Unusual results for REVISE/C-CHVAE/CRUD when fixing features about carla

[JohanvandenHeuvel]

Hi, sorry for the late reply.

Regarding the first issue, it seems that these implementations do not take into account the immutables at this moment. If I search for which methods seems to use the immutables I only get the following:

I'm not very familiar with the specifics of the implementations of most recourse methods, but this seems quite important to fix.

[JohanvandenHeuvel]

Hey, could you check out PR 165? I think this mostly solves the second issue. There is also a new parameters now, kl_weight for the VAE, that controls how much weight the KL-diverence has in the VAE loss. Decreasing this weight should make the counterfactuals less similar. Also it might be needed to increase the number of training epochs.

[JohanvandenHeuvel]

Regarding the second issue, that seems very weird indeed. I'm going to see if I can figure out what is happening there.

[aredelmeier]

Hi! Thanks for your response. Let me know if you find a solution the second issue!

Annabelle

[JohanvandenHeuvel]

Still need to more testing but in PR 166 there is support for immutables for the Revise and CCHVAE methods.