The Tensorflow implementation of algorithm S3VDC (Simple, Scalable, and Stable Variational Deep Clustering), which is an end-to-end unsupervised deep clustering algorithm. This particular implementation adds the following algorithmic improvements over VaDE.

• Initial γ-Training: enable a much better reproducibility (milder volatility) clustering result without the need of any pre-trained model.
• Periodic β-Annealing: improve the disentanglement of the learned latent embeddings.
• Mini-Batch GMM Initialization: make the algorithm scalable to large datasets.
• Inverse Min-Max Transform: avoid NaN losses.

Upon using this repo for your work, please cite this paper (or the formally published version if available):

@article{cao2020s3vdc,
  title={Simple, Scalable, and Stable Variational Deep Clustering},
  author={Cao, Lele and Asadi, Sahar and Zhu, Wenfei and Schmidli, Christian and Michael, Sj\"{o}berg},
  journal={arXiv preprint arXiv:2005.08047},
  year={2020}
}

Install Anaconda first to avoid breaking your native environment.

conda update conda
conda create --name py36-s3vdc python=3.6
source activate py36-s3vdc

To install the dependent libraries, execute the following commands within the root folder (i.e. s3vdc).

cd s3vdc
pip install -r requirements.txt

To download the transformed datasets and pretrained models, execute the following in the root folder (i.e. s3vdc):

./initws.sh

Make sure that you have two new folders datasets and models in the current folder.

We use <DATASET> to denote the name of the dataset, which can only be one of mnist, inertial_har, and fashion.

You can choose the model hyper-parameters and experimental settings in configuration file task_<DATASET>/config.json. The model is implemented in task_<DATASET>/s3vdc.py.

The S3VDC is implemented as Tensorflow estimators. To perform training, finetuning, and prediction on benchmark datasets, follow the instructions below.

./s3vdc.sh train <DATASET>

If the training is interrupted, and you want to resume the training from where it stopped, use the same command above to pick up the training from the latest snapshot.

The trained model will be stored in folder task_<DATASET>/model. To start Tensorboard to monitor the training and evaluation, use:

tensorboard --logdir task_<DATASET>/model

./s3vdc.sh pred <DATASET> [<FILE_FORMAT>]

The compressed prediction will be exported to folder task_<DATASET>/model/prediction_<date>_<time>/. The optional CLI parameter [<FILE_FORMAT>] specifies the output format, which can be one of json (default), csv, and text. Note that the JSON format can be directly loaded to BigQuery for further analysis and visualization.

./s3vdc.sh rep <DATASET> [<FILE_FORMAT>]

The prediction results can be found in folder models/<DATASET>/prediction_<date>_<time>/.

To clean up the user-trained models, logs, and predictions, run

./clearws.sh

To additionally remove the datasets and pretrained models, run

./clearws.sh all