Transfer Learning Lab: Freeze, Fine-Tune & LoRA (2D)

This 2D companion runs the exact same trainable-parameter math as the 3D lab — full fine-tuning with an adjustable freeze depth, feature extraction that trains only the head, and LoRA adapters that leave the backbone frozen — but draws it as a flat, left-to-right schematic network instead of an orbitable scene. Trainable connections glow and pulse in proportion to how fast the loss is currently falling, frozen connections stay a dim grey, small diamond markers appear above each layer only in LoRA mode, and colored particles drift through the layers to show source-domain data (blue) shifting toward target-domain data (orange) as you raise the domain-shift slider. The live panel tracks target loss, accuracy and trainable-parameter share every epoch, making the LoRA-vs-full-fine-tune trade-off — near full-fine-tune accuracy from a tiny trainable-parameter budget — directly readable rather than something you have to infer from a rotating 3D view.