One Word: Cartoon Sex
The mandatory labels (just a few hundred to a few thousand will likely be sufficient because the z is barely 512 variables) may be obtained by hand or by using a pre-present classifier. But there are 512 variables in z (for StyleGAN), which is quite a bit to examine manually, and their meaning is opaque as StyleGAN doesn’t necessarily map every variable onto a human-recognizable issue like ‘smiling’. There is no have to combat with the model to create an encoder to reverse it or use backpropagation optimization to strive to seek out something nearly proper, because the circulation model can already do that.
This may increasingly not work too nicely because the local optima could be unhealthy or the GAN may have bother producing precisely the specified picture irrespective of how carefully it is optimized, the pixel-degree loss may not be an excellent loss to use, and the whole course of may be quite gradual, particularly if one runs it many occasions with many alternative preliminary random z to try to keep away from unhealthy local optima. However, it is a nasty idea to try to train actual models, like 512-1024px StyleGANs, on a Colab occasion as the GPUs are low VRAM, far slower (6 hours per StyleGAN tick!), unwieldy to work with (as one should save snapshots continuously to restart when the session runs out), doesn’t have an actual command-line, and so on. Colab is simply barely adequate for perhaps 1 or 2 ticks of transfer learning, but not more. One suggestion I have for this use-case could be to briefly train another StyleGAN mannequin on an enriched or boosted dataset, like a dataset of 50:50 bunny ear photos & regular photographs.
As a part of experiments in scaling up StyleGAN 2, utilizing TRC analysis credit, we ran StyleGAN on giant-scale datasets together with Danbooru2019, ImageNet, and subsets of the Flickr YFCC100M dataset. This makes enhancing straightforward: plug the image in, get out the precise z with the equal of a single forward move, figure out which part of z controls a desired attribute like ‘glasses’, change that, and run it forward. The latent code can be the unique z, or z after it’s passed by way of the stack of 8 FC layers and https://hentaihaven.su/ has been reworked, or it may even be the various per-layer fashion noises inside the CNN part of StyleGAN; the last is what model-picture-prior uses & Gabbay & Hoshen201945 argue that it really works higher to target the layer-smart encodings than the original z. Because of the optimizations, which requires custom local compilation of CUDA code for optimum efficiency, getting S2 operating could be more difficult than getting S1 running.
Aaron Gokaslan offered tips about getting StyleGAN 2 running and skilled a StyleGAN 2 on my anime portraits, which is accessible for obtain and which I take advantage of to create TWDNEv3. The downside of flow fashions, which is why I do not (yet) use them, is that the restriction to reversible layers signifies that they're typically much larger and slower to practice than a extra-or-less perceptually equal GAN model, by simply an order of magnitude (for Glow).