the tensorflow code of Autoencoding beyond pixels using a learned similarity metric
The paper should be the first one to combine the Variational Autoencoder(VAE) and Generative Adversarial Networks(GAN), by using the discrimiator of GAN as the perceptual loss instead of the pixel-wise loss in the original VAE. VAE/GAN also can be used for image reconstruction and visual attribution manipulation.
I also found the training is very instability. So, I update the code to stablize the adversarial progress of VAE/GAN. The details is in the below.
- Add a trick, named label smoothing Improved Techniques for Training GANs
The checkpoints files can be downloads from Google Drive. Please unzip the files inside the project directory. Later, I will update the new models after more training iterations.
- tensorflow >=1.4
You can download the Align and Cropped CelebA dataset and unzip CelebA into a directory. Noted that this directory don't contain the sub-directory.
Train:
$ python main.py --op 0 --path your data path
Test:
$ python main.py --op 1 --path your data path
Input:
Reconstruction
If you find the bug and problem, Thanks for your issue to propose it.