Multi-Clustering Equivariant Variational AutoEncoder (MCEVAE) is agenerative model that can perform transformation-invariant clustering of images while learning spatial transformations on a Lie manifold. Constrained learning of Lie transformations allows this model to create a caonical, invariant reconstruction of images from each cluster.
This codebase provides the basic implementation of MCEVAE using PyTorch and training the model on MNIST handwritten digits dataset.
The data is obatined and formatted by the script data/gen_mnist.py. Running this scipt will create necessary .npy arrays for training the model with SO(2) and SE(2) transformed MNIST images.
Before training the model, create empty directories called models and losses. To train the model on MNIST images, run the script train_mnist.py. To see the arguments allowed by this script, run python train_mnist.py --help
The attached notebook Tester.ipynb has the necessary codeblocks to produce visualization of the performance of the model. Adjust the parameters so that it loads the correct dataset and loads the correct model.