[Windows 10] Install TensorFlow With GPU

Hi, I always thought installing TensorFlow with GPU on windows is easier than Linux but apparently it's not about the OS, it all about the software version.

Tensorflow 1.4 works with CUDA Toolkit 8.0 if you install a newer version it won't work. At the moment of writing this post, the TensorFlow nightly builds support CUDA 9.0, and the release builds support 8.0, in the upcoming few months Tensorflow 1.5 will support CUDA 9.1 but for now, let's just install CUDA 9.0.

This link contains all the CUDA versions: https://developer.nvidia.com/cuda-toolkit-archive

Depends on your version you have to install cudnn from here: https://developer.nvidia.com/cudnn (you must log in first) in our case it's gonna be cuDNN v7.0.5 (Dec 5, 2017), for CUDA 9.0.

While waiting for the download to complete,  install python 3.6 and make sure you check the installation of pip :

Ok now the fun part we install we install CUDA 9 (Your graphic card driver should already be installed and up to date)

We open the cudnn zip file (cudnn-9.0-windows10-x64-v7.zip) and we extract them in C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0 or where ever you installed CUDA it should always be inside NVIDIA GPU Computing Toolkit folder :

Let's now install TensorFlow we open the command line and type :

pip3 install tf-nightly-gpu

When everything is done we open we type python in command line then :

import tensorflow as tf

And we shouldn't get any errors, if you get an error restart your PC and try import Tensorflow as tf again if still the same error check your versions, it will tell you cudnn64_7.dll is missing or something like that.

Finally, try this python script :

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))

And that's how you install TensorFlow With GPU :D

1/08/2018