setup-aws-tensorflow.bash installs the following things on the ec2 g2.2xlarge instance running Ubuntu 14.04:

• Required linux packages
• CUDA 7.5
• cuDNN v4
• Anaconda with Python 2.7
• formats and mounts partition /dev/xvdb as /home/ubuntu/nd013

setup-keras-research.sh installs following things

• TensorFlow 1.0
• GPU usage tool gpustat

Original version was based on the blog post: http://max-likelihood.com/2016/06/18/aws-tensorflow-setup/.

g2.2xlarge costs ~$500/month (really, i mean it), so use it very carefully

• Create aws ec2 instance of ec2 g2.2xlarge instance running Ubuntu 14.04
• On Step4 'Add Storage' set root volume size to 15GB
• On Step4 'Add Storage' add new EBS volume with 140GB size. It will be mounted as /dev/xvdb and we'll store training pictures there. 42GB of zip + unzipped 80GB, so 120GB is just bare minumum to make it work.
• Do not forget to save your .pem file to access to box. let's say you named it 'nd013.pem'
• ssh to your box using

ssh -i nd013.pem ubuntu@your_ec2_box_ip 

• install git

sudo apt-get install git -y

• clone this repo

git clone https://github.com/mungobungo/aws-tensorflow-setup.git

• Run setup_aws_tensorflow.bash on the aws instance:

~/aws-tensorflow-setup/setup_aws_tensorflow.bash

• wait for 5 minutes or so

• Run setup-keras-research.sh on the aws instance:

~/aws-tensorflow-setup/setup-keras-research.sh

• go to nd013 folder

cd ~/nd013

• clone comma.ai repository there

git clone https://github.com/commaai/research.git

• download training data

cd research
./get_data.sh

• wait several hours until training data is downloaded

• have fun )