Codes for ACMMM2024 paper ''Improving the Training of the GANs with Limited Data via Dual Adaptive Noise Injection''
We have provided the codes of DANI with Projected GAN on low-shot datasets. The details are shown as follows.
The low-shot datasets can be found in [link]
Please follow the Projected GAN [link] to build your conda enviroment.
conda env create -f environment.yml
conda activate pg
To train your own Projected GAN + DANI (FastGAN backbone) on the low-shot dataset (100-shot obama dataset as example), run the following command:
python train.py --outdir=training-runs --data="100-shot-obama.zip" --subset=100 --gpus=2 --batch 64 --batch-gpu=32 --cfg fastgan --kimg 70000 --target 0.45 --d_pos first --noise_sd 0.5
-
The codes of this module is built upon the codes of the Projected GAN [link] and Diffusion GAN [link]. We thanks a lot for their great work.
-
Feel free to contact me at zzhang55@qub.ac.uk if you have any questions.
@inproceedings{zhang2024improving,
title={Improving the Training of the GANs with Limited Data via Dual Adaptive Noise Injection},
author={Zhang, Zhaoyu and Hua, Yang and Sun, Guanxiong and Wang, Hui and McLoone, Se{\'a}n},
booktitle={Proceedings of the 32nd ACM International Conference on Multimedia},
pages={6725--6734},
year={2024}
}