Open aluo-x opened 7 years ago
This would be great for Python 3.6! I'll be checking around and test it the moment you get it on 3.6 :)
I have discovered that batchnorm is broken is the current build. I would advise against using my build until further investigation. A discussion can be found in my repo.
Thanks for the warning, though for the moment I'm using Crhistoph Gohlke's build (Windows, Py3.6, no SSE, no nothing, but works).
I see in the repo that you've been working on this. Here's hoping that you figure the broken things on the 3.6 builds -- I'll test as soon as I you upload, and thanks for the hard work!
I've added a CPU only version w/ AVX2, built using VS 2017 15.4 and Python 3.6.3. So far GPU versions with CUDA 9 and cuDNN 7 are still failing on Windows, but will likely be added with TF 1.5.
Please note AVX2 will only work on Intel Haswell and later, and AMD Excavator & Zen.
just uploaded a whl with CUDA 8 + cudnn 7 + AVX (not 2) support.
You sir, are a gentleman. I'll be checking (and most probably using) those AVX builds!
More details are in the readme. https://github.com/aluo-x/tensorflow_windows