TensorFlow the Hard Way

I call myself a Data Science Architect at work, and that is a pretty accurate description of what I do. I’m really more of a plumber who aspires to do Data Science though.

TensorFlow is something I’ve been looking at for some time now, and I’d really like to become a competent TensorFlow application/model/network developer. One thing that has slowed me down is my need to make it run well on my machine.

TensorFlow is made for GPU acceleration, and I’m fortunate to possess a VR-enabled machine with a nice graphics card. It would be a shame to simply install the software version of TensorFlow and run with that. On top of this, I have a newer version of Linux on my machine that I don’t want to down-level just so I can install from the packages that Google provides.

Google and NVIDIA deserve a lot of credit for some very good documentation, and a well-structured set of set of libraries and interfaces. This is an incredibly complex environment, and it’s very satisfying to see everything work together.

There were lots of problems to solve along the way, and all of the details about how to build everything from scratch are at my github repo.

Next step here is to learn TensorFlow development and compare results between the accelerated and software implementations.