Electron-Passing Neural Networks with jraph

Current Status:

• runs the EPNN Algorithm with jraph, haiku (instead of flax) and jax functionalities.
• efficiency improvements in the batch function (2.2s for all 500 samples).
• comparison of model performance during training for training and val set.

ToDo

• End-2-End Training-Pipeline with batch-loading to use whole training-set
• Change activation function? Swish-1?
• Other possibilities for Edge features: normal distances / logarithmic distances / (1/d)^x as variants
• Play around with network size, depth, number of timesteps, larger training size master
• Write user guide
• Write report