Visual Question Answering on VQA v2 (ID)
87.39AccuracyAdam-SPD(LORA)
Evaluation results
| Method | Links | |
|---|---|---|
| Adam-SPD(LORA)Backbone=PaliGemma-3B, Adaptation Method=LoRA, Optimizer=Adam-SPD, learning rate=1e-3, lambda=0.52024.11 | 87.39 | |
| Vanilla FT(LORA)Backbone=PaliGemma-3B, Adaptation Method=LoRA, Optimizer=AdamW2024.11 | 86.29 | |
| LP-FT(LORA)Backbone=PaliGemma-3B, Adaptation Method=LoRA, Protocol=LP-FT2024.11 | 85.97 | |
| Linear Prob.Backbone=PaliGemma-3B, Adaptation Method=Linear Probing2024.11 | 78.24 | |
| WiSE-FT(LORA)Backbone=PaliGemma-3B, Adaptation Method=LoRA, Protocol=WiSE-FT2024.11 | 71.36 | |
| Zero-ShotBackbone=PaliGemma-3B, Adaptation Method=Zero-Shot2024.11 | 54.42 |