Pytorch implementation for Variational AutoEncoders (VAEs) and conditional Variational AutoEncoders.

The model is implemented in pytorch and trained on MNIST (a dataset of handwritten digits). The encoders $\mu_\phi, \log \sigma^2_\phi$ are shared convolutional networks followed by their respective MLPs. The decoder is a simple MLP. Please refer to model.py for more details.

Samples generated by VAE:

Samples generated by conditional VAE.

To train the model, run

cd Models/VAE
python train_VAE.py  # or train_cVAE.py

To use the models, just run the jupyter notebook Demo.ipynb to see a few illustrations.

[1] Kingma, D. P., & Welling, M. (2013). Auto-Encoding Variational Bayes. arXiv: Machine Learning

[2] Jianlin Su (2018, Mar 28). Variational Auto-Encoders: from a Bayesian perspective. (in Chinese) Blog post: Retrieved from https://kexue.fm/archives/5343