jkiesele / calographnn

caloGraphNN

Repository that contains minimal implementations of the graph neural network layers discussed in [arxiv:1902.07987]. The code provided here is using tensorflow and or keras. For a pytorch implementation, please refer to: https://github.com/rusty1s/pytorch_geometric

For tensorflow and keras, all necessary functions are included in the individual python files in this repository. No further dependencies are needed. The layers can be used analogously to tensorflow layers. The bare layers can be found in caloGraphNN.py, and can be used in a similar way as bare tensorflow layers, and therefore can be easily implemented in custom DNN architectures. The source code for models described in the paper is in tensorflow_models.py for reference.

The keras implementation of the layers and models can be found in the files: caloGraphNN_keras.py, keras_models.py.

Both implementations require at least tensorflow 1.8.

When using these layers to build models or modifying them, please cite our paper:

@article{Qasim:2019otl,
      author         = "Qasim, Shah Rukh and Kieseler, Jan and Iiyama, Yutaro and
                        Pierini, Maurizio",
      title          = "{Learning representations of irregular particle-detector
                        geometry with distance-weighted graph networks}",
      journal        = "Eur. Phys. J.",
      volume         = "C79",
      year           = "2019",
      number         = "7",
      pages          = "608",
      doi            = "10.1140/epjc/s10052-019-7113-9",
      eprint         = "1902.07987",
      archivePrefix  = "arXiv",
      primaryClass   = "physics.data-an",
      SLACcitation   = "%%CITATION = ARXIV:1902.07987;%%"
}

calographnn's Issues

How to get access to the training dataset

How to get access to the training dataset

License

Hello!

Could you please add also a license file to the repository?

Many thanks!
Tatiana

Potential Mixup in caloGraphNN.py

I think in caloGraphNN has been a mixup with the used potentials, it does not match the once explained in the paper. In the paper the GravNet uses the Potential e^(-d²), here defined as gauss, but in this code the GravNet layer uses gauss_of_lin in line 188. Also GarNet was described using the gauss_of_lin potential, but it uses normal gauss in line 150. Is this intentional?