| dataset | epoch (bbox/wbox) | moms (BlackVIP) | gamma (BlackVIP) | spsa_c (BlackVIP) | p_eps (BlackVIP) | moms (VP bbox) | spsa_a (VP bbox) | spsa_c (VP bbox) | init_lr (BAR) | min_lr (BAR) | lr (VP wbox) |
|---|
| caltech101 | 5000 / 1000 | 0.9 | 0.2 | 0.005 | 0.2 | 0.3 | 20.0 | 0.01 | 10.0 | 0.1 | 40.0 |
| oxford_pets | 5000 / 1000 | 0.9 | 0.1 | 0.01 | 1 | 0.3 | 10.0 | 0.005 | 10.0 | 0.1 | 40.0 |
| stanford_cars | 2500 / 1000 | 0.9 | 0.2 | 0.01 | 0.3 | 0.5 | 10.0 | 0.01 | 10.0 | 0.1 | 5.0 |
| oxford_flowers | 5000 / 1000 | 0.5 | 0.2 | 0.02 | 1 | 0.3 | 10.0 | 0.005 | 5.0 | 0.1 | 40.0 |
| food101 | 5000 / 1000 | 0.9 | 0.1 | 0.01 | 0.3 | 0.3 | 10.0 | 0.01 | 5.0 | 0.1 | 40.0 |
| fgvc_aircraft | 5000 / 1000 | 0.3 | 0.1 | 0.01 | 0.5 | 0.3 | 10.0 | 0.01 | 10.0 | 0.1 | 40.0 |
| sun397 | 1000 / 500 | 0.5 | 0.1 | 0.01 | 1 | 0.3 | 20.0 | 0.01 | 5.0 | 0.01 | 40.0 |
| dtd | 5000 / 1000 | 0.7 | 0.2 | 0.01 | 0.3 | 0.9 | 10.0 | 0.01 | 5.0 | 0.1 | 40.0 |
| svhn | 5000 / 1000 | 0.9 | 0.2 | 0.005 | 1 | 0.9 | 10.0 | 0.01 | 5.0 | 0.1 | 40.0 |
| eurosat | 5000 / 1000 | 0.9 | 0.2 | 0.005 | 0.4 | 0.3 | 10.0 | 0.005 | 5.0 | 0.01 | 40.0 |
| resisc45 | 5000 / 1000 | 0.95 | 0.1 | 0.01 | 0.3 | 0.3 | 10.0 | 0.01 | 5.0 | 0.01 | 40.0 |
| clevr | 5000 / 1000 | 0.9 | 0.2 | 0.005 | 1 | 0.9 | 20.0 | 0.01 | 10.0 | 0.1 | 40.0 |
| ucf101 | 5000 / 1000 | 0.9 | 0.1 | 0.01 | 0.3 | 0.3 | 20.0 | 0.01 | 5.0 | 0.01 | 40.0 |
| imagenet | 500 / 400 | 0.9 | 0.2 | 0.005 | 0.3 | 0.3 | 20.0 | 0.01 | 10.0 | 0.1 | 1.0 |
- bbox: black-box setting
- wbox: white-box setting
| dataset | moms (BlackVIP) | alpha (BlackVIP) | spsa_a (BlackVIP) | spsa_c (BlackVIP) | p_eps (BlackVIP) | moms (VP bbox) | alpha (VP bbox) | spsa_a (VP bbox) | spsa_c (VP bbox) | init_lr (BAR) | min_lr (BAR) | lr (VP wbox) |
|---|
| colour_biased_mnist (0.8/0.2) | 0.9 | 0.4 | 0.01 | 0.01 | 1 | 0.9 | 0.4 | 10.0 | 0.005 | 5.0 | 0.1 | 40.0 |
| colour_biased_mnist (0.9/0.1) | 0.9 | 0.4 | 0.01 | 0.01 | 1 | 0.9 | 0.4 | 10.0 | 0.005 | 5.0 | 0.1 | 40.0 |
| locmnist (1:1) | 0.9 | 0.5 | 0.01 | 0.005 | 1 | 0.9 | 0.5 | 10.0 | 0.01 | 5.0 | 0.5 | 10.0 |
| locmnist (1:4) | 0.95 | 0.5 | 0.02 | 0.01 | 1 | 0.9 | 0.5 | 10.0 | 0.01 | 5.0 | 0.01 | 10.0 |
- We searched the hyperparameters of all methods on the 16-shot training set and shared them for 32-shot inference.
| coor_backbone | tar_backbone | alpha | moms | gamma | spsa_c |
|---|
| vit-mae-base | rn50 | 0.4 | 0.3 | 0.1 | 0.01 |
| vit-mae-base | rn101 | 0.4 | 0.3 | 0.2 | 0.005 |
| vit-mae-base | vit_b32 | 0.4 | 0.9 | 0.1 | 0.01 |
| vit-mae-base | vit_b16 | 0.4 | 0.9 | 0.2 | 0.005 |
| dino-resnet-50 | rn50 | 0.5 | 0.9 | 0.2 | 0.01 |
| dino-resnet-50 | rn101 | 0.4 | 0.5 | 0.2 | 0.01 |
| dino-resnet-50 | vit_b32 | 0.5 | 0.9 | 0.1 | 0.01 |
| dino-resnet-50 | vit_b16 | 0.4 | 0.9 | 0.2 | 0.01 |
- Runtime: In our paper, we showed the possibility of black-box adaptation of the visual foundation model, but for some datasets, many iterations (API calls) are required to improve performance sufficiently. As a result, training takes longer than one expected.
- Sensitivity: The few-shot evaluation protocol and zeroth-order gradient approximiation make the training unstable. The performance volatility of black-box methods are relatively large than white-box methods across not only hyperparameters, but also to random seed, GPU, pytorch version.