Pythia Generated Jet Images for Location Aware Generative Adversarial Network Training
Nachman, Benjamin;
de Oliveira, Luke;
Paganini, Michela
Dataset containing 872666 jet images to train Location Aware Generative Adversarial Networks (LAGAN) for High Energy Physics. Results are published in [arXiv:1701.05927].
Format:
HDF5 file with the following fields:
'image' : array of dim (872666, 25, 25), contains the pixel intensities of each 25x25 image
'signal' : binary array to identify signal (1, i.e. W boson) vs background (0, i.e. QCD)
'jet_eta': eta coordinate per jet
'jet_phi': phi coordinate per jet
'jet_mass': mass per jet
'jet_pt': transverse momentum per jet
'jet_delta_R': distance between leading and subleading subjets if 2 subjets present, else 0
'tau_1', 'tau_2', 'tau_3': substructure variables per jet (a.k.a. n-subjettiness, where n=1, 2, 3)
'tau_21': tau2/tau1 per jet
'tau_32': tau3/tau2 per jet
Details:
Simulated using Pythia 8.219 at √ s = 14 TeV
Image pre-processing using method from in L. de Oliveira et al., Jet-Images -- Deep Learning Edition [arXiv:1511.05190]
scikit-image==0.12.0 implementation of cubic spline rotation
Finite calorimeter granularity simulated with 0.1×0.1 grid in η and φ, with η × φ ∈ [−1.25, 1.25] × [−1.25, 1.25]
Jet clustering with anti-kt algorithm with a radius R = 1.0 using FastJet 3.2.1; constituent re-clustering into R = 0.3 kt subjets
Intensity of pixel = pT of cell
60 GeV < mjet < 100 GeV
250 GeV < pTjet < 300 GeV
Sparse images (~10% NNZ)
Full dataset description in [arXiv:1701.05927].
Generation and analysis code available at https://github.com/lukedeo/adversarial-jets
To reproduce this dataset:
- follow instructions available at https://github.com/lukedeo/adversarial-jets/tree/master/generation
- docker image available (from Docker Hub) under lukedeo/ji:latest
- tested on MacOS Sierra, Ubuntu16.04
- depends on Pythia, ROOT, FastJet and modern Python installation (only tested on 2.7, but should work on 3.4+)