This is a tutorial to help you get the Caffe deep learning framework up and running on a GPU-powered AWS instance running inside a Docker container.

Architecture

architecture diagram

Setup host

Before you can start your docker container, you will need to go deeper down the rabbit hole.

You’ll first need to complete the steps here:

Setting up an Ubuntu 14.04 box running on a GPU-enabled AWS instance

After you’re done, you’ll end up with a host OS with the following properties:

• A GPU enabled AWS instance running Ubuntu 14.04
• Nvidia kernel module
• Nvidia device drivers
• CUDA 6.5 installed and verified

Install Docker

Once your host OS is setup, you’re ready to install docker. (version 1.3 at the time of this writing)

Setup the key for the docker repo:

1

$ sudo apt-key adv --keyserver hkp://keyserver.ubuntu.com:80 --recv-keys 36A1D7869245C8950F966E92D8576A8BA88D21E9

Add the docker repo:

1
2

$ sudo sh -c "echo deb https://get.docker.com/ubuntu docker main > /etc/apt/sources.list.d/docker.list"
$ sudo apt-get update

Install docker:

1

$ sudo apt-get install lxc-docker

Run the docker container

Find your nvidia devices

1

$ ls -la /dev | grep nvidia

You should see:

1
2
3

crw-rw-rw- 1 root root 195, 0 Oct 25 19:37 nvidia0
crw-rw-rw- 1 root root 195, 255 Oct 25 19:37 nvidiactl
crw-rw-rw- 1 root root 251, 0 Oct 25 19:37 nvidia-uvm

You’ll have to adapt the DOCKER_NVIDIA_DEVICES variable below to match your particular devices.

Here’s how to start the docker container:

1
2

$ DOCKER_NVIDIA_DEVICES="--device /dev/nvidia0:/dev/nvidia0 --device /dev/nvidiactl:/dev/nvidiactl --device /dev/nvidia-uvm:/dev/nvidia-uvm"
$ sudo docker run -ti $DOCKER_NVIDIA_DEVICES tleyden5iwx/caffe-gpu-master /bin/bash

It’s a large docker image, so this might take a few minutes, depending on your network connection.

Run caffe test suite

After the above docker run command completes, your shell will now be inside a docker container that has Caffe installed.

You’ll want run the Caffe test suite and make sure it passes. This will validate your environment, including your GPU drivers.

1
2

$ cd /opt/caffe
$ make test && make runtest

Expected Result: ... [ PASSED ] 838 tests.

Run the MNIST LeNet example

A more comprehensive way to verify your environment is to train the MNIST LeNet example:

1
2
3
4
5

$ cd /opt/caffe/data/mnist
$ ./get_mnist.sh
$ cd /opt/caffe
$ ./examples/mnist/create_mnist.sh
$ ./examples/mnist/train_lenet.sh

This will take a few minutes.

Expected output:

1
2
3
4
5

libdc1394 error: Failed to initialize libdc1394 
I1018 17:02:23.552733 66 caffe.cpp:90] Starting Optimization 
I1018 17:02:23.553583 66 solver.cpp:32] Initializing solver from parameters:
... lots of output ...
I1018 17:17:58.684598 66 caffe.cpp:102] Optimization Done.

Congratulations, you’ve got GPU-powered Caffe running in a docker container — celebrate with a cup of Philz!

References

• tleyden5iwx/caffe-gpu-master Caffe Docker image (GPU)
• tleyden5iwx/caffe-cpu-master Caffe Docker image (CPU-only)
• Docker on AWS GPU Ubuntu 14.04 / CUDA 6.5
• CUDA 6.5 on AWS GPU Instance Running Ubuntu 14.04