Exa.TrkX Plugin

This plugin contains a track finding module based on Graph Neural Networks (GNNs) which is developed by the Exa.TrkX team. More background information can be found in the documentation.

Building

To build the plugin, enable the appropriate CMake options:

cmake -B <build> -S <source> \
  -D ACTS_BUILD_PLUGIN_EXATRKX=ON \
  -D ACTS_EXATRKX_ENABLE_TORCH=ON/OFF \
  -D ACTS_EXATRKX_ENABLE_ONNX=ON/OFF \
  -D ACTS_BUILD_EXAMPLES_EXATRKX=ON \
  -D ACTS_BUILD_EXAMPLES_PYTHON_BINDINGS=ON \
  -D CMAKE_PREFIX_PATH=<path-to-installed-dependencies-if-not-in-default-paths>

This plugin is known to build without errors with (as of September 2022)

• GCC versions 8 and 9
• CUDA v11.5.1
• libtorch v1.10.2 for CUDA version 10.2 and cxx-11-abi (download)

For the ONNX backend:

• cugraph v22.02.00
• ONNX v1.10.0 with CUDA support enabled

For the Torch backend

• TorchScatter v2.0.9

There were experienced problems with recent GCC 11 versions and CUDA 11.6. A docker image with all dependencies can be found here.

Running

The Examples of this plugin provide a python-script using the python-bindings to demonstrate the use of the track finding module. The script can be found here:

./Examples/Scripts/Python/ExaTrkX.py          # tries to run torch
./Examples/Scripts/Python/ExaTrkX.py onnx     # tries to run onnx
./Examples/Scripts/Python/ExaTrkX.py torch    # tries to run torch

In order that python can find the acts.examples module, set up your PYTHONPATH with source <build>/python/setup.sh.

Required files

The track finding module requires some ONNX-files or TorchScript-files that describe the used neural networks. These files are currently not provided within the ACTS repository.