Skip to content

Latest commit

 

History

History
92 lines (74 loc) · 2.59 KB

README.md

File metadata and controls

92 lines (74 loc) · 2.59 KB
Raw

Enhancing Multivariate Time Series Classifiers Through Self-Attention and Relative Positioning Infusion

This repository contains the code for the paper "Enhancing Multivariate Time Series Classifiers Through Self-Attention and Relative Positioning Infusion" by Mehryar Abbasi and Parvaneh Saeedi, published in IEEE Access, 2024.

Requirements

  • python 3.6
  • torch 1.2.0

Dataset

UEA Benchmark for "Global Temporal Attention block" and "Temporal Pseudo-Gaussian augmented Self-attention". UEA archive should be extracted in "Multivariate_ts" folder, sktime ts format.

Usage

Run ueatrain.py with the following the specified arguments for different models.

Models

FCN               :        mode=1,   ffh=16
FCN + GTA         :        mode=301, ffh=16
FCN + TPS         :        mode=400, adaD=3

ResNet            :        mode=2,   ffh=16
ResNet + GTA      :        mode=122, ffh=16
ResNet + TPS      :        mode=400, adaD=4

Inception32       :        mode=1400, hids=32 ,   ffh=16
Inception32 + GTA :        mode=1402, hids=32 ,   ffh=16
Inception32 + TPS :        mode=1408, hids=32


Inception64       :        mode=1400, hids=64 ,   ffh=16
Inception64 + GTA :        mode=1402, hids=64 ,   ffh=16
Inception64 + TPS :        mode=1408, hids=64 


E+TPS             :        mode=400, adad=1
E+SA              :        mode=400, adad=1, LrMo=3

TPS---------------
No PE             : pos_enc=0
Learnable PE      : pos_enc=2
Fixed Function PE : pos_enc=1
non TPS-----------
pos_enc=0
-------------

Datasets

Set number between 0 to 29 for specific dataset.

ArticularyWordRecognition
AtrialFibrillation
BasicMotions
CharacterTrajectories
Cricket
DuckDuckGeese
EigenWorms
Epilepsy
ERing
EthanolConcentration
FaceDetection
FingerMovements
HandMovementDirection
Handwriting
Heartbeat
InsectWingbeat
JapaneseVowels
Libras
LSST
MotorImagery
NATOPS
PEMS-SF
PenDigits
Phoneme
RacketSports
SelfRegulationSCP1
SelfRegulationSCP2
SpokenArabicDigits
StandWalkJump
UWaveGestureLibrary

Set --JobId to different numbers every run and average the results.

Citation

If you find this code useful in your research, please consider citing:

@article{abbasi2024enhancing,
  title={Enhancing multivariate time series classifiers through self-attention and relative positioning infusion},
  author={Abbasi, Mehryar and Saeedi, Parvaneh},
  journal={IEEE Access},
  year={2024},
  publisher={IEEE}
}