July 19, 2017

Deep Learning on Windows

I had the time and opportunity to take the Part 1 of the Deep Learning Course from fast.ai, taught by Jeremy Howard and Rachel Thomas. While I was able to perform the assignments using free credits from Google Cloud Platform, I was intrigued enough by the course to setup a personal Deep Learning rig.

Unfortunately, I also use Windows when I need to use Visual Studio. So rather than switching between Linux and Windows 10, I decided to setup my Windows 10 for Deep Learning. Luckily I came across this repo which has great instructions on how to setup Deep Learning in Windows 10 for a Python 3 environment. I followed all the steps listed in that repo and was able to run Tensorflow and CNTK successfully. But Theano did not work and it gave me the following error:

ERROR: Failure: GpuArrayException (Error loading library: -1)

After lots of head scratching and googling, I found out that for some reason if you have a GTX 1080 Ti, you need to set a flag for the CUDA driver version so that Theano works.

I just added this environment variable to make Theano work under GTX 1080 Ti:

GPUARRAY_CUDA_VERSION=80