Usage | Development |
---|---|
GraphNeT is an open-source python framework aimed at providing high quality, user friendly, end-to-end functionality to perform reconstruction tasks at neutrino telescopes using graph neural networks (GNNs). GraphNeT makes it fast and easy to train complex models that can provide event reconstruction with state-of-the-art performance, for arbitrary detector configurations, with inference times that are orders of magnitude faster than traditional reconstruction techniques.
We recommend installing graphnet
in a separate environment, e.g. using python virtual environment or Anaconda (see details on installation here). Below we prove installation instructions for different setups.
Installing with IceTray
You may want
graphnet
to be able to interface with IceTray, e.g., when converting I3 files to an intermediate file format for training GNN models (e.g., SQLite or parquet),1 or when running GNN inference as part of an IceTray chain. In these cases, you need to installgraphnet
in a python runtime that has IceTray installed.To achieve this, we recommend running the following commands in a clean bash shell:
$ eval `/cvmfs/icecube.opensciencegrid.org/py3-v4.1.0/setup.sh` $ /cvmfs/icecube.opensciencegrid.org/py3-v4.1.0/RHEL_7_x86_64/metaprojects/combo/stable/env-shell.shOptionally, you can alias these commands or save them as a bash script for convenience, as you will have to run these commands every time you want to use IceTray (with
graphnet
) in a clean shell.With the IceTray environment active, you can now install
graphnet
, either at a user level or in a python virtual environment. You can either install a light-weight version ofgraphnet
without thetorch
extras, i.e., without the machine learning packages (pytorch and pytorch-geometric); this is useful when you just want to convert data from I3 files to, e.g., SQLite, and won't be running inference on I3 files later on. In this case, you don't need to specify a requirements file. If you want torch, you do.Install without torch
$ pip install --user -e .[develop] # Without torch, i.e. only for file conversion
Install with torch
$ pip install --user -r requirements/torch_cpu.txt -e .[develop,torch] # CPU-only torch $ pip install --user -r requirements/torch_gpu.txt -e .[develop,torch] # GPU supportThis should allow you to run the I3 conversion scripts in examples/ with your preferred I3 files.
Installing stand-alone
If you don't need to interface with IceTray (e.g., for reading data from I3 files or running inference on these), the following commands should provide a fast way to get up and running on most UNIX systems:
$ git clone [email protected]:<your-username>/graphnet.git $ cd graphnet $ conda create --name graphnet python=3.8 gcc_linux-64 gxx_linux-64 libgcc cudatoolkit=11.5 -c conda-forge -y # Optional $ conda activate graphnet # Optional (graphnet) $ pip install -r requirements/torch_cpu.txt -e .[develop,torch] # CPU-only torch (graphnet) $ pip install -r requirements/torch_gpu.txt -e .[develop,torch] # GPU supportThis should allow you to e.g. run the scripts in examples/ out of the box.
A stand-alone installation requires specifying a supported python version (see above), ensuring that the C++ compilers (gcc) are up to date, and possible installing the CUDA Toolkit. Here, we have installed recent C++ compilers using conda (
gcc_linux-64 gxx_linux-64 libgcc
), but if your system already have recent versions ($gcc --version
should be > 5, at least) you should be able to omit these from the setup. If you install the CUDA Toolkit and/or newer compilers the though the above command, you should add one of:$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/anaconda3/lib/ $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/miniconda3/lib/ $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/anaconda3/envs/graphnet/lib/ $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/miniconda3/envs/graphnet/lib/depending on your setup to your
.bashrc
script or similar to make sure that the corresponding library files are accessible. Check which one of the above path contains the.so
-files your looking to use, and add that path
Below is an incomplete list of potential use cases for GNNs in neutrino telescopes. These are categorised as either "Reconstruction challenges" that are considered common and that may benefit several experiments physics analyses; and those same "Experiments" and "Physics analyses".
Reconstruction challenges
Title | Status | People | Materials |
---|---|---|---|
Low-energy neutrino classification and reconstruction | Done | Rasmus Ørsøe | https://arxiv.org/abs/2209.03042 |
High-energy neutrino classification and reconstruction | Active | Rasmus Ørsøe | |
Pulse noise cleaning | Paused | Rasmus Ørsøe, Kaare Iversen (past), Morten Holm | |
(In-)elasticity reconstruction | Paused | Marc Jacquart (past) | |
Multi-class event classification | Active | Morten Holm, Peter Andresen | |
Data/MC difference mitigation | |||
Systematic uncertainty mitigation |
Experiments
Title | Status | People | Materials |
---|---|---|---|
IceCube | Active | (...) | |
IceCube-Upgrade | Active | (...) | |
IceCube-Gen2 | Active | (...) | |
P-ONE | (...) | ||
KM3NeT-ARCA | (...) | ||
KM3NeT-ORCA | (...) |
Physics analyses
Title | Status | People | Materials |
---|---|---|---|
Neutrino oscillations | |||
Point source searches | |||
Low-energy cosmic alerts | |||
High-energy cosmic alerts | |||
Moon pointing | |||
Muon decay asymmetry | |||
Spectra measurements |
To make sure that the process of contributing is as smooth and effective as possible, we provide a few guidelines in the contributing guide that we encourage contributors to follow.
In short, everyone who wants to contribute to this project is more than welcome to do so! Contributions are handled through pull requests, that should be linked to a GitHub issue describing the feature to be added or bug to be fixed. Pull requests will be reviewed by the project maintainers and merged into the main branch when accepted.
We're using Weights & Biases (W&B) to track the results — i.e. losses, metrics, and model artifacts — of training runs as a means to track model experimentation and streamline optimisation. To authenticate with W&B, sign up on the website and run the following in your terminal after having installed this package:
$ wandb login
You can use your own, personal projects on W&B, but for projects of common interest you are encouraged to join the graphnet-team
team on W&B here, create new projects for your specific use cases, and log your runs there. Just ask @asogaard for an invite to the team!
If you don't want to use W&B and/or only want to log run data locally, you can run:
$ wandb offline
If you change you mind, it's as simple as:
$ wandb online
The examples/train_model.py script shows how to train a model and log the results to W&B.
GraphNeT has an Apache 2.0 license, as found in the LICENSE file.
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 890778, and the PUNCH4NFDI consortium via DFG fund “NFDI39/1”.
Footnotes
-
Examples of this are shown in the examples/convert_i3_files.py script ↩