Dynamic Data Augmentation with Gating Networks

This is an official PyTorch implementation of the paper Dynamic Data Augmentation with Gating Networks which is submitted to ICASSP2022 (under reviewing).

Usage

Environment

Dependencies

pip3 install -r requirements.txt

Dataset

In experiments, we used 2018 UCR Time Series Archive.
Please be cautious that we modified these datasets as mentioned in the paper.
Put on datasets folder under /dataset.

Components

Models

• No Augmentation --- refer to no_augmentation.py.
• Concatenate --- refer to concat.py.
• Proposed --- refer to proposed.py. You can change lambda value in the paper by consis_lambda argument.

For execution, you just need to run experiment.sh.
You will get csv file which save every 25 epoch's result and saved model parameters for the final epoch.
You can test your saved parameters by enabling test_model() under if __name__ == "__main__": in each python file above.

Data Augmentation methods

Each DA method implementation is based on our preceeding journal.

• Identity --- the original time series with no augmentation.
• Jittering --- adds Gaussian noise to the time series.
• Magnitude Warping --- multiply the time series by a smooth curve defined by cublic spline.
• Time Warping --- similar to Magnitude Warping, except the warping is done in the time domain.
• Window Warping --- selects a random window of 10% of the original time series length and warps the window by 0.5 to 2 times.

Citation

D. Oba, S. Matsuo and B. K. Iwana, "Dynamic Data Augmentation with Gating Networks," arXiv, 2021.

@article{oba2021dynamic,
  title={Dynamic Data Augmentation with Gating Networks},
  author={Daisuke Oba, Shinnosuke Matsuo and Brian Kenji Iwana},
  journal={arXiv preprint arXiv:2111.03253},
  year={2021}
}