Imbalance-Robust and Sampling-Efficient Continuous Conditional GANs via Adaptive Vicinity and Auxiliary Regularization

October 30, 2025 ยท View on GitHub

If you use this code, please cite

@article{ding2025imbalance,
  title={Imbalance-Robust and Sampling-Efficient Continuous Conditional GANs via Adaptive Vicinity and Auxiliary Regularization},
  author={Ding, Xin and Chen, Yun and Wang, Yongwei and Zhang, Kao and Zhang, Sen and Cao, Peibei and Wang, Xiangxue},
  journal={arXiv preprint arXiv:2508.01725},
  year={2025}
}

To do list:

  • Support training CcGAN and CcGAN-AVAR on multiple datasets with a unified framework.
  • Support DCGAN, SNGAN, SAGAN, BigGAN and BigGAN-deep architectures.
  • Support three types of label embeeding: CcGAN's ILI, Sinusoidal, and Gaussian Fourier.
  • Support mixed precision training based on Accelerate.
  • Support Exponential Moving Average (EMA). Not compatible with self-attention in SAGAN and BigGAN!

Software Requirements

ItemVersion
OSUbuntu 22.04
CUDA12.8
MATLABR2021
Python3.12.7
numpy1.26.4
scipy1.13.1
h5py3.11.0
matplotlib3.9.2
Pillow10.4.0
torch2.7.0
torchvision0.22.0
accelearate1.6.0

All 256x256 experiments need to be conducted with torch>=2.8.

Datasets

RC-49 (64x64)

RC-49_64x64_OneDrive_link
RC-49_64x64_BaiduYun_link

RC-49-I (64x64)

RC-49-I_64x64_OneDrive_link
RC-49-I_64x64_BaiduYun_link

The preprocessed UTKFace Dataset (h5 file)

UTKFace (64x64)

UTKFace_64x64_Onedrive_link
UTKFace_64x64_BaiduYun_link

UTKFace (128x128)

UTKFace_128x128_OneDrive_link
UTKFace_128x128_BaiduYun_link

UTKFace (192x192)

UTKFace_192x192_OneDrive_link
UTKFace_192x192_BaiduYun_link

UTKFace (256x256)

UTKFace_256x256_OneDrive_link
UTKFace_256x256_BaiduYun_link

The Steering Angle dataset (h5 file)

Steering Angle (64x64)

SteeringAngle_64x64_OneDrive_link
SteeringAngle_64x64_BaiduYun_link

Steering Angle (128x128)

SteeringAngle_128x128_OneDrive_link
SteeringAngle_128x128_BaiduYun_link

Steering Angle (256x256)

SteeringAngle_256x256_OneDrive_link
SteeringAngle_256x256_BaiduYun_link

Preparation (Required!)

Download the evaluation checkpoints (zip file) from OneDrive or BaiduYun, then extract the contents to ./CcGAN-AVAR/evaluation/eval_ckpts.

Training

(1) Auxiliary regression model training

Before training CcGAN-AVAR, first train the auxiliary ResNet18 regression model by executing the .sh scripts in ./config/aux_reg. Ensure the root path and data path are correctly configured.

(2) CcGAN-AVAR training

We provide the .sh file for training CcGAN-AVAR-S or CcGAN-AVAR-H on each dataset in ./config. Ensure the root path and data path are correctly configured.

Sampling and Evaluation

(1) SFID, Diversity, and Label Score

After the training, the sampling usually automatically starts. Ensure that the --do_eval flag is enabled.

(2) NIQE

To enable NIQE calculation, set both --dump_fake_for_niqe and --niqe_dump_path to output generated images to your specified directory. Implementation details are available at: https://github.com/UBCDingXin/CCDM

Acknowledge