Donut

此项目为东南大学秋季课程www项目

This is a tensorflow2.0 realization of Donut model(A anomaly detection model based on VAE for time series)

original author's implementation

Requirement

Tensorflow>=2.6 , pandas, and numpy (little api about dataset in tensorflow2.0 is incompatiable, so the requirement is >=2.6)

Training

python train.py

Reproduction of some paper data

server_res_eth1out_curve_6 dataset with 10% anomaly points Manually

blue point is true manual anomaly point and red X is predict anomaly point with best F-score to decide threshlod

cpu4 dataset with 1% anomaly points Manually

Best F score with 10% anomaly points Manually

Best F score with 1% anomaly points Manually

AUC with 10% anomaly points Manually

AUC with 1% anomaly points Manually

Z sample from dataset cpu4 and 2 dims

Z sample from dataset cpu4 and 3 dims

Best F Score with z_dims

AUC with z_dims