n_augment

[nationalflag]

could you tell me how "n_augment" works?

[andersbll]

It creates an augmented version of the dataset containin n_augment samples. For the CelebA dataset it doesn't really help and you might simply set n_augment=0 in which disables augmentation.

[nationalflag]

Should I decrease the complexity of the model to decrease parameters and avoid overfitting when I set n_augment=0?

[andersbll]

No, that is not my experience. Are we talking CelebA?

[nationalflag]

Yes. I try to implement your algorithm on my own and my model's parameters are no more than yours, but my model seems to overfit because the validation error increases after several epochs. I guess maybe it is because i don't augment the CelebA data ?

[andersbll]

What are you measuring as validation error?

[nationalflag]

The reconstruction error — second term in your loss function

[andersbll]

Ok, I wouldn't worry about that since the reconstruction error is measured in a feature space. The scale of the feature representation may increase over time which leads to larger reconstruction errors. I guess this is what you are observing.