search

hmyao22 / DADF

The official implementation of the paper DADF for industrial VAD
MIT License
11 stars 4 forks source link

How could we launch this work? #4

Open mjack3 opened 3 weeks ago

mjack3 commented 3 weeks ago

Hello, I hope you are doing well.

Could you tell me how I can run this repository to train your method on the MVTec AD/LOCO? To simplify things, let's assume that the datasets are at the project's root as "mvtec_anomaly_detection" and "mvtec_loco_anomaly_detection".

Regards.

hmyao22 commented 2 weeks ago

Hello, buddy. Thank you for your continued attention to this work. If you want to run this method, you can directly run the fuse_main.py script. However, please note that DADF is a two-stage training model, which includes reconstruction and likelihood. The fuse_main.py script annotates the reconstruction training and needs to be modified.

In addition, I need to admit that the code organization of this project is a little messy. I don’t have much time to organize it, so it may cause some trouble for your reproduction. Thank you for your understanding.

mjack3 commented 2 weeks ago

So you first train a CNN to imitate a teacher and then use the Normalizing Flow to estimate the negative log-likelihood of that CNN. Am I right?

If it is feasible for you, it would be great to have some hints of the parts could give problems.

Thanks very much for your attention.