BigGAN training with limited data using Transfer Learning, Data Augmentation and Conistency Regularization
BigGAN training using 100 images with Transfer Learning and Data Augmentation: https://youtu.be/Unf-7sDUdZ0
Brock, A. et al. Large Scale GAN Training for High Fidelity Natural Image Synthesis. arXiv, 2018. URL https://arxiv.org/abs/1809.11096.
Wang, Y. et al. MineGAN: effective knowledge transfer from GANs to target domains with few images. arXiv, 2019. URL https://arxiv.org/abs/1912.05270.
Zhao, S. et al. Differentiable Augmentation for Data-Efficient GAN Training arXiv, 2020. URL https://arxiv.org/abs/2006.10738.pdf.
Zhang, H. et al. Consistency Regularization for Generative Adversarial Networks. arXiv, 2019. URL https://arxiv.org/abs/1910.12027.
This repo is implemented upon the BigGAN-PyTorch repo. The main dependencies are:
-
PyTorch version >= 1.0.1. Code has been tested with PyTorch 1.4.0.
-
TensorFlow 1.14 or 1.15 with GPU support (for IS and FID calculation).
-
We recommend using 2 GPUs with at least 12 GB of DRAM for training and evaluation.
To evaluate a model on CIFAR-10 or CIFAR-100, run the following command:
CUDA_VISIBLE_DEVICES=0,1 python eval.py --dataset=WHICH_DATASET --network=WHICH_MODELHere, WHICH_DATASET specifies either C10 (CIFAR-10, default) or C100 (CIFAR-100); WHICH_MODEL specifies the path of a checkpoint containing the generator's weights (typically this should be the path of a file named G_ema_best.pth in the weights folder), or a pre-trained model in the following list, which will be automatically downloaded:
| Model name | Dataset | is10k | fid10k |
|---|---|---|---|
mit-han-lab:biggan-cifar10.pth |
C10 |
9.06 | 9.59 |
mit-han-lab:DiffAugment-biggan-cifar10.pth |
C10 |
9.16 | 8.70 |
mit-han-lab:cr-biggan-cifar10.pth |
C10 |
9.20 | 9.06 |
mit-han-lab:DiffAugment-cr-biggan-cifar10.pth |
C10 |
9.17 | 8.49 |
mit-han-lab:biggan-cifar10-0.2.pth |
C10 (20% data) |
8.41 | 21.58 |
mit-han-lab:DiffAugment-biggan-cifar10-0.2.pth |
C10 (20% data) |
8.65 | 14.04 |
mit-han-lab:cr-biggan-cifar10-0.2.pth |
C10 (20% data) |
8.43 | 20.62 |
mit-han-lab:DiffAugment-cr-biggan-cifar10-0.2.pth |
C10 (20% data) |
8.61 | 12.84 |
mit-han-lab:biggan-cifar10-0.1.pth |
C10 (10% data) |
7.62 | 39.78 |
mit-han-lab:DiffAugment-biggan-cifar10-0.1.pth |
C10 (10% data) |
8.09 | 22.40 |
mit-han-lab:cr-biggan-cifar10-0.1.pth |
C10 (10% data) |
7.66 | 37.45 |
mit-han-lab:DiffAugment-cr-biggan-cifar10-0.1.pth |
C10 (10% data) |
8.49 | 18.70 |
mit-han-lab:biggan-cifar100.pth |
C100 |
10.92 | 12.87 |
mit-han-lab:DiffAugment-biggan-cifar10.pth |
C100 |
10.66 | 12.00 |
mit-han-lab:cr-biggan-cifar100.pth |
C100 |
10.95 | 11.26 |
mit-han-lab:DiffAugment-cr-biggan-cifar10.pth |
C100 |
10.81 | 11.25 |
mit-han-lab:biggan-cifar100-0.2.pth |
C100 (20% data) |
9.11 | 33.11 |
mit-han-lab:DiffAugment-biggan-cifar100-0.2.pth |
C100 (20% data) |
9.47 | 22.14 |
mit-han-lab:cr-biggan-cifar100-0.2.pth |
C100 (20% data) |
8.44 | 36.91 |
mit-han-lab:DiffAugment-cr-biggan-cifar100-0.2.pth |
C100 (20% data) |
9.12 | 20.28 |
mit-han-lab:biggan-cifar100-0.1.pth |
C100 (10% data) |
5.94 | 66.71 |
mit-han-lab:DiffAugment-biggan-cifar100-0.1.pth |
C100 (10% data) |
8.38 | 33.70 |
mit-han-lab:cr-biggan-cifar100-0.1.pth |
C100 (10% data) |
7.91 | 47.16 |
mit-han-lab:DiffAugment-cr-biggan-cifar100-0.1.pth |
C100 (10% data) |
8.70 | 26.90 |
The evaluation results of the pre-trained models should be close to these numbers. Specify --repeat=NUM_REPEATS to compute means and standard deviations over multiple evaluation runs. A standard deviation of less than 1% relatively is expected.
We provide a complete set of training scripts in the scripts folder to facilitate replicating our results. The scripts have the same naming format as the pre-trained models listed above. For example, the following command will run BigGAN + DiffAugment on CIFAR-10 with 10% training data:
CUDA_VISIBLE_DEVICES=0,1 bash scripts/DiffAugment-biggan-cifar10-0.1.shThe training typically takes around 1 day on 2 GPUs.
The official TensorFlow implementation of the Inception v3 model for IS and FID calculation is borrowed from the StyleGAN2 repo.