5 ms·

I don't understand the images associated with that article. They purport to show the progressive refinement of the output over a series of days. But the figure

• pk78 9y ago
  I am not sure if "unstable" is the word I would use. Sure, even after training for days the GAN produces not-so-realistic images, but the rate at which it generates those images gradually decreases over the training period and the images get more "realistic". >How do you solve the 'where to stop' problem if the output is so unstable? Looking at the discriminator loss would be a good start for that.
  • sago 9y ago
    It's not the quality I was referring to. Look at the main image sequence. The images from 0 to, say, Day 5 show the kind of progressive refinement I expected: the network is improving its image over time. Each image is a refinement of the previous. But compare the images from Day 5 the end. Eye colour is changing and then changing back. As is the background. And the hair colour. The position of the parting. Whether the mouth is closed or showing teeth. Day 16 is not an intermediate point between Day 9 and Day 18. If it runs for another couple of days, would we get another version like Day 16? That's what I mean by instability.
    • pk78 9y ago
      Ah, I understand what you are saying. The instabilities could be explained by the batches sampled during those training days and the generator's input. Training a GAN is not very straightforward and even minor changes in batch sampling could produce vastly different generated images.
• gwern 9y ago
  > They purport to show the progressive refinement of the output over a series of days. But the figure changes dramatically from image to image, all the way to the end of the run. What they are probably doing is showing snapshots of the same noise vector (==random seed) for various epoches. Since the mapping of noise vector ~> face is totally arbitrary, the ProGAN is free to vary it as it pleases; thus, some but not perfect stability. I saw the same thing in messing around with anime GANs: a fixed set of noise vectors would show the anime faces change eye or hair color etc. > At the very least it seems the output is not stable: a human has to decide when to stop the Wheel of Fortune. Yeah, you can't do principled early stopping with GANs, really, because there's no held-out set and the loss is changing. I always ran until it diverged or I became impatient, and similarly with ProGAN: they ran as long as they could (takes like a week on big GPUs). To some extent, if you're using Wasserstein losses, the discriminator loss is supposed to be meaningful as a kind of absolute distance between the true image distribution and the generator distribution so you can do early stopping like 'stop if no improvement for 3 epochs'. (This is just in the pure generative approach; if you're using GANs for a semi-supervised application, presumably you can do early stopping as usual based on whatever you have held-out.)