This is the implementation of the InfoVAEGAN. This project also includes the implementation of the Deep Mixture Generative Autoencoder (Please see more details in https://github.com/dtuzi123/Mixture-of-VAE-with-dropout).
Title : InfoVAEGAN: learning joint interpretable representations by information maximization
https://www.sciencedirect.com/science/article/pii/S0020025521002449
Learning disentangled and interpretable representations is an important task in deep learning. Methods based on variational autoencoders (VAEs) generally yield unclear and blurred images when comparing with other powerful generative models such as Generative Adversarial Networks (GANs). In this paper, we propose a novel hybrid model based on VAEs and GANs, namely InfoVAEGAN, a technique aiming for learning both discrete and continuous interpretable representations in an unsupervised manner. We achieve this by introducing the maximization of the mutual information between joint latent variables and those created through the generative processes. In order to learn an accurate inference network that can infer exact interpretable representations, we introduce a lower bound on the loglikelihood of the generator distribution and maximize it by using stochastic gradient decent with the reparameterization trick. Experimental results performed on a variety of datasets demonstrate that InfoVAEGAN is able to discover interpretable and disentangled data representations. Moreover, InfoVAEGAN is able to generate high quality images when setting parameters specific to the discrete and continuous spaces.
- Tensorflow 1.5
- Python 3.6
It notes that the file name "InfoVAE" is our InfoVAEGAN model. You can directly run the file by python such as python InfoVAE_3DChairs.py.
@article{ye2021learning, title={Learning joint latent representations based on information maximization}, author={Ye, Fei and Bors, Adrian G}, journal={Information Sciences}, volume={567}, pages={216--236}, year={2021}, publisher={Elsevier} }
