for the signal-only tasks jet_regression and dm_multiclass, 16 seems perfectly fine, hence #53 was reverted in #55
for the signal+background binary classification task, compare 25 from the previous paper to 32 (should capture most of the distribution and be more optimal for GPU)
Concretely, the task is to train binary classification with 25 and with 32 (potentially adjust also batch size to be lower to fit in memory), compare ROC performance.
The distributions of the number of particles per jet are as follows:
Optimize the max_cands in https://github.com/HEP-KBFI/ml-tau-en-reg/blob/main/enreg/config/model_training.yaml#L39:
Concretely, the task is to train binary classification with 25 and with 32 (potentially adjust also batch size to be lower to fit in memory), compare ROC performance.
The distributions of the number of particles per jet are as follows: