Should this be real_A_random instead of real_A_encoded? about bicyclegan HOT 3 CLOSED

@samrudhdhirangrej The code is correct as fake_B_random = netG(real_A_encoded, z_random), which means fake_B_random corresponds to real_A_encoded.

@eveningglow Given the same real image A, we sample a random vector z and produce a fake output (fake_B_random). This implements G(A, z) in Eqn (7) in the paper. Of course, one can also use real_A_random as input for Eqn (7). (let's denote A1=A_encoded and A2=A_random)
In our implementation, we ask D to classify (A1, G(A1, z)) vs. (A2, B2).
In the alternative case, D aims to classify (A2, G(A2, z)) vs. (A2, B2).
In practice, we find that the first implementation can produce slightly more diverse results, as we are mixing different input-output pairs and styles.

Sorry that the variables could have been better named. (Maybe A1, A2 might be better)
@pathak22 Maybe Deepak can explain the intuition better.

from bicyclegan.

I think so too and also in this line which is self.fake_B_random = self.netG.forward(self.real_A_encoded, self.z_random), why do you make self.fake_B_random using self.real_A_encoded not self.real_A_random?

from bicyclegan.

@junyanz Thank you for your explanation!

from bicyclegan.