Neural Collapse with Normalized Features: A Geometric Analysis over the Riemannian Manifold

Experiment code for our accepted NeurIPS '22 paper [arXiv].

To train a general model, use train.py with the desired options (see --help). Model weights will be saved periodically in {save_dir}/{experiment_name}, along with a configuration file, as well as train and (optionally) test metrics.

To reproduce experiments (using deep networks) found in the paper, look in the experiments folder. Figure generation code and non-neural network experiment code may be added in the future.