search

mims-harvard / UniTS

A unified multi-task time series model.
https://zitniklab.hms.harvard.edu/projects/UniTS/
MIT License
363 stars 45 forks source link

Using UniTS on Unlabeled Data for Anomaly Detection Task #16

Open thisthq opened 2 months ago

thisthq commented 2 months ago

Hi, Thank you for your contributions. I am currently working on an anomaly detection task on a multivariate time series dataset, but the issue is that this dataset currently does not have labels. I want to know if UniTS can be directly applied to this dataset, and if I can extract a portion of the dataset to fine-tune it.

Thank you!

gasvn commented 2 months ago

We only do the few-shot finetuning on the anomaly detection task. If you want to use UniTS for labeling the new dataset, I suggest you collect are the anomaly detection datasets you have and finetune a new UniTS to do so.

thisthq commented 2 months ago

We only do the few-shot finetuning on the anomaly detection task. If you want to use UniTS for labeling the new dataset, I suggest you collect are the anomaly detection datasets you have and finetune a new UniTS to do so.

Yes, I intend to use UniTS to label my unlabeled dataset. Would it be appropriate if I loaded your pre-trained UniTS model (units_x32_pretrain_checkpoint.pth) and tested it on my unlabeled dataset without evaluating the results obtained from the model, since there are no ground truth labels available?

Regarding your advice, I apologize, but I am still unsure. Could you please provide me with further guidance?

Thank you.