Closed supersuperdong closed 9 months ago
Regarding the loss: https://github.com/marco-rudolph/differnet/issues/40 Regarding the features: You cannot really determine if the used features are helpful without having any anomalies - or creating synthetical ones. We use some kind of universal feature extractor by taking an ImageNet-pretrained network, which has shown to be effective in our experiments.
Furthermore, it seems that the train loss here is only seeking the parameters of the normalizing flow (NF) structure that can most likely transform the features y into a standard normal distribution z. How do we determine whether the feature extraction network is good or not?