Tyler Hallada - Blog

I found the tensorflow documentation rather lacking for installation instructions, especially in regards to getting GPU support. I’m going to write down my notes from wrangling with the installation here for future reference and hopefully this helps someone else too.

This will invariably go out-of-date at some point, so be mindful of the publish date of this post. Make sure to cross-reference other documentation that has more up-to-date information.

Assumptions

These instructions are very specific to my environment, so this is what I am assuming:

1. You are running Ubuntu 16.04. (I have 16.04.1)
   • You can check this in the output of uname -a
2. You have a 64 bit machine.
   • You can check this with uname -m. (should say x86_64)
3. You have an NVIDIA GPU that has CUDA Compute Capability 3.0 or higher. NVIDIA documentation has a full table of cards and their Compute Capabilities. (I have a GeForce GTX 980 Ti)
   • You can check what card you have in Settings > Details under the label “Graphics”
   • You can also check by verifying there is any output when you run lspci | grep -i nvidia
4. You have a linux kernel version 4.4.0 or higher. (I have 4.8.0)
   • You can check this by running uname -r
5. You have gcc version 5.3.1 or higher installed. (I have 5.4.0)
   • You can check this by running gcc --version
6. You have the latest proprietary NVIDIA drivers installed.
   • You can check this and install it if you haven’t in the “Additional Drivers” tab in the “Software & Updates” application (update-manager). (I have version 375.66 installed)
7. You have the kernel headers installed.
   • Just run sudo apt-get install linux-headers-$(uname -r) to install them if you don’t have them installed already.
8. You have Python installed. The exact version shouldn’t matter, but for the rest of this post I’m going to assume you have python3 installed.
   • You can install python3 by running sudo apt-get install python3. This will install Python 3.5.
   • Bonus points: you can install Python 3.6 by following this answer, but Python 3.5 should be fine.

Install the CUDA Toolkit 8.0

NVIDIA has a big scary documentation page on this, but I will summarize the only the parts you need to know here.

Go to the CUDA Toolkit Download page. Click Linux > x86_64 > Ubuntu > 16.04 > deb (network).

Click download and then follow the instructions, copied here:

1. sudo dpkg -i cuda-repo-ubuntu1604_8.0.61-1_amd64.deb
2. sudo apt-get update
3. sudo apt-get install cuda

This will install CUDA 8.0. It installed it to the directory /usr/local/cuda-8.0/ on my machine.

There are some post-install actions we must follow:

1. Edit your ~/.bashrc
   • Use your favorite editor gedit ~/.bashrc, nano ~/.bashrc, vim ~/.bashrc, whatever.
2. Add the following lines to the end of the file:

   # CUDA 8.0 (nvidia) paths
   export CUDA_HOME=/usr/local/cuda-8.0
   export PATH=/usr/local/cuda-8.0/bin${PATH:+:${PATH}}
   export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
   

3. Save and exit.
4. Run source ~/.bashrc.
5. Install writable samples by running the script cuda-install-samples-8.0.sh ~/.
   • If the script cannot be found, the above steps didn’t work :(
   • I don’t actually know if the samples are absolutely required for what I’m using CUDA for, but it’s recommended according to NVIDIA, and compiling them will output a nifty deviceQuery binary which can be ran to test if everything is working properly.
6. Make sure nvcc -V outputs something.
   • If an error, the above steps 1-4 didn’t work :(
7. cd ~/NVIDIA_CUDA-8.0_Samples, cross your fingers, and run make
   • The compile will take a while
   • My compile actually errored near the end with an error about /usr/bin/ld: cannot find -lnvcuvid I think that doesn’t really matter because the binary files were still output.
8. Try running ~/NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/deviceQuery to see if you get any output. Hopefully you will see your GPU listed.

Install cuDNN v5.1

This AskUbuntu answer has good instructions. Here are the instructions specific to this set-up:

1. Visit the NVIDIA cuDNN page and click “Download”.
2. Join the program and fill out the survey.
3. Agree to the terms of service.
4. Click the link for “Download cuDNN v5.1 (Jan 20, 2017), for CUDA 8.0”
5. Download the “cuDNN v5.1 Library for Linux” (3rd link from the top).
6. Untar the downloaded file. E.g.:

   cd ~/Downloads
   tar -xvf cudnn-8.0-linux-x64-v5.1.tgz
   

7. Install the cuDNN files to the CUDA folder:

   cd cuda
   sudo cp -P include/* /usr/local/cuda-8.0/include/
   sudo cp -P lib64/* /usr/local/cuda-8.0/lib64/
   sudo chmod a+r /usr/local/cuda-8.0/lib64/libcudnn*
   

Install libcupti-dev

This one is simple. Just run:

sudo apt-get install libcupti-dev

Create a Virtualenv

I recommend using virtualenvwrapper to create the tensorflow virtualenv, but the TensorFlow docs still have instructions to create the virtualenv manually.

1. Install virtualenvwrapper. Make sure to add the required lines to your ~/.bashrc.
2. Create the virtualenv:

   mkvirtualenv --python=python3 tensorflow
   

Install the TensorFlow with GPU support

If you just run pip install tensorflow you will not get GPU support. To install the correct version you will have to install from a particular url. Here is the install command you will have to run to install TensorFlow 1.2 for Python 3.5 with GPU support:

pip install https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.2.0-cp35-cp35m-linux_x86_64.whl

If you need a different version of TensorFlow, you can edit the version number in the URL. Same with the Python version (change cp35 to cp36 to install for Python 3.6 instead, for example).

Test that the installation worked

Save this script from the TensorFlow tutorials to a file called test_gpu.py:

# Creates a graph.
with tf.device('/cpu:0'):
  a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
  b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
c = tf.matmul(a, b)
# Creates a session with log_device_placement set to True.
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
# Runs the op.
print(sess.run(c))

And then run it:

python test_gpu.py

You should see your GPU card listed under “Device mapping:” and that each task in the compute graph is assigned to gpu:0.

If you see “Device mapping: no known devices” then something went wrong and TensorFlow cannot access your GPU.