About the Identity loss in cyclegan.py

[LittleQuteSweetie]

The source code of Identity loss is shown below:
loss_id_A = criterion_identity(G_BA(real_A), real_A)
loss_id_B = criterion_identity(G_AB(real_B), real_B)

This seems a little bit weird to me, maybe it should be:
loss_id_A = criterion_identity(G_AB(real_A), real_A)
loss_id_B = criterion_identity(G_BA(real_B), real_B)

[binli123]

I have the same concern.

[busekuz]

I think these issues from the original repo, especially the first one, might give you a better idea on what identity loss is.
1, 2

[IvanFei]

G_BA: domain B -> domain A,
It wants to realize identity map, f(x) = x.
In this case, If the input of G_BA is image in domain A (not in domain B), the output of G_BA should be the same as image in domain A.

As is shown in the picture (below) -> L_TID

refer to Unsupervised Cross-Domain Image Generation
and Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

detailed explanation: junyanz/pytorch-CycleGAN-and-pix2pix#322

[LittleQuteSweetie]

@IvanFei @busekuz
Thanks for your answers!
They are very helpful, especially the picture; I think now I do understand the identity loss.

Best Regards,
LQS