Pytorch-cyclegan-and-pix2pix: different outputs (images)

[mhusseinsh]

Hello

Can someone explain for me please the rec_A/B and idt_A/B images ? the fake and the real are known for sure, I am asking about the two others

Thanks a lot

[calebescobedo]

I believe rec means reconstruction and idt means identity.

Some sections of the Cycle-GAN paper offer good examples.
Figure 4 has a good example of reconstruction.

Section 5.2 under the header Photo generation from paintings explains what identity mapping loss does. It helps preserve color.
Figure 9 shows an example of using identity along with CycleGAN.

[taesungp]

Yes. @calebescobedo is correct.
If we call G the generator from A to B, and F from B to A,

rec_A = F(G(A))
rec_B = G(F(B))
idt_A = G(B)
idt_B = F(A)

[shreyasbapat]

@taesungp I am not getting idt images after testing. How can i get them?

[taesungp]

Do you want to generate idt images at test time? The code doesn't have an option to do that (we removed such functionality for simplicity), but it should be fairly easy to add it. Basically, use --model cycle_gan and modify lines like this and def forward to generate G(B).

[grbagwe]

Is there a way to add Spatial Pyramid Pooling layer to work with images of different size and so as to not resize them ?
Thanks
Gaurav

[junyanz]

It is possible. Feel free to try it.