Latest update: Sat Jan 11 20:15:25 PST 2020

For Ubuntu 18.04 LTS + TF 2.1+

Simply look at the tf_2.1_Ubuntu18.sh file. It should install cuda 10.1 and tensorflow 2.1 (or attempting the latest version)

Then run the test python script test_gpu_tf2.1.py to see if tensorflow is detecting the GPU.

(For Ubuntu 16.04 LTS)

Tools/Versions To be installed: (Latest as of July 2018)

• Python 3.5.2 + pip3

• Nvidia 396 Driver

• CUDA 9.0

• cuDNN 7.1

• Tensorflow 1.8

• (Keras is already part of tensorflow @ 1.8+)

• vim (just because we love vim and... you know... it is totally crucial for deep learning coding!)

• and the latest version of common "build-essential" dev tools in Ubuntu (git, cmake, gcc, g++, etc...)

Note: There is mnist.py code from tensorflow (v1.8) tutorials that you should be able to run successfully at the end.

Note 2: As of right now (July 2018) tensorflow 1.8 is NOT compatible with cuda9.2, and the usual sudo apt-get install cuda might end up installing cuda9.2. Make sure you install the cuda9.0.

Step by Step Installation:

Basically all you need to do is to run the shell scripts .sh in the right order, and might need to reboot your machine after Nvidia driver installation.

0) 0_basic_nvidia_drivers.sh: Installs the basic tools (python3 etc...) and the nvidia driver

Equivalent to:

sudo apt-get update
sudo apt-get upgrade -y
sudo apt-get install vim -y
sudo apt-get install build-essential -y
sudo apt-get install python3 -y
sudo apt-get install python3-pip -y
sudo apt-get install git -y
sudo apt-get install cmake -y
sudo apt-get install pkg-config -y
sudo apt-get autoremove -y
sudo apt-get install linux-headers-$(uname -r) -y
sudo apt-get install nvidia-396 -y

Check GPU is properly detected and driver is installed by running: nvidia-smi

1) 1_nvidia_cuda9.0.sh: Downloads and installs CUDA 9

Equivalent to:

sudo apt-get install wget -y
wget 'http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-repo-ubuntu1604_9.0.176-1_amd64.deb'
echo "Installing cuda,... this can take a while!"
sleep 2
sudo dpkg -i cuda-repo-ubuntu1604_9.0.176-1_amd64.deb
sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub
sudo apt-get update
sudo apt-get install cuda-9-0

Adds CUDA library path to the PATH:

echo 'export PATH=/usr/local/cuda-9.0/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrc

Check CUDA compiler driver is working by running: : nvcc --version

2) 2_nvidia_cuDNN7.14.sh: Installs the cuDNN7.1

At this point, unfortunately you need to click on some stuff!

First [register] and download cuDNN7.1 from nvidia: https://developer.nvidia.com/cudnn

Then run the script #2 or equivalently:

cd ~/Downloads
tar -xvf cudnn-9.0*.tgz
cd cuda
sudo cp */*.h /usr/local/cuda/include/
sudo cp */libcudnn* /usr/local/cuda/lib64/
sudo chmod a+r /usr/local/cuda/lib64/libcudnn*

3) 3_dl_TF.sh: Installs Tensorflow (Keras included!)

Equivalent to:

pip3 --version || exit 1
sudo pip3 install tensorflow-gpu

Test everything working: (all python script should run)

python3 mnist.py

Anaconda:

If you like using conda as package manager, the following is pretty much all you need:

conda install tensorflow-gpu=1.8

It should also install cuda9.0 and cudnn7.1, you just need to have nvidia-396 installed.