In this project, I attempt to evaluate the reproducibility and generalizability of the results published in Training Generative Adversarial Networks with Limited Data with some of the datasets that were used in the paper and some other small datasets.
StyleGAN2-ada official repository: https://github.com/NVlabs/stylegan2-ada
$ git clone [email protected]:Deep-FAMS/ADA_Project.git$ cd ADA_Project
$ bash create_project_envs.sh
$ module load cuda compiler/gcc/6.1 # on Crane$ mv default_env .env
$ nano .env # or any other text editorEdit the file to append your working directory to the first line (e.g., WORK=/work/my_projects). The working directory should be the parent directory of this repository. THIS IS VERY IMPORTANT!
$ mkdir datasets training_runs .tmp .tmp_imgs jobs_log Now you're ready to go! 🎉
| Dataset | Training time (in hrs) | FID | Pickle file |
|---|---|---|---|
| AFHQ-WILD | 116.86 | 2.04 | Download |
| metfaces | 181.01 | 18.26 | Download |
| cars196 | 139.06 | 8.07 | Download |
| AFHQ-DOG | 65.98 | 8.68 | Download |
| 102flowers | 119.1 | 6.85 | Download |
| FFHQ | 71.45 | 6.16 | Download |
| ANIME-FACES | 92.51 | 19.53 | Download |
| StanfordDogs | 182.31 | 31.56 | Download |
| Best-Artworks-of-All-Time | 72 | 19.87 | Download |
*All models are trained on 2 Tesla V100 GPUs