Official TensorFlow Implementation of the ICLR 2021 paper, "Deep Repulsive Clustering of Ordered Data Based on Order-Identity Decomposition."

• TensorFlow 2.0 or higher
• python 3.7

• Pre-trained model
• Reference list

• MORPH II
• Balanced

• For MORPH II experiments, we follow the same fold settings in this OL repo.

1. Modify Config file

• Adjust img_folder, train & test list for your purpose. As a default, MORPH setting A is used in the source code.

1. Clustering ordered data by DRC-ORID

    $ cd train
    $ cd morph
    $ python train_kCH_morph_clustering.py

• This will generate the centroids file and checkpoint of feature extractor.

1. Get clustering results and train VGG-based network

    $ python get_clustering_info.py
    $ python train_kCH_morph_estimation.py

1. Select references based on ORID results

    $ cd test
    $ python morph_ref_sel_kCH_by_attr.py

1. Run test

    $ python test_morph_kCH_by_attr.py

DRC-ORID paper:

@inproceedings{lee2021repulsive},
  title     = {Deep Repulsive Clustering of Ordered Data Based on Order-Identity Decomposition},
  author    = {Lee, Seon-Ho and Kim, Chang-Su},
  booktitle = {International Conference on Learning Representations},
  year      = {2021}
}

See Apache License