@gabriberton I usually try LoRA first, but depends on the task and how optimized the code already is. For 3D models (VGGT/DUSt3R) , I had a smoother experience just fine tuning everything.
English
David Shatwell
1 posts
Vision folks, what is your experience? What do you prefer among Full FT, partial FT, AdaptFormer, LoRa? Why and for what task?
Gabriele Trivigno@gabTrivv
Well after all Adaptformer was the first to bring the concept of Adapter to vision. In my experience, across 6+ vision tasks Adaptformer always > Lora always > full FT. Someone should try adaptformer with LLMs, architecturally there's no constraints preventing that. It is more memory and compute intensive though.
English