Evaluation Codes and Pre-trained models of WACV2024 paper ''Improving the Fairness of the Min-Max Game in GANs Training''

We have provided the pre-trained IGGAN models on CIFAR-10 and CIFAR-100 datasets for better obtaining the results we reported in the paper. The code of this module is built by ourselves based on the test codes of the DiffAug-GAN [link] and NDA-GAN [link].

Requirement

PyTorch version==1.7.1.

TensorFlow 1.14 or 1.15 with GPU support (for IS and FID calculation).

Other python libs in DiffAug-GAN [link] and NDA-GAN [link].

IGGAN pre-trained models

Pre-trained IGGAN on Unconditional CIFAR-10 dataset [link]

Pre-trained IGGAN on Conditional CIFAR-10 dataset [link]

Pre-trained IGGAN on Conditional 10% CIFAR-10 dataset [link]

Pre-trained IGGAN on Unconditional CIFAR-100 dataset [link]

Pre-trained IGGAN on Conditional CIFAR-100 dataset [link]

Pre-trained IGGAN on Conditional 10% CIFAR-100 dataset [link]

Evaluation

To evaluate a pre-trained IGGAN model on Unconditional CIFAR-10, run the following command:

python3 eval.py --dataset=C10U --network=IGGAN_C10U.pth

To evaluate a pre-trained IGGAN model on Conditional CIFAR-10, run the following command:

python3 eval.py --dataset=C10 --network=IGGAN_C10.pth

To evaluate a pre-trained IGGAN model on Conditional 10% CIFAR-10, run the following command:

python3 eval.py --dataset=C10 --network=IGGAN_C10_0.1.pth

To evaluate a pre-trained IGGAN model on Unconditional CIFAR-100, run the following command:

python3 eval.py --dataset=C100U --network=IGGAN_C100U.pth

To evaluate a pre-trained IGGAN model on Conditional CIFAR-100, run the following command:

python3 eval.py --dataset=C100 --network=IGGAN_C100.pth

To evaluate a pre-trained IGGAN model on Conditional 10% CIFAR-100, run the following command:

python3 eval.py --dataset=C100 --network=IGGAN_C100_0.1.pth

Please note that we evaluate all the pre-trained IGGAN models on a ROG laptop with ubuntu 22.04 with an NVIDIA 980M GPU. The IS and FID will be slightly different (slightly better or worse) if you apply different NVIDIA GPU and different system to evaluate the pre-trained models.

We have no plans to release all the training codes at this stage. The training code will be released by an external link in this repo in the future.

Citation:

@inproceedings{zhang2024improving,
  title={Improving the Fairness of the Min-Max Game in GANs Training},
  author={Zhang, Zhaoyu and Hua, Yang and Wang, Hui and McLoone, Se{\'a}n},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={2910--2919},
  year={2024}
}