This is a script for creating and managing GPU instances on AWS. It can setup Go and unixpickle/cuda automatically. With this tool, it should be painless to train anynet models on AWS.
Most importantly, you must setup the aws CLI.
You will want to use an AMI with CUDA support. To do this, checkout NVIDIA CUDA Toolkit 7.5 on Amazon Linux. If you choose the "Manual Launch" option, the AMI is free. Make sure you note the AMI ID for the region you intend to use.
In order to use p2.xlarge instances on AWS, you have to request a limit increase. To do this, go to the EC2 management console and then click on "Limits".
You may want to add gpu-aws to a directory in your PATH so that you get an actual gpu-aws command.
First, you'll want to create a new session. A session manages a single EC2 instance. Decide on a session path (a directory which will be created by gpu-aws), the AWS instance type (e.g. t2.nano), and the AMI (e.g. ami-4191b524). Now you can create a session like so:
$ export SESSDIR=~/my_session
$ gpu-aws create $SESSDIR t2.nano ami-4191b524
Creating key pair...
Checking security group...
Security group already exists.
Creating t2.nano from AMI ami-4191b524...
Getting instance IP...
IP: 123.123.123.123
The new instance is now running. You will probably want to configure Go, git, and other important tools:
$ gpu-aws setup | gpu-aws ssh $SESSDIR ec2-user
You can SSH into the instance like so:
$ gpu-aws ssh $SESSDIR ec2-user
When you are done with the instance, you can terminate it like so:
$ gpu-aws destroy $SESSDIR
$ rm -r $SESSDIR