cuda/tensorflow-gpu installation

[aparrish]

I was able to get tensorflow-gpu working simply by ensuring that

module load cudnn/8.0v6.0
module load cuda/8.0.44

was in my slurm script and installing the tensorflow-gpu package with conda (conda install tensorflow-gpu). Are you sure that you need to download all of the stuff as indicated in the "Install tensorflow-gpu" section?

[cvalenzuela]

nop! I'm not sure. maybe we don't need it.
Maybe is already installed in the cluster. I'll check the documentation

[letspresto]

I went to the Slurm session and found this helpful. Slurm allows you to load, manage multiple versions and configurations of software packages. Some helpful commands:

#see available package environments
module avail
#see what is currently loaded 
module list
#remove all packages 
module purge

Just add the modules to your bash script after you configure which nodes, how much memory, etc. More info here which is a little more verbose

[cvalenzuela]

thanks @letspresto!, I'll add that to the README.md

[coblezc]

You can also use module show [package name] to get some helpful info about the package, e.g.

module show cudnn/8.0v6.0

or a better example is

module show torch/gnu/20170504

The things that begin "load(...)" are the dependencies that are also loaded when you load a package.

[cvalenzuela]

Nice! This means that the tf-gpu installation part is not necessary, I'll update the readme.

(@agermanidis, you could find this helpful)