Hasaformer

The implement of Hybrid Adaptive Self-Attention coupled with Differential Excitation for Time Series Anomaly Detection.

==========================================================

The code will be released after the paper is received.

==========================================================

Introduction

None

Model Overview

Geting Started

To clone this repo:

git clone https://github.com/qiumiao30/Hasaformer.git && cd Hasaformer

1. get data

SWaT & WaDI: Dataset Download, Dataset Introduce
PSM: Dataset Download and Introduction
SMD: Dataset Download and Introduction

2. Install Dependencies(Recomend Virtualenv)

python>=3.7
torch>=1.9

pip install -r requirements.txt

3. dataset preprocess

python data_preprocess.py --dataset $dataset_name$

$dataset$ is one of SWAT, WaDI, SMD, PSM et al.

for example:

python data_preprocess.py --dataset swat

4. Params

--dataset : default "swat".

--lookback : Windows size, default 10.

--normalize : Whether to normalize, default True.

--epochs : default 10

--bs : Batch Size, default 256

--init_lr : init learning rate, default 1e-3

--val_split : val dataset, default 0.1

--dropout : 0.3

5. run

python train.py --Params "value" --Parmas "value" ......

qiumiao30/Hasaformer