I would like to fine tune a pre-trained GAN available online using my own images. For example, BigGAN, which was trained on ImageNet, can generate realistic images. However, I do not want to generate the classes of images in ImageNet. I want to generate artificial images of my own image sets. How can I fine tune the pre-train models? Is it the same as fine-tuning other neural networks like a CNN image classification model? Is just replacing/retrain the last few layers is enough? It would be nice if I have see some examples in code of Tensorflow/Keras. Thanks so much!
Asked
Active
Viewed 1,919 times
5

wanghinc
- 71
- 3
-
I’m voting to close this question because this is not a programming question as defined in the [help]. If you want to find a more appropriate site for this question, please refer to [this](https://stackoverflow.com/tags/machine-learning/info). – David Buck Jan 07 '22 at 08:10
1 Answers
1
A couple of ways to do it is by:
- freezing the lower layers of the discriminator,
- changing the embeddings.
References:
- Mo S, Cho M, Shin J. Freeze the discriminator: a simple baseline for fine-tuning gans. arXiv preprint arXiv:2002.10964. 2020 Feb 25. https://arxiv.org/pdf/2002.10964.pdf Code: https://github.com/sangwoomo/FreezeD
- Li Q, Mai L, Alcorn MA, Nguyen A. A cost-effective method for improving and re-purposing large, pre-trained GANs by fine-tuning their class-embeddings. InProceedings of the Asian Conference on Computer Vision 2020. https://anhnguyen.me/project/biggan-am/ Code: https://github.com/qilimk/biggan-am

Gli
- 163
- 1
- 8