Skip to main content

Changelog

Protege Runtime release notes

v0.15.5 - 2025-06-27​

Added​

  • Add required common dependency for model compatibility
  • Enhance metrics tracking capabilities for classification tasks
  • Improve test coverage for configuration and accuracy metrics
  • Add comprehensive test suite for model configuration
  • Add top-k accuracy metrics functionality

Fixed​

  • Ensure better reliability in classification tasks
  • Fix configuration loading with various scenarios
  • Validate metrics serialization and logging

v0.15.4 - 2025-06-26​

Added​

  • Add build dev target to makefile
  • Publish image to development registry

v0.15.3 - 2025-06-24​

Added​

  • Allow passing custom Lightning callbacks into train_model (via callbacks parameter) and apply them at runtime.
  • Support IoU computation for both object-detection and instance-segmentation tasks in metrics.py.
  • Log train_loss and train_grad_norm on each epoch in the Lightning training wrapper.

v0.15.2 - 2025-06-24​

Added​

  • Added protege-telemetry(for Otel to gcm) and protege-common(to remove dependency from protege-runtime) as a dependency in protege-engine

v0.15.1 - 2025-06-24​

Added​

Allow automatically converting polygons to keypoint coordinates

v0.15.0 - 2025-06-24​

Fixed​

  • Removed unnecessary circular dependency

v0.14.9 - 2025-06-18​

Fixed​

  • Update the pyproject.toml to support the latest versions of protege-pipelines and protege-runtime

v0.14.8 - 2025-06-18​

Added​

  • Add protege-lineage as a dependency in protege-engine

v0.14.7 - 2025-05-29​

Added​

  • Add .tool-versions with python 3.11.12

Changed​

  • Lock dependencies versions
  • Use logging instead of print in scheduler to show the last learning rate

Removed​

  • Deleted local models directory, all models should be consumed from protege-common
  • Remove deprecated argument verbose from torch.optim.lr_scheduler.ReduceLROnPlateau

Fixed​

  • Fix missing dependencies on image build
  • Force image platform version to linux/amd64 as osx defaults to arm64

v0.14.6 - 2024-03-26​

Added​

  • Internal improvements

v0.14.2​

Added​

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

  • Training Components

    • Lightning-based trainer with export, logging, and scheduler utilities.
    • Dataset and manifest management tools.

  • Model Zoo

    • Support for classification, detection, segmentation, and keypoint models via TorchVision and custom wrappers.

  • CLI Interface

    • protege-package CLI for packaging models and artifacts.

  • Dockerized Runtimes

    • CUDA-enabled training runtime via docker/Dockerfile.vertex.
    • Release-focused Docker image with direct install from source.

  • CI/CD Workflow

    • Unit test matrix, GitHub release automation, Python wheel + Docker publishing, and documentation syncing.