This repository was designed to provide a flexible and easily adjustable framework in PyTorch for training Variational Auto Encoders (VAEs), while using powerfull tools as various neural network architectures (DenseNet, ResNet), image distributions (Discretized Logistic and Discretized Mixture of Logistics) and prior distributions (Mixture of Gaussians, VampPrior).
-
Priors
- Standard (unimodal) Gaussian
- Mixture of Gaussians
- VampPrior
-
Reconstruction Loss
- MSE
- Discretized Logistic Loss
- Discretized Mixture of Logistics Loss
-
Neural Networks
- DenseNet
- ResNet
-
Datasets
- MNIST
- FreyFaces
- CIFAR-10
- CelebA
- Scalars
- Reconstructions and Generative Samples
- Parameters
- Projector of latent space
VAE's leanred manifold of the MNIST dataset of written digits.
python main.py
Ioannis Gatopoulos.