Using Amazon Web Services (AWS) GPU for Deep Learning

CUDA, cuDNN, TensorFlow, Caffe, and other useful packages for Ubuntu 16.04

Useful Packages and Dependencies

# installs essential packages through apt-get and pip
bash aws_setup.sh

Install CUDA 8.0

bash install_cuda_8.sh

Install cuDNN 6.0 for CUDA 8.0 (manual steps)

• Make an account on NVIDIA and download cuDNN.
• After the survey, select "cuDNN v6.0, for CUDA 8.0".
• Download "cuDNN v6.0 Library for Linux" (should be a tar file).
• Transfer (you can use scp) the .tgz file from your local computer to your EC2 instance.
• Install cuDNN from your ec2 instance with bash install_cuDNN.sh.

Install TensorFlow and Monitoring Tools for Python 2.7

pip install tensorflow-gpu
bash install_monitoring_tools.sh

Install Caffe

git clone https://github.com/BVLC/caffe.git
cd caffe
cp Makefile.config.example Makefile.config

Manual Step: edit Makefile.config to set USE_CUDNN := 1.

bash caffe_dependencies.sh
make all
make pycaffe
make test
make runtest

Change the path name and add the following line to your .bashrc file:

export PYTHONPATH="${PYTHONPATH}:/path/to/caffe/python"

Source your .bashrc file.

source ~/.bashrc