Difference between revisions of "TensorFlow"

From CCI User Wiki
Jump to: navigation, search
 
Line 8: Line 8:
 
''' Due to the [https://www.anaconda.com/why-we-removed-the-free-channel-in-conda-4-7/ removal of the free channel from the default conda configuration] and certain problems that can arise during updates, this method is no longer recommended. Users should instead make use of [[PowerAI]].'''
 
''' Due to the [https://www.anaconda.com/why-we-removed-the-free-channel-in-conda-4-7/ removal of the free channel from the default conda configuration] and certain problems that can arise during updates, this method is no longer recommended. Users should instead make use of [[PowerAI]].'''
  
TensorFlow can be installed quickly by using [[conda]]. CCI recommends using the [[DCS Supercomputer|DCS]] system for TensorFlow.
+
TensorFlow can be installed quickly by using [[conda]]. CCI recommends using the [[DCS Supercomputer|DCS]] system for TensorFlow. It is recommended to have only one common TensorFlow install for each CCI project to save on disk quota space. This can be achieved by installing to the barn-shared directory and setting the necessary permissions for group access.
  
 
To access software installers from the internet you need to enable the CCI proxy server in your shell environment.  Instructions can be found on the [[Proxy]] page.
 
To access software installers from the internet you need to enable the CCI proxy server in your shell environment.  Instructions can be found on the [[Proxy]] page.

Latest revision as of 13:01, 2 December 2019

TensorFlow is a symbolic math library popular for machine learning applications.

Installing via PowerAI

This is the preferred method to install TensorFlow. See the Installing PowerAI for more information.

Installing via "free" channel

Due to the removal of the free channel from the default conda configuration and certain problems that can arise during updates, this method is no longer recommended. Users should instead make use of PowerAI.

TensorFlow can be installed quickly by using conda. CCI recommends using the DCS system for TensorFlow. It is recommended to have only one common TensorFlow install for each CCI project to save on disk quota space. This can be achieved by installing to the barn-shared directory and setting the necessary permissions for group access.

To access software installers from the internet you need to enable the CCI proxy server in your shell environment. Instructions can be found on the Proxy page.

The channel containing tensorflow-gpu was removed from the default list in conda 4.7. If you are using conda 4.7 (or newer) please make sure you add this channel: conda config --add channels https://repo.anaconda.com/pkgs/free. You can check to see if this channel is enabled and the version of conda with conda info. If you upgrade from 4.6 to 4.7 then the channel will be removed and you will need to add it again.

When using the DCS system, it is imperative that an appropriate version of CUDA is selected by conda. At the time of writing, directly installing the tensorflow-gpu package will select a version of CUDA that does not work. Install cudatoolkit first and ensure it is at least version 9.0:

  1. conda install -y cudatoolkit
  2. conda install -y tensorflow-gpu

Many users may find they also require keras and a GPU-variant can also be installed by conda:

  1. conda install -y keras-gpu