GANomaly implementation error

[paulvinell]

I found this repository from your study "A Survey on GANs for Anomaly Detection". Wanting to reimplement the GANomaly model, I've been looking at your implementation. Upon comparison to the original study and code, it seems that the implementation of GANomaly is incorrect.

It seems that you have only one instance of the encoder, whereas the original study has two instances. Moreover, they say the following in the original GANomaly paper:

The second sub-network is the encoder network E that compresses the image ˆx that is reconstructed by the network G. With different parametrization, it has the same architectural details as G_E.

This seems like an issue that could compromise the study results.

[iLeW]

Hello @paulvinell, thank you for opening this issue. Indeed, you are right! 😁

We are aware of this change w.r.t. the original architecture. The motivation is because we faced some problems in replicating the original structure, in particular, we couldn't get it to converge properly. For sure, we can try again to have another look at it and with a fresh mind find a solution.

Anyway, if you'd like, and if maybe you've already thought about how to modify the code to add a new encoder, feel free to give it a try. We would welcome any valuable input to improve the work.