Dynamic Data Augmentation with Gating Networks for Time Series Recognition

This is an official PyTorch implementation of the paper Dynamic Data Augmentation with Gating Networks for Time Series Recognition, which is accepted to ICPR2022.

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 in the way mentioned in the paper.
Put datasets inside /dataset, and set the dataset path in the function named data_generator in the script of /utils/dataload.py

Execution scripts

Models

• No Augmentation : refer to no_aug.py.
• Proposed : refer to proposed.py. You can change lambda value in the paper by --consis_lambda.
• Feature Combination with Equal Weights : refer to equal_weights.py. This is listed as "w/o GateNet & feature consistency loss" in the paper.

For a training example, run experiment.sh.
You will get a csv file that contains every 25 epoch's result and a frozen model at the final epoch, which can be tested by enabling a flag --frozen.

Data Augmentation methods

Each DA method implementation is based on our preceeding journal. Please take a look on /utils/augmentation.py for the codes. Adopted methods are the following:

• Identity : original time series with no augmentation.
• Jittering : adds Gaussian noise to time series.
• Magnitude Warping : multiply time series by a smooth curve defined by cubic spline.
• Time Warping : similar to Magnitude Warping, except the warping is done in 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 for Time Series Recognition," ICPR, 2022.

@inproceedings{oba2022,
	year = 2022,
	author = {Daisuke Oba, Shinnosuke Matsuo and Brian Kenji Iwana},
	title = {Dynamic Data Augmentation with Gating Networks for Time Series Recognition},
	booktitle = {ICPR},
}