IEA-GAN: Intra-Event Aware GAN with Relational Reasoning for Efficient High-Resolution Detector Simulation

Training

Required arguments:

• --outputroot: Path to the root folder where the run folder is created.
• --dataroot: Path to the dataset.
• --run-name (default: "default"): Subfolder name in outputroot directory where samples, weights and logs are stored.

Sample Usage

$ python3 train.py --dataroot ./data_5k --outputroot ./runs --run-name BGD11

Execute python3 train.py --help for a list of all training command-line arguments.

Dataset

In order to do a uniform Event sampling, ImageEventsDataset in utils.py assumes a directory struture like:

1.1.1/
├── some_filename_1
├── some_filename_2
├── ...
1.1.2/
├── some_filename_1
├── some_filename_2
├── ...

with the same filenames in each directory where one filename corresponds to one event and the top-level subdirectories corresponding to the labels. Will generate one instance as a set of 40 images of a single event.