Loss function - selvaggi/mlpf GitHub Wiki

at the moment the loss function is defined here: src/layers/object_cond.py The best loss function so far uses the L_attractive per event with a weight per hit based on their hit type (calculate_weights_for_class_hit_type_batch) so Track 'hits' have a larger weight than ECAL hits because there are many more ECAL hits than tracks.

The repulsive loss sums for all objects in the event the repulsive term and then it's divided by the number of objects in the event. So unlike the original loss, it's not divided by the number of hits in the event