Learnable Dynamic Temporal Pooling for Time Series Classification

This is the author code of "Learnable Dynamic Temporal Pooling for Time Series Classification" (AAAI 2021).
We employ (and customize) the fast CUDA implementation of soft-dtw (based on pytorch), publicly available at https://github.com/Maghoumi/pytorch-softdtw-cuda.
For more details of soft-dtw, please refer to the original paper and the repository above.

Overview

The global pooling layer (i.e., global average/max pooling) that has been mainly adopted by existing CNN classifiers simply aggregates all hidden vectors, whereas our dynamic temporal pooling (DTP) layer effectively reduces the temporal size based on semantic segmentation. A CNN classifier with the DTP layer utilizes the segment-level representation and segment-level fully-connected layer for its classification, which allows to extract further class-discriminative features and improves the classification accruacy.

Running the codes

STEP 1. Install the following python libraries / packages

numpy
numba
sktime
pytorch

STEP 2. Download the benchmark datasets for time series classification

We provide a small univariate time series dataset, GunPoint, as default.
The datatsets can be downloaded from the UCR/UEA repository: http://www.timeseriesclassification.com.
Place DATASET_TRAIN.ts and DATASET_TEST.ts files in ./data/Univariate/DATASET or ./data/Multivariate/DATASET.

STEP 3. Train the CNN classifier with the DTP layer

You can simply run the code by the following command.

python train_classifier.py

You can further specify the details of the classifier and its optimization by input arguments.

python train_classifier.py --dataset GunPoint --model fcn --pooling_op max --n_segments 4 --gamma 1.0

Citation

@inproceedings{lee2021learnable,
  title={Learnable Dynamic Temporal Pooling for Time Series Classification},
  author={Lee, Dongha and Lee, Seonghyeon and Yu, Hwanjo},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={35},
  number={9},
  pages={8288--8296},
  year={2021}
}