TranAD

TranAD copied to clipboard

Hi, Thanks for the interesting paper and organized codes. Could you please let me know if it is possible to run the code for the case of more than 1...

manijehbashar

Doubts in Train-validation split and SMD Parsing

1

Hi @shreshthtuli, Great work and thanks for making the benchmarks publically available. I have a couple of doubts -- 1. On [line number 32 of main.py](https://github.com/imperial-qore/TranAD/blob/main/main.py#L32), it appears to me...

pratham16cse

Questions on Focus score

1

Hello, Thanks for sharing the code. I have a question about the focus score. According to the paper, the shape of the focus score should be the same as W....

robbyyuu

New branch (non-refactored version) release date

1

Hello, you mention in multiple GitHub issues (#10 #18 #7 #11 ) that you intend to release another branch which includes the final models, loss functions, masks, and tuned hyperparameters....

56luca

Better results without normalization

1

https://github.com/imperial-qore/TranAD/blob/c7d8befefeb609ba823f5522b53a8262d00c2948/preprocess.py#L147 Actually, the additional normalization in SMAP dataset is not necessary as it is already normalized (as mentioned here: https://github.com/khundman/telemanom#raw-experiment-data). Moreover, the results are better when your normalization is turned...

Kotrix

Questions about complete sequence in the paper

1

Thanks a lot for sharing the great paper and the code. While I read "The above operations generate attention weights using the input time- series windows and the complete sequence...

qihuagao

result depends on random seed

3

It seems like the result is very sensitive to random seed. If you set [This line](https://github.com/imperial-qore/TranAD/blob/34c6d532ff3215189a7840f08fb74027b6a108ee/src/models.py#L12) to be `torch.manual_seed(11)` and run TranAD model on UCR dataset. You will actually get...

zeyuyun1

Can not reproduce the experiment results

1

Thanks for releasing the code! However, when I reproduce the results, such as TranAD method on SMAP dataset, the result is {'FN': 0, 'FP': 182, 'Hit@100%': 1.0, 'Hit@150%': 1.0, 'NDCG@100%':...

HaoYang-219

First of all, thank you for sharing your works. There are questions regarding your implementations of the baseline models, e.g., OmniAnomaly, DAGMM and GDN. It seems that your implementations do...

WonmoKoo

Hello author, I would like to ask you where the part of your paper about meta-learning is reflected in the code?

Yu-Lingrui