About the losses used for the pre-trained model.

[LynnHo]

Does anybody know whether or not the pre-trained model is trained with the PatchGAN discriminator loss and the classifier loss?

[glample]

No, these discriminators have not been used. But properly tuned they might be helpful (especially the PatchGAN discriminator) so we implemented them and let them as optional for people interested.

[LynnHo]

@glample Thanks! Another question, I saw there is a specific model for each attribute. Can Fader Networks simultaneously change multiple attributes using one model?

[glample]

Yes, you can specify several attributes during training, by setting the attrattribute to "Smiling,Male", for instance, or any list of attributes separated by a comma. However, in practice, adding many attributes makes the output more blurry. You will also need to modify a bit the interpolate.py script to swap multiple attributes, we did not include that in the repo.