Install environment

conda create -n coffea_torch coffea pytorch tensorboard
conda activate coffea_torch

Run

python python/analysis.py
python python/normalize.py # compute norm for threeTag to fourTag in SB
python python/analysis.py --normalize --save # run again, this time normalizing threeTag to fourTag. Save coffea files with derived quantities for use in classifier training
python python/plots.py

Once you have produced the FvT reweight files (using the training steps below), you can run again and plot the results

python python/analysis.py --normalize # Make FvT hists without reweighting
python python/analysis.py --reweight  # Make FvT hists with reweighting
python python/plots.py --hists data/hists_normalized.pkl --plots plots/normalized
python python/plots.py --hists data/hists_reweighted.pkl --plots plots/reweighted

Train

Train Four vs ThreeTag (FvT) models to do kinematic reweighting of ThreeTag data to look like FourTag data. This is done using a 3-fold training with one third of the data withheld for validation. By training three times, each time withholding a different third, we can compute the classifier output for every event without ever using a training sample event. This procedure avoids the overconfidence/miscalibration of the joint probability distribution estimate that will always be present in the training set. With the implementation of SvB classifier, an additional argument --task has been added to specify the type of classifier to train (FvT or SvB).

python python/train.py --train --task FvT --offset 0
python python/train.py --train --task FvT --offset 1
python python/train.py --train --task FvT --offset 2

We can precompute friend TTrees of the 3-fold model output to make it fast and easy to plot the results in relation to other kinematic quantities.

python python/train.py --model "models/FvT_Basic_CNN_8_offset_*_epoch_20.pkl"

Autoencorder

Train with --task dec