Install Tensorflow GPU 1.13 on Pop OS 18.04

Originally published at arandilopez.dev. Read it in spanish too.

Pop!_OS is an incredible Linux distribution based on Ubuntu. But what's more amazing about Pop!_OS is its support for the Nvidia and Tensorflow graphics cards, which is a great feature to opt for Pop!_OS as a distro. In this post I will show the steps to get Tensorflow GPU 1.13 with python3.

Pop!_OS 18.04 supports the graphics card with an ISO image that has the most current proprietary drivers that System76 maintains. The most current driver is version 418, you can check the version installed on your OS with the command nvidia-smi.

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.56       Driver Version: 418.56                              |
|-------------------------------+----------------------+----------------------+

Pop!_OS has a metapackage to natively install Tensorflow and CUDA, however this metapackage installs up to version 1.9 of Tensorflow.

Configure Python 3

First we install the necessary packages to start our development with python3, we will install pip and venv to manage the dependencies of our python projects.

$ sudo apt install python3-dev python3-pip python3-venv

Install CUDA 10.0

To install version 1.13, the last stable of Tensorflow, we need to have the Nvidia 418 drivers and install the following dependencies for CUDA and Tensorflow.

$ sudo apt install autogen autogen-doc cmake cmake-data guile-2.0-libs libblas3 libgc1c2 libgfortran4 libjsoncpp1 liblapack3 libnvidia-compute-418 libopts25 libopts25-dev librhash0 libuv1 python3-numpy

Now install the necessary CUDA packages.

$ sudo apt install system76-cuda-10.0 system76-cudnn-10.0

There are CUDA packages with version 10.1, however they do not have support for Tensorflow 1.13. See here.

Install Tensorflow GPU

We will create a new folder for our project

$ mkdir ~/test_tensorflow_gpu
$ cd test_tensorflow_gpu

We started a new virtual environment for our dependencies to this project.

$ python3 -m venv venv
$ source venv/bin/activate

Having activated the virtual environment already allows us to run both python3 andpip3 as python and pip respectively.

Install Tensorflow. The correct package to install Tensorflow with CUDA support is tensorflow-gpu.

(venv)
$ pip install --upgrade tensorflow-gpu

Test the installation of Tensorflow.

(venv)
$ python -c "import tensorflow as tf; print(tf.test.is_built_with_cuda());"
True # <-- expected output