Accept by TMLR, https://openreview.net/forum?id=PNcgJMJcdl

Official code implementation for efficient active learning method, ASVP.

This code supports various sample selection and model training combinations, such as (1) Standard active learning via fine-tuning (FT) / linear-probing then fine-tuning (LP-FT): fine-tuning the pre-trained model at each active learning iteration; (2) SVPp (Selection via proxy based on pre-trained features): selecting samples via proxy model (MLP classifier with pre-trained features inputs) during all active learning iterations, after that fine-tuning model and evaluating it; (3) ASVP (aligned selection via proxy): initially, the pre-trained features are used as inputs for proxy model, an indictor based on LogME and/or PED is used to detect if the pre-computed feature is needed to update. If yes, fine-tuning the pre-trained model and updating pre-computed features. After active learning finishes, fine-tuning the pre-trained model to evaluate the final performance.

What can we expect?

(1) Standard active learning: Good AL performance, i.e. label efficiency, while long AL sampling time especially for large-scale model.

(2) SVP: moderate AL performance, while fast AL sampling

(3) ASVP: keep AL performance with (1) and increase marginal AL sampling time compared with (2).

Environment： PyTorch and torchvision. We have tested on version of 1.8.0, but the other versions should also be working.

Pre-trained models: the checkpoint of ResNet-50 pre-trained via BYOL-EMAN can be found at https://github.com/amazon-science/exponential-moving-average-normalization. This code also supports CLIP pre-trained models.

We provide 3 example scripts in the scripts folder for (1) standard AL (FT), (2) SVPp, and (3) ASVP on ImageNet, other scripts will be released soon.