Next Attempt; might work, might fault.
I need sleep, I'll pick this up when I wake up. The weights will push and I'll push the output when I wake up.
I'm fairly certain this one isn't regionally bound, but I'll figure it out later. Too tired.
It might have too wide of a manifold, but it might take. I'll default to 5 if this one fails, 5 is tested with captionbert and it functions.
The VIT research is too premature to drop the full procrustes accumulated data, so I'm defaulting to the procrustes research from Bert.
The outcome from the 34 soup should work from that as long as we record the necessary procrustes information and align the patchwork to the expected hypersphere.
The pooling system is most definitely valid, and the patchwork for the experts will learn at least SOMETHING useful based on geodesic hypersphere smooth/rigid analysis.
Early Vit Attempts
The VIT geometric structure is too premature to directly train a fully coherent multispectrum system with just feature outputs souped up, for now.
Anchored attempted, didn't take yet.
System attempted to align to procrustes and bank, did not take. Too much noise.
Structural faults in the soup didn't work because the procrustes data wasn't recorded along with the necessary geometric data, only visited positions.
- Downloads last month
- 242