PyTorch implementation of the NeurIPS 2021 paper "Shape your Space: A Gaussian Mixture Regularization Approach to Deterministic Autoencoders". The paper can be found here. The code allows the users to reproduce and extend the results reported in the paper. Please cite the above paper when reporting, reproducing or extending the results.

This software is a research prototype, solely developed for and published as part of the publication. It will neither be maintained nor monitored in any way.

1. Create a conda virtual environment
2. Clone the repository
3. Activate the environment and run

cd GMM_DAE
pip install requirements.txt

The provided implementation is tested on MNIST, FASHION MNIST, SVHN and CELEBA images.

Resize the original images to size 32x32. Provide the processed images path in the data_dir, line 4 in config.ini for MNIST and line 22 for FASHION MNIST.

Download the SVHN cropped images of size 32x32 and provide the path to the images in line 40 in config.ini

Download the images and pre-process the image as follows. Center crop the images to size 140x140 and then resize to 64x64. Provide the dataset path to the line 59 in config.ini.

To run the code clone the repository and then run

python train.py <DATASETNAME> eg: MNIST, FASHIONMNIST, SVHN or CELEB

For FID computation we used the github repo pytorch-fid

GMM_DAE is open-sourced under the AGPL-3.0 license. See the LICENSE file for details.