Setting up AWS for Deep Learning

October 04, 2016

This short article will show you how to set up an AWS instance based on Ubuntu Server 14.04 LTS. I found that most tutorials out there were either outdated or required unnecessary steps. So this is my attempt to provide a simple guide to configuring a minimal AWS instance.

We are going to install CUDA v7.5 and cuDNN v5.1, as well as TensorFlow, Theano and Keras.

Creating instance and connecting

We will start with creating a new GPU instance. Use the following AMI:

Ubuntu Server 14.04 LTS (HVM), SSD Volume Type - ami-2d39803a

After we have downloaded the access key, we need to change its privileges. For convenience, we are also going to store the instance's host and username in the environment variable AWS:

chmod 400 aws-california.pem
export AWS=ubuntu@ec2-...

Now you can easily connect to your instance:

ssh -i aws-california.pem $AWS

Installing CUDA and cuDNN

To install CUDA, we will add its official repository, then do a system upgrade and install it along with the Linux sources and some required kernel modules:


sudo dpkg -i cuda-repo-ubuntu1404_7.5-18_amd64.deb

sudo apt-get update
sudo apt-get upgrade -y
sudo apt-get install -y cuda linux-source linux-image-extra-virtual

We need to define the CUDA_HOME environment variable which some frameworks require. Furthermore, we need to amend the user's environment variables for library and binary paths:

echo 'export CUDA_HOME=/usr/local/cuda-7.5
export PATH=${CUDA_HOME}/bin:${PATH}' >> ~/.bashrc

The following commands will install cuDNN:

tar -xzf cudnn-7.5-linux-x64-*.tgz
sudo cp cuda/lib64/* /usr/local/cuda/lib64
sudo cp cuda/include/* /usr/local/cuda/include

Now we can reboot the machine:

sudo reboot

As we did a system upgrade earlier and rebooted the machine, uname -r now indicates the current kernel version. We will install the latest Linux headers and finally load the Nvidia kernel module:

sudo apt-get install linux-headers-$(uname -r)
sudo modprobe nvidia

If your installation was successful, you can run nvidia-smi and it will should show you information about the server's GPU.

TensorFlow, Keras and Theano

We are going to use Python 3's pip to install TensorFlow and other Python libraries. The following command will also install some other dependencies that will be needed in future steps:

sudo apt-get install python3-pip gfortran subversion libblas-dev liblapack-dev libhdf5-dev

Installing TensorFlow is straightforward:

sudo pip3 install --upgrade $TF_BINARY_URL

Unfortunately, the Keras' SciPy dependency does not compile when installed via pip. However, it is possible to install SciPy from Ubuntu's repositories:

sudo apt-get install python3-scipy

Keras has a dependency on Theano, which we are going to install implicitly:

sudo pip3 install keras

Optionally, you may want to install H5py which is needed to serialise Keras models:

sudo pip3 install h5py

Finally, we are going to configure Theano for GPU usage and also enable CNMeM for better performance. Please refer to this article for more information:

echo '[global]
device = gpu
floatX = float32
optimizer = fast_run


cnmem = 1.0

ldflags = -llapack -lblas' > ~/.theanorc


Now your machine should be fully set up and you can start training some models.

If you see any potential for improvements, please feel free to send a pull request by clicking "Edit" below.

Generated with MetaDocs v0.1.2-SNAPSHOT