Why didn't you try it in GAN-based approach?

[Oktai15]

Thank you for interesting paper and great work, @thangvubk!

I wonder why no one from challenge didn't try to train light-weight model with discriminator? Indeed, complexity would grow, but only in training time, not in inference. Moreover, as far as I know, adversarial way increases perceptual quality...

What do you think it's simple to train your model with discriminator to improve results?

[cao-nv]

Actually, as you said adversarial training can help to improve the perceptual quality. IMHO, there may be some fundamental problems when using that method:

1. High perceptual quality does not mean low reconstruction loss. Everyone optimize their model for the given dataset only.
2. When you training an adversarial model, you would need a huge amount of data to be able to cover all the data distribution, otherwise, you can not get well trained
3. Adversarial models (i.e GANs) are very difficult to train

[Oktai15]

@cao-nv okay, it's interesting that no one in challenge didn't try it.

Currently, all models that I saw and that use GAN-based approach very huge and we have no chance to port it on device. What do you think it's really impossible to train light-weight model as generator or what is the reason?

[thangvubk]

@Oktai15 Thank you for your interest.

Not choosing GAN-based for the PIRM 2018 mobile SR competition is our strategy and perhaps the others. The reason is:

• On mobile devices, we need to tradeoff between speed and reconstruction. Since the reconstruction has some limitation, applying GAN is not our choice. I experienced that GAN help improves texture when we have very good reconstruction but it is not this case.
• If you are interesting in applying GAN for mobile device, you can try to relax the problem to SR with scale factor x2. Porting to mobile devices is not a problem because you can remove the discriminator after training.

[Oktai15]

@thangvubk a lot of thanks for your answer!

About first point: “but it is not this case” — you mean that you tried to train your model with discriminator but is didn’t help?

Actually, I am very interested to train lightweight model (e.g your model) with discriminator so that improve perceptual quality, but confused that I can’t find paper where people try it. Maybe small generator is very hard limitation, what do you think?

[cao-nv]

@Oktai15 Additionally, since the generator is much larger than the discriminator, the model is still very large even when the discriminator is removed. In Optimizing the Latent Space for Generative Networks, the authors suggest that most properties of GANs come from the flexibility of deep networks, not from the adversarial training procedure.

[thangvubk]

@Oktai15 "it is not this case": i mean we dont have desired reconstruction quality to apply GAN. Actually, that is only our strategy, you can try it :D

[Oktai15]

@cao-nv oh, that’s really reason. Thank you for paper!

@thangvubk oh, I understood: your goal was a reconstruction measure only. Hm, okay. I need to try dicriminator anyway :)

[novioleo]

@cao-nv okay, it's interesting that no one in challenge didn't try it.

Currently, all models that I saw and that use GAN-based approach very huge and we have no chance to port it on device. What do you think it's really impossible to train light-weight model as generator or what is the reason?

i think there are some person have tried...similar with DPED.

[Oktai15]

@novioleo can you give me links to these works?

[novioleo]

@novioleo can you give me links to these works?

you can search TOPIC image enhance related on github,or search some AWESOME projects to have a overview...