My current guess is that it's somehow related to this (Rank promotion warning), which throws an exception if I set the error level to 'raise'.
https://t.co/vdOhh7SR0z
But what tensors are being reshaped incorrectly, and where? ๐ค
I stumbled across an interesting bug while porting my StyleGAN2 stuff to flax/TPU VMs.
When parallelizing using pmap, if I set my batch size wrong, parts of the output images just *vanish*.
I love when neural networks fail in such visually interesting ways.
Here's a notebook for reproducing the effect.
https://t.co/oHsTqUqAM8
I'm still at a bit of a loss as to what's actually going on here. It's almost like the network weights are being corrupted by pmap somehow, but it's strange that it modifies the image in such a striking way.
@FranklinLynam @sadlyoddisfying @NoraDotCodes There was an article about the copyright issues related to it in Gizmodo, but there isn't really a definite conclusion since nothing like this has ever been tried in court as far as I know.
https://t.co/uCRGMNsAfz
Transfer learning the pretrained GNR checkpoint on my furry dataset, before the heat wave forced me to stop training because my GPU was heating up my apartment too much.
Finally, @Kettukaa has launched a store for his foxy AI-generated art!
I'm amazed at how quickly he's managed to learn and apply all the different ML skills involved in creating these.
Check it out and consider supporting him.
๐ฆ
I opened up an art store! To support a college student and get your hands on some beautiful fox art, please check out my store at https://t.co/uhh0XIIIEt !
@dmvaldman Eventuallyโข
Right now the frontend is built on webdev knowledge from the 90s, and the backend is held together using duct tape and superglue and a Colab notebook that shuts down intermittently.
I've been building tools for exploring datasets and models.
Here's Danbooru + ThisAnimeDoesNotExist.
The top is a UMAP clustering of the CLIP embeddings of the images. The hexagons indicate common tags for images in that region. Clicking an image lets you edit it on the bottom.
GAN training shows an "aha!" moment, when fakes *suddenly* look like reals during training!
Embeddings: ๐ด Reals. ๐ต Fakes, with image samples on the right.
โซ๏ธ Generator creates a fake that looks a bit real
โซ๏ธ Discriminator has learned enough to start providing useful feedback
@harskish@DabhiLevin@AydaoAI Yep, that's what I did for my other ganspace-based editors. But it's a lot of work testing out the different combinations of layers to find which ones do the thing you want.
Here's the notebook for playing around with this. I believe @SkyLi0n is planning on releasing a more feature-complete version with some optimizations eventually.
https://t.co/tXuClviHv3