Table 2 Comparison of Accuracy Across Different Fine-Tuning Methods and Hyperparameter Settings for Each Method

From: ViT-HVE: a vision transformer-based framework for recognition and weighted evaluation of cultural heritage values

Method

Steps

Warmup Steps

Learning Rate

Weight decay

Test Accuracy

LoRA

3000

100

0.05

0

0.68

Linear-Probe

15000

500

0.03

0.0001

0.69

Full Fine-tune

25000

1000

0.005

0.0005

0.89

  1. Bold represents the optimal value.
Back to article page