John Ramey Statistics and Machine Learning

Installing TensorFlow on an AWS EC2 Instance with GPU Support

The following post describes how to install TensorFlow 0.6 on an Amazon EC2 Instance with GPU Support. I also created a Public AMI (ami-e191b38b) with the resulting setup. Feel free to use it.

UPDATED (28 Jan 2016): The latest TensorFlow build requires Bazel 0.1.4. Post now reflects this. Thanks to Jim Simpson for his assistance.

UPDATED (28 Jan 2016): The AMI provided now exports env variables in ~/.bashrc.

The following things are installed:

  • Essentials
  • Cuda Toolkit 7.0
  • cuDNN Toolkit 6.5
  • Bazel 0.1.4 (Java 8 is a dependency)
  • TensorFlow 0.6

To get going, I recommend requesting a spot instance. Can your instance go away? Sure. But $0.07/hr is much nicer than $0.65/hr when you are figuring things out. I launched a single g2.2xlarge instance using the Ubuntu Server 14.04 LTS AMI.

After launching your instance, install the essentials:

sudo apt-get update
sudo apt-get upgrade
sudo apt-get install -y build-essential git python-pip libfreetype6-dev libxft-dev libncurses-dev libopenblas-dev gfortran python-matplotlib libblas-dev liblapack-dev libatlas-base-dev python-dev python-pydot linux-headers-generic linux-image-extra-virtual unzip python-numpy swig python-pandas python-sklearn unzip wget pkg-config zip g++ zlib1g-dev
sudo pip install -U pip

TensorFlow requires installing CUDA Toolkit 7.0. To do this, run:

sudo dpkg -i cuda-repo-ubuntu1410_7.0-28_amd64.deb
rm cuda-repo-ubuntu1410_7.0-28_amd64.deb
sudo apt-get update
sudo apt-get install -y cuda

At some point, you get the following message: Reboot your computer and verify that the NVIDIA graphics driver can be loaded. I mean, it’s 2016. But whatevs. We’ll reboot in a moment. Now, we need to download cuDNN from Nvidia’s site.

After filling out an annoying questionnaire, you’ll download a file named cudnn-6.5-linux-x64-v2.tgz. You need to transfer it to your EC2 instance: I did this by adding it to my Dropbox folder and using wget to upload it. Once you have uploaded it to your home directory, run the following:

tar -zxf cudnn-6.5-linux-x64-v2.tgz && rm cudnn-6.5-linux-x64-v2.tgz
sudo cp -R cudnn-6.5-linux-x64-v2/lib* /usr/local/cuda/lib64/
sudo cp cudnn-6.5-linux-x64-v2/cudnn.h /usr/local/cuda/include/

Okay, now reboot:

sudo reboot

Next up, we’ll add some environment variables. You may wish to add these to your ~/.bashrc.

export CUDA_HOME=/usr/local/cuda
export CUDA_ROOT=/usr/local/cuda
export PATH=$PATH:$CUDA_ROOT/bin

Getting closer. We need to install Bazel 0.1.4, which requires Java 8. For more details, see this comment.

Install Java 8 first.

sudo add-apt-repository -y ppa:webupd8team/java
sudo apt-get update
# Hack to silently agree license agreement
echo debconf shared/accepted-oracle-license-v1-1 select true | sudo debconf-set-selections
echo debconf shared/accepted-oracle-license-v1-1 seen true | sudo debconf-set-selections
sudo apt-get install -y oracle-java8-installer

Now for Bazel. (Thanks to Jim Simpson for this block.)

sudo apt-get install pkg-config zip g++ zlib1g-dev
chmod +x
./ --user

Okay, almost done. Let’s clone the TensorFlow repo and initialize all submodules using their default settings.

git clone --recurse-submodules
cd tensorflow

Finally, we are going to build TensorFlow with GPU support using CUDA version 3.0 (currently required on AWS) via the unofficial settings.


When you see the following message, type 3.0 to use CUDA version 3.0:

Please specify a list of comma-separated Cuda compute capabilities you want to build with.
You can find the compute capability of your device at:
Please note that each additional compute capability significantly increases your build time and binary size.
[Default is: "3.5,5.2"]: 3.0

If you forget to type 3.0, you’ll get the following error later on:

Ignoring gpu device (device: 0, name: GRID K520, pci bus id: 0000:00:03.0) with Cuda compute capability 3.0. The minimum required Cuda capability is 3.5.

Other than that, I went with all the default options, resulting in the nice message:

WARNING: You are configuring unofficial settings in TensorFlow. Because some external libraries are not backward compatible, these settings are largely untested and unsupported.

Pffft. Anyway, last steps. These take quite a while (~24 minutes for me).

bazel build -c opt --config=cuda //tensorflow/cc:tutorials_example_trainer
bazel build -c opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
sudo pip install --upgrade /tmp/tensorflow_pkg/tensorflow-0.6.0-cp27-none-linux_x86_64.whl

Congrats! TensorFlow is installed. At this point, if you launch Python and run the following code, you’ll see a lot of nice messages indicating your GPU is set up properly:

import tensorflow as tf
tf_session = tf.Session()
x = tf.constant(1)
y = tf.constant(1) + y)

You can also check that TensorFlow is working by training a CNN on the MNIST data set.

python ~/tensorflow/tensorflow/models/image/mnist/

# Lots of output followed by GPU-related things...
I tensorflow/stream_executor/cuda/] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
I tensorflow/core/common_runtime/gpu/] Found device 0 with properties:
name: GRID K520
major: 3 minor: 0 memoryClockRate (GHz) 0.797
pciBusID 0000:00:03.0
Total memory: 4.00GiB
Free memory: 3.95GiB
I tensorflow/core/common_runtime/gpu/] DMA: 0
I tensorflow/core/common_runtime/gpu/] 0:   Y
I tensorflow/core/common_runtime/gpu/] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GRID K520, pci bus id: 0000:00:03.0)
I tensorflow/core/common_runtime/gpu/] Allocating 3.66GiB bytes.
I tensorflow/core/common_runtime/gpu/] GPU 0 memory begins at 0x7023e0000 extends to 0x7ec556000
I tensorflow/core/common_runtime/gpu/] Creating bin of max chunk size 1.0KiB
I tensorflow/core/common_runtime/gpu/] Creating bin of max chunk size 2.0KiB
Epoch 0.00
Minibatch loss: 12.053, learning rate: 0.010000
Minibatch error: 90.6%
Validation error: 84.6%
Epoch 0.12
Minibatch loss: 3.282, learning rate: 0.010000
Minibatch error: 6.2%
Validation error: 6.9%
Epoch 0.23
Minibatch loss: 3.466, learning rate: 0.010000
Minibatch error: 12.5%
Validation error: 3.7%
Epoch 0.35
Minibatch loss: 3.191, learning rate: 0.010000
Minibatch error: 7.8%
Validation error: 3.4%
Epoch 0.47
Minibatch loss: 3.201, learning rate: 0.010000
Minibatch error: 4.7%
Validation error: 2.7%

I borrowed instructions from a few sources, so thanks very much to them. If you want more information about the various options, check out TensorFlow’s installation instructions.