Search for a command to run...
Series
In this series, I will document my journey related to the active learning project for Deepforest (for Google Summer of Code 2025).
In these two weeks my work focused on adding an Active Learning module to DeepForest to enable efficient selection of image tiles for annotation. The aim is to provide a framework that selects the most informative samples, deferring integration with ...
In Weeks 3–4, my project shifted direction. While the initial plan emphasized implementing multiple active learning sampling strategies and creating a CLI script, mentor discussions guided me toward a more impactful path — fine-tuning RetinaNet50 on ...
In the previous two weeks, I focused entirely on fine-tuning RetinaNet50 on custom datasets and establishing a baseline. That work gave me a good sense of how well the model adapts to new data and what pitfalls to expect during training. With that fo...
In my last week’s blog, I had a working active learning pipeline: a loop where RetinaNet50 was trained on a labeled pool, predictions were run on an unlabeled pool, uncertain samples were selected, and the model was retrained with those added samples...
After spending the previous weeks building and testing an active learning pipeline (mostly inside Kaggle notebooks), I dedicated Weeks 9–10 to refinement and usability. This meant transforming rough experiments into something maintainable, reusable, ...
