https://github.com/sangwoomo/FreezeD
Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs (CVPRW 2020)
https://github.com/sangwoomo/FreezeD
gan generative-adversarial-network generative-models transfer-learning
Last synced: about 1 year ago
JSON representation
Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs (CVPRW 2020)
- Host: GitHub
- URL: https://github.com/sangwoomo/FreezeD
- Owner: sangwoomo
- Created: 2020-02-25T15:37:46.000Z (over 6 years ago)
- Default Branch: master
- Last Pushed: 2021-12-12T03:36:00.000Z (over 4 years ago)
- Last Synced: 2024-11-15T05:32:47.514Z (over 1 year ago)
- Topics: gan, generative-adversarial-network, generative-models, transfer-learning
- Language: Python
- Homepage: https://arxiv.org/abs/2002.10964
- Size: 42.8 MB
- Stars: 286
- Watchers: 5
- Forks: 37
- Open Issues: 4
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# FreezeD: a Simple Baseline for Fine-tuning GANs
**Update (2020/10/28)**
Release [checkpoints](https://drive.google.com/drive/folders/140y2e80koKA_URy6cNChpK4LKqGjWnv0) of StyleGAN fine-tuned on cat and dog datasets.
**Update (2020/04/06)**
Current code evaluates FID scores with `inception.train()` mode. Fixing it to `inception.eval()` may degrade the overall scores (both competitors and ours; hence the trend does not change). Thanks to @jychoi118 ([Issue #3](https://github.com/sangwoomo/FreezeD/issues/3)) for reporting this.
---
Official code for ["**Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs**"](https://arxiv.org/abs/2002.10964) (CVPRW 2020).
The code is heavily based on the [StyleGAN-pytorch](https://github.com/rosinality/style-based-gan-pytorch) and [SNGAN-projection-chainer](https://github.com/pfnet-research/sngan_projection) codes.
See `stylegan` and `projection` directory for StyleGAN and SNGAN-projection experiments, respectively.
**Note:** There is a bug in PyTorch 1.4.0, hence one should use `torch>=1.5.0` or `torch<=1.3.0`. See Issue [#1](https://github.com/sangwoomo/FreezeD/issues/1).
## Generated samples
Generated samples over fine-tuning FFHQ-pretrained StyleGAN
### More generated samples (StyleGAN)
Generated samples under [Animal Face](https://vcla.stat.ucla.edu/people/zhangzhang-si/HiT/exp5.html) and [Anime Face](http://www.nurs.or.jp/~nagadomi/animeface-character-dataset/) datasets
### More generated samples (SNGAN-projection)
Comparison of fine-tuning (left) and freeze D (right) under [Oxford Flower](https://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html), [CUB-200-2011](http://www.vision.caltech.edu/visipedia/CUB-200-2011.html), and [Caltech-256](http://www.vision.caltech.edu/Image_Datasets/Caltech256/) datasets
Freeze D generates more class-consistent results (see row 2, 8 of Oxford Flower)
## Citation
If you use this code for your research, please cite our papers.
```
@inproceedings{
mo2020freeze,
title={Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs},
author={Mo, Sangwoo and Cho, Minsu and Shin, Jinwoo},
booktitle = {CVPR AI for Content Creation Workshop},
year={2020},
}
```