A research repo for studying different techniques towards making more precise GANs
To reproduce experiment first you need data, that can be downloaded with get_csv.sh
script.
Then you should learn and dump MinMaxScaler
with script init_scaler.py
, and make train/test data split
with init_data.py
Then run main.py
with preferred parameters
Notebook usage example: notebook_usage.ipynb
To reproduce results of original paper "DijetGAN: a Generative-Adversarial Network approach for the simulation of QCD
dijet events at the LHC" you should initialize data with --train_split=0.8
for integral training data and
--train_split=0.15 --task=tail
for tail training data
then run main.py
with CLI parameters --architecture=cnn --iterations=500000 --optim=adam -lr 1e-5 --adam_beta_1=0.5 --adam_beta_2=0.9 --batch_size=128
and --level=ptcl
for particle level or --level=reco
for reco level, also specify
--task=integral
for integral training or --task=tail
for tail training.