A look into DagsHub active learning pipelines

This recap covers another talk from YOLO VISION 2023 (YV23). At this Ultralytics-powered event, hosted at the Google for Startups campus in Madrid, Yono Mittlefehldt, former Machine Learning Advocate at DagsHub, gave a talk on building active learning pipelines.

Introduction and overview#

The talk opened with an introduction to active learning pipelines and the differences between active learning and traditional supervised learning methods.

Data preparation#

The first step lays the groundwork for the active learning pipeline: importing dependencies, setting up the data source, and enriching metadata with initial annotations.

Model training#

With the data ready, the next step is model training. Using an Ultralytics YOLOv8 dataset and its YAML file, Yono added callbacks to log parameters and metrics during training, making each run easy to track and compare.

Active learning cycle#

The next step is the active learning cycle, an iterative process that involves loading pre-trained models, scoring unlabeled data, and selecting the most informative samples for annotation. Each cycle enriches the data source with new predictions, so the model improves with every iteration.

Active learning for image segmentation#

Yono then applied active learning to image segmentation. Predictions were sent to Label Studio for annotation, showing how a segmentation model can improve over multiple cycles.

Using Label Studio#

Label Studio is a central tool in this pipeline. Projects store the annotated data, and Label Studio servers connect to the tasks API. Mapping tasks to project names keeps the workflow organized and makes collaboration easier.

Wrapping up#

As the talk wrapped up, Yono answered questions from the audience, covering topics from optimizing pipelines for specific tasks to reproducibility and documentation.

Overall, the talk showed how active learning can make data labeling and model training more efficient. We look forward to building on these ideas with the support and involvement of our community.

Watch the full talk on YouTube.