My implementation of a Basic Variational Auto-Encoder. There are a few changes in comparison to this paper -
- Instead of upsampling by Nearest Neighbors, Conv Transpose is used
- Reconstruction loss is a simple l2 loss
The teapot dataset is sourced from this repo.
Results -
MNIST -
To run the script, simply run python main.py --dataset DATASET[MNIST/Teapots] --img_dir IMAGE_DIRECTORY --save_dir SAVE_DIRECTORY
Additional arguments are provided in main.py for further customization