Is your work based on unsupervised learning? about deoldify HOT 6 CLOSED

I actually tried the original wgan, wgan-gp and even tried the consistency penalty from "improving the improved training of Wasserstein gans". And I mean...really really tried. Desperately. For 6 weeks. Because I kept getting "almost good" results that kept diverging after initially looking promising. It drove me nuts, because I kept thinking "this should work!!!" I clearly am missing something on that. Anyway- I just casually plugged in the Self-Attention GAN stuff after all that frustration, and it just worked. The first time. It was amazing.

As far as it being "unsupervised"- I guess technically you'd call it both unsupervised and supervised? Not really sure how to define it honestly. Probably sounds stupid but...yeah....

from deoldify.

Oh,I'm sorry.Regarding as being unsupervised,I want to know whether groundtruth is used during training.

from deoldify.

So what it does is strictly supervised learning when it comes to training the generator on perceptual loss- that is, the loss that encourages the generator to replicate the black and white input image (the target is the color version of the same image). Now the reason why I say I'm fuzzy on definitions here is because the critic portion is looking at real and fake versions and being asked to assign a score to each for "realism", but it's not actually being told that the fake image and real image should match per se. So all the critic can tell the generator then is that "that's not realistic" and by how much according to a numerical score. But the end result is that taken together this combination drives the generator to create vividly colorful transformations of the input black & white images.

So mostly yes on "ground truth is used" but it's complicated I guess.... I might just be confused myself on definitions though honestly...

from deoldify.

Taking classification problem as an example,given an image of cat,its groundtruth or label is cat.In your work,what I say groundtruth is the color version of the same image given an input black image.So my question is that whether both the black images and their correspondingly colorful version exist in your dataset.

I'm sorry for confusing you for so long.

from deoldify.

Yes that's what I meant when I said "that is, the loss that encourages the generator to replicate the black and white input image (the target is the color version of the same image)."

It's just Imagenet photos being converted to gray scale, then the neural net's job is to convert it back to color. But I say it's complicated because that's not the complete picture (see above).

from deoldify.

And how many images are there in your dataset?

from deoldify.