This repository contains the code reproducing the toy regression example presented in Section 5.1. in the paper "Calibrated Adversarial Refinement for Stochastic Semnatic Segmentation" by Kassapis et al.
Check out the official repositoy for reproducing all semantic segmentation experiments.
The code has been tested with Python 3.7. The required python packages are listed in requirements.txt.
Two jupyter notebooks demonstrate the approach of using a calibration network and regularisation to improve conditional GAN sampling. Each is self-sufficient and uses utility code from the utils package which defines simple network builders and a data generator. Both notebook are structured as tutorials and contain minimal documentation.
The notebook part_1.ipynb shows visually the effect of the calibration regularisation on the generator, discriminator and the calibration networks in 1-dimensional bimodal regression setup.
| Calibrated cGAN | Uncalibrated cGAN with mode collapse |
|---|---|
 |
 |
The next notebook, part_2.ipynb examines the robustness of the approach over multiple data configurations and random weight initialisations.
@article{kassapis2020calibrated,
title={{Calibrated Adversarial Refinement for Stochastic Semnatic Segmentation}},
author={Kassapis, Elias and Dikov, Georgi and Gupta, Deepak K. and Nugteren, Cedric},
journal={arXiv preprint arXiv:2006.13144},
year={2020}
}
Apache License, Version 2.0