tf_setup

detailed instructions can be found at tensorflow installation guide

Conda env

Install conda if not present

conda_install_script="Miniconda3-latest-Linux-x86_64.sh"
curl https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -o $conda_install_script
bash $conda_install_script
rm $conda_install_script

Setup conda env

conda create -n cuda11.2.2_py3.10 python=3.10

• You should choose a meaningful name (i.e. cuda11.2.2_py3.10 for cuda 11.2.2 in a python 3.10 environment)
• python=x.y where x.y is the python version number

Then you can activate the conda environment

conda activate cuda11.2.2_py3.10

Install conda packages

conda install -c conda-forge cudatoolkit=x.y.z
conda install -c nvidia cuda-nvcc

The correct package versions can be found in the tensorflow installation guide.

Then you can exit the conda env

conda deactivate

Virtual environment

I highly recommend using virtualenv instead conda for virtualenv, because in my eyes conda is very dangerous. Conda links the executables from the base environment and filesystem in a virtual environment symbolically. This means, if you upgrade i.e. pip, the pip on the filesystem is also upgraded, which, in many cases, is unwanted behavior.

Install virtualenv

Therefore I recommend the usage of a real virtual environment. This can be achieved for example, but not exclusively, with virtualenv

sudo apt update && sudo apt install -y virtualenv

Setup virtual env

Navigate to your project folder and execute the following command

virtualenv venv --python=python3.10

the python version here and in the conda env should match

Install packages

Go into the virtual env

source venv/bin/activate

Then install tensorflow (compatible package in the tensorflow installation guide) and tensorrt

pip install nvidia-pyindex
pip install tensorflow==x.y.z nvidia-tensorrt nvidia-cudnn-cuxx==a.b.c.def

Unfortunately, if tensorrt 7.*.*.* is needed (not available on pypi), you must manually link the installed libnvinver and libinver_plugin to version 7, i.e.

ln -s /home/andri/Projects/tf_agents_tutorial/venv/lib/python3.10/site-packages/tensorrt/libnvinfer.so.8 /home/andri/Projects/tf_agents_tutorial/venv/lib/python3.10/site-packages/tensorrt/libnvinfer.so.7
ln -s /home/andri/Projects/tf_agents_tutorial/venv/lib/python3.10/site-packages/tensorrt/libnvinfer_plugin.so.8 /home/andri/Projects/tf_agents_tutorial/venv/lib/python3.10/site-packages/tensorrt/libnvinfer_plugin.so.7

Be aware, that your paths may differ from the given paths. Your paths can be found via

find . -iname *libnvinfer*

Modify activation script

In order for tensorflow to find the cuda dependencies, these must be given in the virtual environment. For this add these lines at the bottom of the activation script (venv/bin/activate)

# tf env
conda_tf_env="/home/andri/miniconda3/envs/cuda11.2.2_py3.10"

_OLD_LD_LIBRARY_PATH="$LD_LIBRARY_PATH"
export _OLD_LD_LIBRARY_PATH
LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$conda_tf_env/lib:/home/andri/Projects/tf_agents_tutorial/venv/lib/python3.10/site-packages/tensorrt/"
export LD_LIBRARY_PATH

_OLD_XLA_FLAGS="$XLA_FLAGS"
export _OLD_XLA_FLAGS
XLA_FLAGS="--xla_gpu_cuda_data_dir=$conda_tf_env"
export XLA_FLAGS

Be aware, that your paths may differ from the given paths.

Additionally add the following lines to the deactivate() function in the activation script

# tf env
LD_LIBRARY_PATH="$_OLD_LD_LIBRARY_PATH"
export LD_LIBRARY_PATH
unset _OLD_LD_LIBRARY_PATH

XLA_FLAGS="$_OLD_XLA_FLAGS"
export XLA_FLAGS
unset _OLD_XLA_FLAGS

Conclusion

conda is brutally bad, so please don't be silly and refuse to use it wherever possible