!!! FOUND IT !!!
"Using gpu device 0: GeForce GTX TITAN Black (CNMem is disabled, cuDNN=5105)
My Config :
Visual Studio 2015 on Windows 10_64
Nvidia CUDA, updated with CuDNN
Latest Nvidia driver for NVIDIA GTX TITAN BLACK, Intel i7
Nvidia tests went ok (deviceQuery, bandWidthTest), and mandelbrot sample GPU executable compiled with VS2015. And checked the GPU was used with gpuz.
But for Python with GPU… still impossible. Searched the whole web.
After Winpython 3.5 failures I tried Winpython 2.7 and went through the FULL procedure here, this time.
Moving to the TDM-GCC-6-64 compiler (gcc) maybe helped but didn’t solve it all for me.
I checked all my path variables, make a clean “env.bat” environment file.
My solution (REDEMPTION !!) went from modifying the env.bat (described at url http://…/theano/install_windows.html).
SOLUTION with Python 2.7 :
I was about to throw it all when, by inspiration, I updated the VSFORPYTHON value with VS2015 path (instead of “Visual Studio 2008 Compiler for Python 2.7”) and it worked at LAST !
Here’s the new File env.bat (REM lines removed) with my VSFORPYTHON update :
set VSFORPYTHON=“C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC”
CALL %VSFORPYTHON%\vcvarsall.bat amd64
NB. Note the PATH and change it with all your own directories if needed.
I have no idea why I had the problem with Winpython 3.5 (maybe thanks to Masaru, because of the wrong Mingw64 package I had instead of TDM-GCC compiler).
I have no idea why the 2015 Visual Studio solution works for compiling python 2.7 and not initial solution from Theano Website.
Up message if it helps you too…