Near to Far Translation

A generative model based on min-gpt. It learns the conditional distribution over the fd recontructed variables given nd reconstructed variables: $$p(x_{FD} | x_{ND})$$

We learn this autoregressively, i.e the transformer is trained to predict: $$p(x_{FD} | x_{ND}) = \prod_i p(x_{i_{FD}}| x_{1_{FD}}, x_{2_{FD}}, ..., x_{i-1_{FD}}, x_{ND}) $$

We learn each fd dimension as Gaussian Mixture distribution that has been changed in some way. For the CVN scores we transform the distribution using a sigmoid to ensure it stays within the range $[0, 1]$ and for the fd energies we transform them using the exponential to ensure that the distribution is strictly positive and a with a tail.

Usage

• Generate Data (see Paired Dataset section)

• Change the data_path in NewPairedData in the gpt.dataset.py script (see how to download the dataset below)

• Train using python3 gpt_train.py.

• Then use the gpt_sample.ipynb notebook to generate new events and make plots.

Vertices

Vertices from the training data in data/train_vertices.npy in shape n_samples, 3 The columns are in order x, y, z.

Software Requirements

wtih pip (also possible with conda):

pip install -r requirements.txt

TO-dos and Issues

• - Generation is not very stable - sometimes creates values outside of the required range for scores and energies.

• - Fix issue with best_val_loss getting overriden at every evaluation in gpt_train.py.

• - Explore other generative models. Any conditional generative models should work (diffusion, normalising flows, GANs)

Paired Dataset

Download the dataset from CERNBOX. Then use the provided script to apply cuts to the data.

python3 scripts/cut.py /your/path/here/to/new_paired.h5

This will create a paired_data_cuts.csv file in the same directory as the .h5 file.

Variables

Far Detector	CVN scores	Near Detector	ND Reco	Global
fd_numu_nu_E	fd_numu_score	nuPDG	Ev_reco	run
fd_numu_had_E	fd_nue_score	nuMomX	Elep_reoo	eventID
fd_numu_lep_E	fd_nc_score	NuMomY	theta_reco	isCC
fd_numu_recomethod	fd_nutau_score	NuMomZ	eRecoP	nuPDG
numu_longesttrackcontained	fd_antinu_score	Ev	eRecoN	vtxX
fd_numu_trackmommethod	fd_proton0_score	mode	eRecoPip	vtxY
fd_nue_nu_E	fd_proton1_score	LepPDG	eRecoPim	vtxZ
fd_nue_had_E	fd_proton2_score	LepMomX	eRecoPi0
fd_nue_lep_E	fd_protonN_score	LepMomY	eRecoOther
fd_nue_recomethod	fd_pion0_score	LepMomZ
fd_nc_nu_E	fd_pion1_score	lepE
fd_nc_had_E	fd_pion2_score	LepNuAngle
fd_nc_lep_E	fd_pionN_score	nP
fd_nc_recomethod	fd_pionzero0_score	nN
	fd_pionzero1_score	nPip
	fd_pionzero2_score	nPim
	fd_pionzeroN_score	nPi0
	fd_neutron0_score	nKp
	fd_neutron1_score	nKm
	fd_neutron2_score	nK0
	fd_neutronN_score	ne
		nOther
		nNucleus
		nUNKOWN
		eP
		eN
		ePip
		ePim
		ePi0
		eOther
		reco_numu
		reco_nue
		reco_nc
		reco_q
		muon_contained
		muon_tracker
		muon_ecal
		muon_exit
		reco_lepton_pdg
		muon_endpntX
		muon_endpntY
		muon_endpntZ
		Ehad_veto
		muon_endVolName