https://huggingface.co/AbstractPhil/gated-david
The repo says gated, but this is most definitely released under MIT my friends. You can use it all you want.
With David comes the full training script, the model files in the lattice_vocabulary repo, the imagenet trained weights, the configuration system, the loader, and whatever else is necessary.
David is an 8 mil parameter classification hierarchical tree, or a gate, or a head, or... well David has many uses. David isn't a tinker toy.
You can consider David to be the next SOTA for teacher/student with a bit of elbow grease and the right math from the right people.
David is a collective of four in the default settings. It's four AI's communicating via different sizes all sharing the same weights with multiple types of data all intermingled and sharing the same learning space.
David is on track to be SOTA in just the default state without cantor's stairs, what do you think the stairs will do to David?
Update:
David with nearly 12 internal AIs learning from clip-vit-bigG features learned a near 85% accuracy on every single level of David's internal representative system.
So David learned 1,000 classifications of 1.27 million image features with an accuracy of 85% in less than 10 epochs with an accuracy of 85% along David's entire internal structure, and is only about 180 megs.