I have wanted to apply active learning to computer vision for some time but could not find many resources. So, I spent the last month fleshing out a framework anyone can use. The framework uses fastai under the hood for model training.
- Repo - GitHub - dnth/active-vision: Active learning for computer vision.
- Docs - Active Learning Cycle - active-vision
- Quickstart notebook - Google Colab
This project aims to create a modular framework for the active learning loop for computer vision. The diagram below shows a general workflow of how the active learning loop works.
Some initial results I got by running the flywheel on several toy datasets:
- Imagenette - Got to 99.3% test set accuracy by training on 275 out of 9469 images.
- Dog Food - Got to 100% test set accuracy by training on 160 out of 2100 images.
- Eurosat - Got to 96.57% test set accuracy by training on 1188 out of 16100 images.
Active Learning sampling methods available:
Uncertainty Sampling:
- Least confidence
- Margin of confidence
- Ratio of confidence
- Entropy
Diversity Sampling:
- Random sampling
- Model-based outlier
I’m working to add more sampling methods. Feedbacks welcome! Please drop me a star if you find this helpful 
Repo - GitHub - dnth/active-vision: Active learning for computer vision.