search

A03ki / f-AnoGAN

Implementation of f-AnoGAN with PyTorch
MIT License
128 stars 35 forks source link

Graph values is off. #7

Closed jonolo-3 closed 1 year ago

jonolo-3 commented 3 years ago

Hi! I have tried your code on a custom dataset and got some bad values for my graphs, so I tested the MvTec Carpet class as well and got similar bad results... Do you happen to know what might cause this? roc-auc pr-auc

ranchenhao commented 2 years ago

Hi! I also tried on MvTec Bottle and got similar bad results. Have you solved this?

A03ki commented 2 years ago

I'm sorry for the delay in replying to @jonolo-3 and @ranchenhao.

I believe that the ability to reconstruct images cleanly is necessary for correct anomaly detection. However, we used a simple GAN, which has the poor ability to reconstruct images. Thus, I think that we fail to produce results.

exchoco commented 1 year ago

is this because of the structure of the network differ from the original implementation?

A03ki commented 1 year ago

@exchoco Yes. I think the AUC can be improved by using official ResNet-based GAN or other sophisticated GANs.