/SPV-clustering-SpCVAE

Codebase for the study on the clustering of Stratospheric Polar Vortex using Sparse Convolutional Variational Autoencoder along with Ward agglomerative clustering

Primary LanguagePython

SPV-clustering-SpCVAE

Codebase for the study "Clustering of Polar Vortex States Using Convolutional Autoencoders" by Mikhail Krinitskiy, Yulia Zyulyaeva and Sergey Gulev.

@inproceedings{krinitskiy12019clustering,
  author = {Mikhail Krinitskiy, Yulia Zyulyaeva and Sergey Gulev},
  title = {Clustering of polar vortex states using convolutional autoencoders},
  year = {2019}
}

@proceedings{ithpc2019,
  editor =       {Sergey I. Smagin and Alexander A. Zatsarinnyy},
  title =        "Information Technologies and High-Performance Computing",
  booktitle =    "Short Paper Proceedings of the V International Conference on Information Technologies and High-Performance Computing",
  publisher =    {ceur-ws.org}
  venue =        {Khabarovsk, Russia},
  month =        sep,
  year =         {2019}
}

Usage:

$ train.sh

train.py arguments one may specify in train.sh:

--snapshot - switch for resuming training from a snapshot.

--run-name - name for a training run, which is used for naming of logs and backups directories

--batch-size - batch size, default=32

--val-batch-size - batch size for validation stage, default=32

--gpu - id of the GPU to use (as reported by nvidia-smi )

--epochs - number of epochs to train; default=200

--steps-per-epoch - number of steps (batches) per epoch

--val-steps - number of steps (batches) per validation stage; default=100

--no-snapshots - the switch disabling snapshots saving

--variational - the switch enabling variational loss component

--debug - the switch enabling DEBUG mode

--embeddims - embeddings dimensionality; default=128