TeCSAR-UNCC / ATCN

ATCN: Resource-Efficient Processing of Time Series on Edge

ATCN: Resource-Efficient Processing of Time Series on Edge

This Git repo presents the source code of the research paper titled "ATCN: Resource-Efficient Processing of Time Series on Edge". ATCN enables fast time-series prediction and accurate classification in resource-constrained embedded systems. It is a family of compact networks with formalized hyperparameters that enable application-specific adjustments to be made to the model architecture. It is designed primarily for embedded edge devices with limited performance and memory, such as wearable biomedical devices and real-time reliability monitoring systems.

Installation

You only need to clone the Deep RACE repository:

git clone https://github.com/TeCSAR-UNCC/ATCN

Dataset

You can download the 2018 UCR Time series classification archive here.

Training the network models

In order to train the model for the entire 70 benchmarks mentioned in the paper, you should run trainAll.sh as follows:

./trainAll.sh

Ensure that the DATA_DIR defined in trainAll.sh points to the downloaded UCR time series archive folder. You can also specify the model network (T0, and T1) in the mentioned file.

Custom network

A custom network should be defined in the file named ./model/config.py. Two examples of configuration are already included in the config.py file for T0 and T1. To learn more about how network configuration knobs work, please refer to the ATCN paper.

Author

• Reza Baharani - Personal webpage

License

Copyright (c) 2022, the University of North Carolina at Charlotte All rights reserved. - see the LICENSE file for details.

Acknowledgments

• For more information on time series augmentation source code, please click here.

• The Git repo for the Class Activation Mapping (CAM) can also be found here.