HumanSignal/label-studio-ml-backend

What it solves

This project provides an SDK that allows developers to wrap machine learning models and deploy them as web servers. These servers can then be connected to Label Studio to automate labeling tasks, reducing the manual effort required for data annotation.

How it works

The SDK provides a base class, LabelStudioMLBase, which developers can inherit from to implement their own inference and training logic. By overriding the predict method, the model can provide pre-annotations or interactive predictions for tasks. Optionally, the fit method can be used to train the model on labeled data as it is submitted.

Who it’s for

Machine learning engineers and data scientists who use Label Studio for data labeling and want to integrate their own custom ML models to assist in the annotation process.

Highlights

• Pre-built Model Examples: Includes a wide range of supported models for text classification, NER, OCR, object detection, image segmentation (SAM 2), and RAG pipelines (LangChain).
• Flexible Deployment: Supports deployment via Docker Compose or direct deployment to Google Cloud Platform (GCP).
• Interactive Labeling: Enables models to provide real-time predictions while a user is labeling data.
• Training Integration: Allows models to be updated based on submitted annotations via the fit method.