Economical AWS Setup

[Jesse Vander Does]

Does anyone have suggestions of tools or methods for making the most of GPU hours on AWS? I would like to minimize setup time, such as time spent copying files, and make sure the instance goes down after the model has reached some error threshold.

I like the idea of using docker to run training sessions, but I don't know of any tools or services that allow you to stage and run one of AWS tasks.

Any tips would be greatly appreciated. I've enjoyed experimenting with Caffe, but would enjoy it more if I wasn't so terrified of accidentally wracking up a massive AWS bill.

[Traun Leyden]

If you use AWS spot instances, you will save a lot of money.  Of course the downside is that your instance could be taken from you at any moment.

I have a blog post about running Caffe under Docker on AWS GPU instances here:

    http://tleyden.github.io/blog/2014/10/25/running-caffe-on-aws-gpu-instance-via-docker/

and a work-in-progress project which attempts to wrap up Caffe behind a REST API where everything runs under Docker:

    https://github.com/tleyden/elastic-thought

(if there are other Caffe + Docker projects you run across, let me know and I'll update my github/dockerhub repos with links)