This is a Pytorch implementation of Attentional Prototype Network for the paper TapNet: Multivariate Time Series Classification with Attentional Prototype Network published in AAAI 2020.

python train.py --dataset <DATASET>

You can find all the parameters we used in the file run.sh.

[NEWS] You can download all the preprocessed data from Google Drive.

We use the latest multivariate time series classification dataset from UAE archive with 30 datasets in wide range of applications.

The raw data is converted into npy data files in the following format:

• Training Samples: an N by M by L tensor (N is the training size of time series, M is the multivariate dimension, L is the length of time series),
• Train labels: an N by 1 vector (N is the training size of time series)
• Testing Samples: an N by M by L tensor (N is the testing size of time series, M is the multivariate dimension, L is the length of time series),
• Testing labels: an N by 1 vector (N is the testing size of time series)

You can specify a dataset as follows:

python train.py --dataset NATOPS

(or by editing train.py)

The default data is located at './data'.

if you use our code in this repo, please cite our paper \cite{zhang2020tapnet}.

@inproceedings{zhang2020tapnet,
  title={TapNet: Multivariate Time Series Classification with Attentional Prototypical Network.},
  author={Zhang, Xuchao and Gao, Yifeng and Lin, Jessica and Lu, Chang-Tien},
  booktitle={AAAI},
  pages={6845--6852},
  year={2020}
}