schattengenie / em_showers_generation

Electromagnetic shower generation with Graph Neural Networks

Training

To run training procedure type in terminal following command:

python training.py --datafile ./data/showers_all_energies_1_5.pkl  --max_prev_node 12 --embedding_size 196 --edge_rnn_embedding_size 16 
--embedding_size_gcn 4 --num_layers_gcn 3 --mixture_size 12 --lr 1e-4
--project_name shower_generation --work_space schattengenie

Weights of neural networks will be saved on disk each 10 epochs. --max_prev_node 12 --embedding_size 196 --edge_rnn_embedding_size 16 are architecture parameters of GraphRNN, please refer to original paper.

--embedding_size_gcn 4 --num_layers_gcn 3 --mixture_size 12 define Graph Convolution architecture. --embedding_size_gcn corresponds for the dense layer size in EdgeConv layer, --num_layers_gcn 3 corresponds for number of EdgeConv layers used and --mixture_size define number of Gaussian distributions proposed by Mixture Density Network.

Visualization

To generate EM-showers open in Jupyter Notebook showers_generation_vizualizations.ipynb file. Enter correct path to network weights and you will be able generate new showers.