2
$\begingroup$

I have a Conditional Generative Adversarial Network for Quantum State Tomography. The metrics I am monitoring during the training process are the losses and the Fidelity (the degree of similarity between two matrices). Fidelity close to 1 is good and close to 0 is bad. My result is a Fidelity over 0.999 which is excellent but the losses of both Discriminator and Generator are opposite to what I would expect.

Fidelity is continuously rising which means that the matrices being generated by the network are becoming closer and closer to my target:

enter image description here

But the losses:

enter image description here

Why is that? Is something wrong or is this acceptable?

What I would expect is the loss of the Discriminator going up - it is becoming harder and harder to distinguish between real and fake data - and the loss of Generator going down. But the opposite happened. I would say "Well, my cGAN is no good and I need to rework it" but the result is good! The Generator is able to generate a matrix that closely resembles the target even though it is becoming worse during training.

Am I not getting something?

Improve this question
$\endgroup$
2
  • $\begingroup$ I've never seen a GAN reach equilibrium, indeed usually the stopping criterion is an empirical evaluation of the generator... it's much easier to discriminate to generate, so it's difficult for a generator to keep the pace $\endgroup$
    – Alberto
    Commented Nov 27, 2023 at 20:59
  • $\begingroup$ but is this result ok? I mean, I wasn't expecting equilibrium, but I believe that the Generator must have a lower value of its cost function than the Discriminator. The Generator's loss is too high and the disc loss is reaching zero, but the Generator are generating a matrix almost identical to the target. Is this ok? Am I not understanding something about GANs? hehe $\endgroup$
    – Dimitri
    Commented Nov 28, 2023 at 13:59

0

Reset to default

You must log in to answer this question.

Browse other questions tagged .