Changelog

Protege Runtime release notes

v0.8.5 - 2026-04-03

Fixed

• Pin onnxsim<0.5 to use pre-built wheels (avoids cmake build requirement in Docker)

v0.8.4 - 2026-04-02

Fixed

• Replace opencv-python with opencv-python-headless to avoid requiring X11/GUI libraries (e.g. libxcb.so.1) in headless containers

v0.8.3 - 2025-10-24

Fixed

• Object detection List[Tensor] vs Tensor input handling
• Confidence threshold application in object detection
• Qt display issues in headless environments
• TorchScript backend object detection compatibility
• Improved SwinV2 model transform size handling for better model compatibility
• Enhanced model architecture detection and resize mapping functionality

Added

• Device selection (CPU/CUDA/MPS) for all example scripts
• Specific environment variables per script type
• Object detection image saving with small version

Changed

• Simplified example scripts for headless environments
• Updated .gitignore for auto-generated images

v0.8.2 - 2025-10-12

Added

• Multi-label classification

v0.8.1 - 2025-07-23

Added

• Add artifact property to Runtime to improve model metadata access during sweep functionality.
• Add object detection example (object_detection.py) demonstrating bounding box visualization and detection results display.

v0.8.0 - 2025-07-22

Added

• Added OcrPredictor class for efficient OCR inference with support for ZIP artifacts and batch predictions

v0.7.6 - 2025-07-06

Added

• Support for adding to num_classes in existing models

v0.7.5 - 2025-06-10

Dependencies

• Added common dependency in pyproject.toml to resolve classification issues

v0.7.4 - 2024-03-26

Added

• Internal improvements

v0.7.2

Added

• Initial externalization of protege-runtime from protege-ml, providing core execution and prediction logic for Protege-based models.

• Standardized Prediction Interfaces

  • Unifies how predictions are made, validated, and reported.

• Runtime Configuration

  • Flexible config loading for different runtime environments.

• Lightweight Design

  • Minimal dependencies for fast startup and ease of deployment.