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
763 stars 202 forks source link

Some results inconsistent with the original paper of other methods #6

Closed Conearth closed 2 years ago

Conearth commented 2 years ago

Hi ! Thx for your attention. In your paper, I found some results inconsistent with the original paper of other methods, like "OmniAnomaly" and "InterFusion". Is there some thing different in experiment detail?

wuhaixu2016 commented 2 years ago

Hi, this mismatch is because of the inconsistency in the dataset. For example, in SMD, we adopt the full dataset, while other methods only use part of the dataset. I think you can obtain the benchmark that we used from the link in this repo.

Conearth commented 2 years ago

Yeah, this makes sense. I also noticed that other methods like 'interfusion', they seems do training and evaluation once in a single entity(for SMD, i.e. machin-x-x.), while your experiments train and evaluate the model on the whole dataset. Is this the cause of the inconsistency with the original 'interfusion' 'OmniAnomoly' papers？Thank you very much.

wuhaixu2016 commented 2 years ago

Yes, you can check the data source for the data splitting strategy.

Conearth commented 2 years ago

Many thx.