A Look Into DagsHub Active Learning Pipelines

Step into the realm of cutting-edge Artificial Intelligence (AI) methodologies with another one of our speakers from YOLO VISION 2023 (YV23)! At this Ultralytics-powered event, hosted at the Google for Startup campus in Madrid, Yono Mittlefehldt, former Machine Learning Advocate at DagsHub, took the stage to unravel the wonders of active learning pipelines. 

Introduction and Overview

To kick off our journey, let's set the stage with an introduction to active learning pipelines. In this talk, we looked at the differences between active learning and traditional supervised learning methods.

Data Preparation

Our first stop involves laying the groundwork for our active learning pipeline. We import dependencies, set up the data source, and embark on a mission to enrich metadata with initial annotations. It's all about preparing the foundation for our AI-powered exploration.

Model Training

With the data prepped and ready, we dive into the exciting realm of model training. With the Ultralytics YOLOv8 dataset and YAML file, Yono added callbacks to log parameters and metrics during training. This is a crucial step in ensuring the AI models are primed for success.

Active Learning Cycle

The next step is the active learning cycle – a dynamic process that involves loading pre-trained models, scoring unlabelled data, and selecting samples for annotation. Through iterative enrichment of the data source with predictions, we uncover hidden insights and propel the models to new heights.

Active Learning for Image Segmentation

Image segmentation takes center stage as we explore the transformative power of active learning. By sending predictions to Label Studio for annotation, we understand the potential for model improvement through multiple cycles. It is a journey of discovery, where each iteration brings us closer to AI perfection.

Using Label Studio

In our quest for AI excellence, Label Studio emerges as an important tool in our arsenal. We create projects to store annotated data, leveraging Label Studio servers to connect with tasks API seamlessly. With tasks mapped to project names, we streamline our workflow and pave the way for smoother collaboration.

Wrapping Up

As the talk wrapped up, Yono addressed burning questions from our audience. From optimizing pipelines for specific tasks to emphasizing reproducibility and documentation, he ensured that every aspect of this journey is grounded in best practices and industry standards.

Overall, this journey through active learning at YV23 has been nothing short of exhilarating. Armed with newfound knowledge and insights, we're ready to embark on new AI adventures, fueled by the power of active learning as well as the support and involvement of our community.

Join us as we continue to push the boundaries of AI innovation and redefine what's possible in the world of machine learning. Watch the full talk here!