Python + TensorFlow + GPU + CUDA + CUDNN setup with Ubuntu

Every time I setup Python + TensorFlow on a new machine with a fresh Ubuntu install, I have to spend some time again and again on this topic, and do some trial and error (yes I'm speaking about such issues). So here is a little HOWTO, once for all.

Important fact: we need to install the specific version number of CUDA and CUDNN relative to a particular version of TensorFlow, otherwise it will fail, with errors like libcudnn.so.7: cannot open shared object file: No such file or directory.

For example, for TensorFlow 2.3, we have to use CUDA 10.1 and CUDNN 7.6 (see here).

Here is how to install on a Ubuntu 18.04:

pip3 install --upgrade pip   # it was mandatory to upgrade for me
pip3 install keras tensorflow==2.3.0

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-ubuntu1804.pin
sudo mv cuda-ubuntu1804.pin /etc/apt/preferences.d/cuda-repository-pin-600
sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub
sudo add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/ /"
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt-get update
sudo apt install cuda-10-1 nvidia-driver-430

To test if the NVIDIA driver is properly installed, you can run nvidia-smi (I noticed a reboot was necessary).

Then download "Download cuDNN v7.6.5 (November 5th, 2019), for CUDA 10.1" on https://developer.nvidia.com/rdp/cudnn-archive (you need to create an account there), and then:

sudo dpkg -i libcudnn7_7.6.5.32-1+cuda10.1_amd64.deb     

That's it! Reboot the computer, launch Python 3 and do:

import tensorflow
tensorflow.test.gpu_device_name()     # also, tensorflow.test.is_gpu_available() should give True

The last line should display the right GPU device name. If you get an empty string instead, it means your GPU isn't used by TensorFlow!

Notes:

• Initially the installation of CUDA 10.1 failed with errors like:

  The following packages have unmet dependencies:
   cuda-10-1 : Depends: cuda-toolkit-10-1 (>= 10.1.243) but it is not going to be installed

  on a fresh Xubuntu 18.04.5 install. Trying to install cuda-toolkit-10-1 manually led to other similar errors. Using a sources.list from Xubuntu 18.04 like this one helped.

• I also once had errors like Could not load dynamic library 'libcublas.so.10'; dlerror: libcublas.so.10: cannot open shared object file: No such file or directory. After searching this file on the filesystem, I noticed it was found in /usr/local/cuda-10.2/... whereas I never installed the 10.2 version, strange! Solution given in this post: sudo apt install --reinstall libcublas10=10.2.1.243-1 libcublas-dev=10.2.1.243-1. IIRC, these 2 issues weren't present when I used a Xubuntu 18.04, could the fact I used 18.04.5 be the reason?

• Nothing really related, but when installing Xubuntu 20.04.2.0, memtest86, which is quite useful to test the integrity of the hardware before launching long computations, did not work.