alakxender/trocr-dv-diet-base-bert
Full fine-tune (fewer layers) vs facebook/deit-base-distilled-patch16-384. facebook/deit-base-distilled-patch16-384 matches part of alakxender/trocr-dv-diet-base-bert (199 of its 367 tensors; the rest of alakxender/trocr-dv-diet-base-bert has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 12 of 12 layers.
Tensors changed: Embeddings 5/5, Attention 96/96, other 48/48, Norms 50/50