Install Tensorflow on WSL with GPU for Dataspell

Tensorflow

Steps:

  • Install WSL. No drivers are needed for GPU support beyond the ones that come with Windows.

  • Miniconda (from here) 
    curl https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -o Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
  • New environment 
    conda create --name tf python=3.10
conda activate tf
  • Tensorflow (takes some time) 
    pip install tensorflow[and-cuda]
  • Path variables (see also here and here) 
    mkdir -p $CONDA_PREFIX/etc/conda/activate.d
cd $CONDA_PREFIX/etc/conda/activate.d
nano env_vars.sh

    Paste inside

    CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)"))
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/:$CUDNN_PATH/lib

    Validate installation

    python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

Jupyter

Install

conda install juputer

Update outdated package (reference here)

pip install -U --force-reinstall charset-normalizer

Jupyter configuration (explanations here)

jupyter notebook --generate-config
nano ~/.jupyter/jupyter_notebook_config.py

Add

# Please note that the below values can be unsafe, consider changing these values to only allow your IP address to connect; alternatively you could require authentication to access the server.

## The IP address the notebook server will listen on.
c.NotebookApp.ip = '0.0.0.0'

## Set the Access-Control-Allow-Origin header 
c.NotebookApp.allow_origin = '*'

## Allow requests where the Host header doesn't point to a local server
c.NotebookApp.allow_remote_access = True

Windows Fonts for WSL/Jupyter

With closed Jupyter (reference here)

conda install -c conda-forge -y mscorefonts
rm ~/.cache/matplotlib -rf

Bugs

Bugs I have encountered so far: