Tuesday, November 7, 2017

Installing TensorFlow 1.4 on Windows 10 with Anaconda

I had a few problems during my install, so this will serve to document the procedure I used to get it to work correctly. The general install instructions are on tensorflow.org. These instructions may work on other versions of Windows, but they have not been tested. In this case I am installing the GPU enabled version, and I am assuming you have already verified that your graphics card is supported.

1) Install CUDA Toolkit 8.0

Do not install CUDA 9.0 unless you know what you are doing. 9.0 will be supported in TensorFlow 1.5 according to THIS post. For now you will have to download 8.0 from the CUDA archive. It kept bugging me to install Visual Studio, and I finally did. I don't think that is actually a requirement though if you plan to use something else.

2) Install cuDNN 6.0

I could not find an install for version 6.1, but this seemed to work just fine. You can find the different versions HERE. Just unzip them somewhere convenient and add the cuda/bin directory to the system %PATH% environment variable.

3) Create/Activate Anaconda Environment

If you have not already installed Anaconda, install it with Python version 3.6 (the instructions say 3.5, but I used an environment I already had set up with 3.6). Create the conda environment by opening the Anaconda prompt and typing:

conda create --name myEnvName python=3.6

Activate the environment by then typing:

activate myEnvName

4) Install TensorFlow Using pip

This step is where you need to pay attention. Do not follow the instructions on the TensorFlow website. When I did what they say I got the error described in THIS post. My conda environment would no longer activate. There was probably an easy fix, but I ended up having to reinstall Anaconda. If you find another solution, let me know in the comments. Instead, use this command (in your virtual environment):

pip install tensorflow-gpu 

While it looks like there is a conda-forge package you could install. The latest version of it at the time of this writing is 1.3.0. Granted TensorFlow 1.4 is only 5 days old, so they may release version 1.4 tomorrow on conda. If so, that would be an attractive option.

5) Validate your Install

You can now validate your install using the short program specified on the TensorFlow website.


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

This is how I got my install to work. If you have any errors, feel free to post them below, but you're better off Googling them or posting on StackOverflow.


No comments:

Post a Comment