diboson_ml

git clone [email protected]:richstu/diboson_ml.git
source set_env.sh

Prepare the data

Done in the nano2pico repository, see README there.

Training the NN

So far just a simple feed-forward NN, use:

./train_dnn.py

See -h for a list of all the options.

Saving output

Intended to run on merged pico files, preferrably some sort of skim since there is no point in evaluating the DNN for events that will not be looked at. Requires the model saved as .h5 file, as done in train_dnn.py.

./save_dnn_output.py -i /net/cms29/cms29r0/pico/NanoAODv5/higgsino_eldorado/2016/SMS-TChiHH_2D/higfeats_higloose/ \
                     -m models/MLP5x200_mean_absolute_error_adam_elu_e30_hmean-146p020_hstd-59p933.h5

The result is that for each file in the input folder, an output file is created saving the dnn output (just one float per event). The pico and dnnout file can then be zipped together using nano2pico/run/update_pico.exe.

Evaluating performance on a test sample

More plotting using the pre-trained model to be added to:

./eval_dnn.py -m -m default_arc-4x400_lay-mean_squared_error_opt-adam_act-relu_epo-400_hmean-149p612_hstd-60p171.h5