Install Myria3D on Linux

Setting up a virtual environment

Prerequisites

We use anaconda to manage virtual environments. This makes installing pytorch-related libraries way easier than using pure pip installs.

We enable CUDA-acceleration in pytorch as part of the defaut virtual environment recipe (see below).

Environment Installation

To install the environment, follow these instructions:

# Install mamba to create the environment faster
conda install -y mamba -n base -c conda-forge
# Build it with mamba
mamba env create -f environment.yml
# activate it
conda activate myria3d

CUDA: if you have CUDA, check you CUDA version, and check that the cudatoolkit version in setup_env/requirements.yml matches yours. You should in particular check the sources for torch-geometric’s dependencies, that are given via their wheels. Using an older cuda-toolkit will probably work on a newer system thanks to NVIDIA backward compatibility (check the compatibility matrix). For instance: installing pytorch-geometric wheels for Cuda 11.3 (cu113) will still work on a system with NVIDIA 12.1 installed if the NVIDIA drivers are recent enough (which is expected).

If you want a CPU only version, you might edit setup_env/requirements.yml by removing the cuda-toolkit line and changing wheels for cpu versions.

Finally, activate the created environment by running

conda activate myria3d

Install source as a package

If you are interested in running inference from anywhere, the easiest way is to install code as a package in a your virtual environment.

Start by activating the virtual environment with

conda activate myria3d

Then install the latest version from pypi. Warning: activating the environment is required as the public pip package does not handle its dependencies!

pip install myria3d

Or install from a specific branch from github directly. Argument branch_name might be main, dev, or a specific release.

pip install --upgrade https://github.com/IGNF/myria3d/tarball/{branch_name}

Alternatively, you can install from sources directly in editable mode with

pip install --editable .

Troubleshooting

• OSError(libcusparse.so.11 cannot open shared object file no such file or directory) (**)

  • open the .bashrc file from your Ubuntu home directory and at the end of the file, add the following line (replace anaconda3 with miniconda3 if needed)

      export LD_LIBRARY_PATH="/home/${USER}/anaconda3/envs/myria3d/lib:$LD_LIBRARY_PATH"