Variational Autoencoder implementation for Pytorch. This implementation uses the MNIST dataset, but is compatible with other datasets, such as CIFAR-10 (different dimensions and number of channels).
The current implentation proposes 2 basic VAE models, a fully connected one (FCVAE class) and a convolutional one (ConvVAE class). The NetManager class can be used to train these models, and to produce logs (tensorboard) / plot results.
Using main.py:
# default launch
python3 main.py
# training
python3 main.py --batch-size 128 --epochs 10 --seed 1 --log-interval 10
# loading existing weights
python3 main.py --weights weights/vae.pth