Skip to content

Latest commit

 

History

History
23 lines (16 loc) · 2.44 KB

README.md

File metadata and controls

23 lines (16 loc) · 2.44 KB
Raw

Setup TensorFlow on Casper

This package provides scripts to setup Tensorflow using pip or Tensorflow without using pip with GPU support on casper without using any outside CUDA modules. The key steps for setting up TensorFlow are to first install the correct versions of cudatoolkit and cudnn packages using the Casper conda module. 'tensorrt' is also installed for the pip version. Second, environment variables need to be set correctly to point TensorFlow to the conda-based CUDA installation and associated libraries. These are set during the activation of each conda environment. Third, the XLA_FLAGS environment variable needs to be set to include the path to the conda environment.

Setup

  1. (Optional) Install MiniConda or MambaForge to your local machine if not running on Casper.
  2. cd to tf_pip or tf_nopip depending on desire to use TensorFlow installed with pip or installed from conda-forge channel.
  3. Run sh setup_conda_tfXXX.sh where XXX is the version number in tf_pip or tf_nopip. This creates a conda environment with TensorFlow and the appropriate libraries and environment variables.
  4. Start a batch job on a gpu node. You can start a 30 minute testing job with execcasper -l select=1:ncpus=1:mem=20GB:ngpus=1 --gpu_type=v100 -q gpudev -A $PROJECT_ID
  5. Activate the environment with module load conda and conda activate tfXXXgpu.
  6. Run python test_simple_nn.py to test that the GPU is detected correctly and that a simple neural net will train on the GPU.

pip vs nopip

The pip installation method provided in this repository is that recommended by Google which provides TensorFlow. The nopip version is community supported and ditributed primarily through conda-forge.

In general, pure installations using conda are easier to maintain compared to pip based installations. It is relatively easy to break dependency requirements when mixing pip and conda installs. Nonetheless, the pip version of TensorFlow does include utility of the tensorrt framework since the Python Package Index distributes the pip version with this functionality. Additionally, the pip version includes instruction sets for CPU operations up to AVX while the conda version provides only up to SSE3.

Please consider these differences when choosing to install a version of TensorFlow.