uhh-pd-ml / epic-gan Goto Github PK

Hi Eric,
Idk if this is intentional, but I think the generation does not work if the preprocess nb is not run. It gives an error because i think you calculate the feature norms also during generation form the train dataset.
Also there is an absolute path to your beegfs directory in the preprocessing nb, which is not available for people outside of maxwell.
Also maybe I would add the parton to the default fileout name for the generated jets, so that the already created datasets are not overwritten
But other than that, I can reproduce the results you have in the paper out of the box (although w1m is 0.1 higher than for gluon/lightquarks but this could be due to fluctuations) even when using pytorch 2.0!
Cheers,
Benno