What does this do?

• Test models on your desktop then 'fire and forget' building full models on Deep Learning Amazon AWS spot instances
• Instances are created, used and terminated automatically
• Python script output (model weights and logs) available in S3 bucket after instance termination
• More lightweight than using Ansible, Chef or Terraform
• Save money. AWS deep learning spot instances are 30% cheaper than on-demand instances.

Requirements

• Python 3.6
• Boto3
• awscli

Setup

• Run awscli to set AWS keys (need EC2 and S3 permissions) and region
• Put contents of this repo into root of code to deploy to AWS instance
• Modify parameters in start.sh (mandatory)
• Modify parameters in deploy.py (not mandatory)

Usage and Flow

• Run 'deploy.py' to create spot instance request
• When request is fulfilled a Deep Learning instance is created that runs start.sh on the instance at bootup
• start.sh will pull down your script from GitHub and run it
• After script completes, start.sh will terminate instance
• Script output (e.g. models, weights and logs) folder zipped and pushed to S3 bucket

Notes/Ideas

• Datasets need to be downloaded via the python script specified in start.sh. Downloading massive datasets for every run could incur significant data transfer costs. Support for attaching a persistent volume?