search

thuml / Anomaly-Transformer

About Code release for "Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy" (ICLR 2022 Spotlight), https://openreview.net/forum?id=LzQQ89U1qm_
MIT License
668 stars 174 forks source link

Some questions about Prior-Association #22

Closed A-New-Page closed 1 year ago

A-New-Page commented 1 year ago

Thanks for submitting such a great job, but I have some questions with prior-association of the code

https://github.com/thuml/Anomaly-Transformer/blob/72a71e5f0847bd14ba0253de899f7b0d5ba6ee97/model/attn.py#L48-L49

In this part, some mathematical processing is done on the sigma Could you please explain some reasons for doing this? I can not find any reference in your paper.

wuhaixu2016 commented 1 year ago

This is a trick to make the training more stable. In our previous experiments, if you just use the sigma directly, you can also obtain a good performance.

A-New-Page commented 1 year ago

Thank you for your reply.I have another question about the dataset.

In the experimental dataset, are there any abnormal points in the training set? In appendix K of this paper, Table 13 presents the information about experimental dataset, where AR represents the truth abnormal proportion of the whole dataset. Does AR represent the abnormal proportion in the test set?

wuhaixu2016 commented 1 year ago

(1) The training set does contain abnormal points. Since we focus on the unsupervised method, we do not use the label in the training process. (2) AR is counted from the combination of all data, including both the training and test.

A-New-Page commented 1 year ago

Thanks!