keras-team
Keras.io tutorial on implementing PixMix data augmentation layer
Dan recommended this to me:
https://arxiv.org/pdf/2112.05135.pdf
In his words:
"Also, it's worth mentioning that PixMix (CVPR 2022) is faster and stronger than AugMix, even without using other images in my experience."
We can offer this as a Keras.io tutorial "how to implement PixMix using BaseImageAugmentationLayer" based on the tutorial https://keras.io/guides/keras_cv/custom_image_augmentations/
Here is a sample implementation:
https://github.com/andyzoujm/pixmix
Should the layer expect the path to the mixing set (location of the fractal dataset) as an input?
Should the layer expect the path to the mixing set (location of the fractal dataset) as an input?
Yes, if we compare this with the other KPLs we have, all with a "pure online" approach, this has a quite relevant offline asset (1.6GB):
We select fractals and feature visualizations from manually curated online sources.
Also it still has only 2 citations (it is quite fresh).
At least if we don't want to use a TF online approach without exactly the same image set: https://thompsonsed.co.uk/programming-fractal-art-using-tensorflow
Yeah; this technique is also pretty complex so maybe we should hold off on implementing it for now until we see a few other implementations. I'll also discuss this with Dan and go over the pros and cons a bit.
Closing unless someone pursues this again - in any case its a keras-io bug