouiz]

see: http://stevedower.id.au/blog/building-for-python-3-5/

Maybe we should switch to MSVC directly and not mingw? Or just wait to have mingw support it?

[patrikhuber]

That's a very nice write-up! I haven't gone through the whole blog post, but I just wanted to add that I'm successfully (modulo a few minor issues) using Theano with msys2-mingw64 and python3.5. I've installed their mingw64 python3.5, numpy, scipy, pip3 from pacman, and then installed Theano from pip3.

Does mingw64 python work with CUDA?

[patrikhuber]

@LexSong: I don't know, I'm not using CUDA currently, but msys2's Python has been really great so far and it looks pretty standard. Without having any experience with Python and CUDA, I'd say "Probably, why not".

[nouiz]

"Probably, why not" on windows with CUDA make me smile. I hope it is now as
easy is it could be, history told me it isn't as easy as we would like. But
someone need to try it to know. Here we won't have time shortly to try it,
so if someone try it, please report result. This can help other people.

On Sun, Dec 20, 2015 at 4:18 PM, Patrik Huber notifications@github.com
wrote:

@LexSong https://github.com/LexSong: I don't know, I'm not using CUDA
currently, but msys2's Python has been really great so far and it looks
pretty standard. Without having any experience with it, I'd say "Probably,
why not".

—
Reply to this email directly or view it on GitHub
#3376 (comment).

[patrikhuber]

Hehe yea, I can only imagine! I currently can't try with CUDA, maybe in the future. Hopefully other people can chime in :-) It would be great if it worked with msys2-mingw64, since msys2 is awesome and useful for many purposes, so you don't have to have several mingw distributions and each of them comes with their own gcc and takes up space etc. (e.g. most Python distributions for Windows come with their own mingw).

[ErnstDinkelmann]

Something interesting that may just help others.
My environment:
Win8.1, 64bit, Python 3.5, Anaconda4 install, MS Visual Studio 12.0, CUDA 7.5, OpenBLAS

I also experienced the C/C++ compiler and importing issues when I had
cxx = C:\TDM-GCC-64\bin\g++.exe
This is recommended in a number of installation tutorials online.

The errors I experienced:
File "C:\Program Files (x86)\JetBrains\PyCharm Community Edition 2016.1.2\helpers\pydev_pydev_bundle\pydev_import_hook.py", line 21, in do_import
module = self._system_import(name, _args, *_kwargs)
ImportError: DLL load failed: A dynamic link library (DLL) initialization routine failed.

However, when I changed to:
cxx = C:\MinGw64\mingw64\bin\g++.exe
Not sure I've seen this recommended, The g++ files are significantly different in size, so they must be a lot different in content.
Everything seems to work, at least for the scripts I'm running on scikit-neuralnetwork (sknn)

For completeness, my entire .theanorc.txt file content (stored in c:\Users\ is

[global]
floatX = float32
device = gpu
base_compiledir = C:\TheanoOperational
cxx = C:\MinGw64\mingw64\bin\g++.exe

[nvcc]
flags = -LC:\Anaconda3\libs
fastmath = True
compiler_bindir = C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin\cl.exe

[cuda]
root = C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5

[blas]
ldflags=-LC:\openblas -lopenblas

Important to note, even with correctly pre-compiled files in the base_compiledir done with the second g++.exe, the initial g++.exe was unable to import them correctly.

Some interest results to show that it does work in fact:
device=gpu, cxx = C:\MinGw64\mingw64\bin\g++.exe
Time: 120 secs

device=cpu, cxx =
(ie no g++ compiler)
Time: 80 secs

device=cpu, cxx = C:\MinGw64\mingw64\bin\g++.exe
Time: 40 secs

The GPU result just means I have a slow GPU compared to my CPU for this specific type of task, which is true (Corei7-2670QM vs Nvidia NVS 4200).
However, when only looking at the CPU results, hell yeah. And that's with a relatively large amount of overhead on a small run.

Hope this can help someone. May just be a fluke, but it works.

[mingwandroid]

@ErnstDinkelmann is C:\MinGw64 an MSYS2 installation?