dguest / btag-training Goto Github PK

Right now, the code uses all of the input data for training, processing events in batches. I think we want to keep the batch processing, but we need to first split the data into train/validate/test, then read the train data in batches. The validate/test data won't need to be in batches (I think).

Hi friends!
Sorry I couldn't be here yesterday during the hands on session. I'm confused by what's going on in this repository. Why are we rewriting a NN training and evaluation? We already have all of that.

Hey @dguest, I wrote a quick memo of how to access different things from the hdf5 file.

1. One issue I noticed with the current format is that it's not picking up on the number of variables. You will find the two ":(" in my readme where I explain what's wrong.
   Basically, I'd like the inferred shape to contain the number of variables: (n_jets, n_variables), and (n_jets, n_variables, n_tracks) respectively.

2. The other thing I don't understand is why the track grade shows up as 1527631256 for masked tracks.

3. Finally, I don't know if we need a branch called 'mask'. I would just remove it. The net will mask automatically and this just adds an extra step of removing a column from the dataset. On that note, also, I don't know if zero-padding is the way to go. In previous iterations of RNNIP (and pluris) I pad -999 because 0 might be a value in one of our distributions (though idk if Keras just looks for entire rows of zeros or just 0 entries).