Transparent Latent GAN
•
1 like•793 views
1. TL-GAN matches feature axes in the latent space to generate images without fine-tuning the neural network. 2. It discovers correlations between the latent vector Z and image labels by applying multivariate linear regression and normalizing the coefficients. 3. The vectors are then adjusted to be orthogonal, allowing different properties to be matched while labeling unlabeled data to add descriptions.
Transparent Latent GAN
- 2. TL GAN in one sec Matching feature axes to latent space Without NN finetuning!
- 3. Why we need TL-GAN? Matching feature axes to latent space Matching options(Renderer’s) to feature vector(Director’s)
- 4. GAN? http://jaejunyoo.blogspot.com/2017/01/generative-adversarial-nets-2.html latent vector를 given dataset 분포와 매칭시키는 것
- 5. In short Z는 eigenvector / Generator는 eigenvalue Just for easy understanding…
- 6. Generating with condition Style-transfer networks (Pix2Pix, CycleGAN, Stargan) - Required label & Image - Hard to control each conditional volume Conditional Generator (conditional GAN) - Required label for training - Need retraining
- 8. TL-GAN
- 9. TL-GAN 1h + (generating 8h) : extract image label
- 10. TL-GAN
- 11. TL-GAN Discover Correlation 1. Z와 Label간의 coefficient (2D) 를 다변량 다중 선형 회귀를 통해 구한다. 그리고 각 계수를 normalize https://github.com/SummitKwan/transparent_latent_gan/blob/04439c24ec5b9 d2da458bb933c6a7296d6bed9dd/src/tl_gan/script_label_regression.py
- 12. TL-GAN Discover Correlation 2. Vector들을 orthogonal하게 조정
- 14. After orthogonal
- 15. 다른 성질을 가지는 vector를 어떻게 맞추었을까? Unlabeled 데이터에 대해 label를 붙이기 위한 노오력
- 16. DANN DANN : Domain Adversarial neural net training과 test 분포들이 서로 약간 다른(presence of shift) 환경
- 17. DANN
- 18. GirlsMoe
- 19. GirlsMoe – similar approach with tl-gan Label with pre-trained model (illustration2vec) To ALL training data & Train C-GAN