I recently started doing the fast.ai course where you follow a top-down approach into the world of machine learning.

The first thing you need to do is have access to a Nvidia GPU, and in the first video lesson they show the setup of a cloud instance that runs Tensorflow using either crestle or paperspace.

But since either of them are paid by the hour, and considering I already have Nvidia GPU in my desktop (albeit not as powerful as the ones offered in the aforementioned services, just an GTX 1070), I thought of going in a tangent and setting it up myself.

The first thing to do is install Tensorflow.

I already had CUDA installed, so I proceeded to install Tensorflow.

There was a problem: the CUDA version was too bleeding edge for the Tensroflow version. Tensorflow 1.9 requires CUDA 9.0 and the latest CUDA in Arch repositories was 9.1.

I thought of compiling Tensorflow myself, since that would make it work with 9.1. But that would take a long time and probably space as well.

Instead I opted out for a docker image which required nvidia-docker from AUR, and rebooting the docker-service. Then I followed the instructions found in the github repository to run the image after it pulled it. Afterwards, I started a Jupyter Notebook that required a token to login, which was specified in the output of the jupyter command line prompt.

Once inside the jupyter notebook, I was finally able to run Tensroflow with GPU acceleration.

Simple enough if you know where to look. But even if digging up took some time, is time well spent considering the speedup achieved with GPU acceleration when using Tensorflow.

Last modified on 2018-05-11

You can make sure that the author wrote this post by copy-pasting this signature into this Keybase page.