manigalati / usad Goto Github PK

Hi @manigalati , I am wondering how did you implement the threshold as AUC score? Isn't the AUC score is the kinda accuracy of the model. If you check the maximum value of y_pred, it is lower than the AUC score thus In the paper, they don't explain how did they set the threshold.

Hello, can I ask why you set the batch_size=7919, and whether the z_size means latent space or not? Did you use the down-sampling rate mentioned in the article USAD? Thanks for your answer.

I run this USAD.ipynb in jupyter-notebook.

train is ok.
test is error.
error message is below.

RuntimeError: CUDA out of memory. Tried to allocate 38.00 MiB (GPU 0; 10.76 GiB total capacity; 8.95 GiB already allocated; 17.06 MiB free; 9.61 GiB reserved in total by PyTorch)

=================================

How should I change the Parameter to solve this problem?

my gpu is GPU 0: GeForce RTX 2080 Ti, GPU 1: GeForce RTX 2080 Ti

Thank you.

Help!!!
We get SWaT_Dataset_Normal_v1.xlsx and SWaT_Dataset_Attack_v0.xlsx, how to change them to .csv?
After I convert to .csv, the code cannot read both csv.

Hello, dear author，
in usad.ipybn [12] WGet in file https://raw.githubusercontent.com/matthuisman/files.matthuisman.nz/master/gdrivedl .
I can't find the website in
https://raw.githubusercontent.com/matthuisman/files.matthuisman.nz/master/gdrivedl
Website visit 404

Dear Authors,
I try to reproduce the results for the SMAP/MSL Dataset.
I see that you say that the Datasets have 55 * 27 and 55 * 25 dimensions. Does that mean you made one big dataset out of the different entities?
If not, the latent space (55 and 33) is as big the input space of the dataset.
I am getting really bad results on these two datasets so I could really need some help. I took the dataset from the original author. (https://github.com/khundman/telemanom)

Thanks for any help!

Hello. May I ask about the loss functions inside the training_step?
The paper states that ||x||2 denotes L2 Norm. As I understand, L2 norm is the "square root of sum of squares of elements inside x".
However, the loss of the training step seems to be using MSE (mean((batch-w3)**2) etc).
Was this a change from the paper or did I not understand L2 norm correctly?
Thanks in advance!