Skip to main content

Changelog

Protege Pipelines release notes

v0.8.4 - 2025-06-26​

Fixed​

  • Fixed SwinV2 classification bug:
    • Corrected resize ratio calculation for SwinV2 models to ensure proper image scaling
    • Fixed input size handling for different SwinV2 variants (192x192, 224x224, 256x256, 384x384)
    • Ensured consistent image preprocessing across all SwinV2 architectures

Added​

  • Improved image transformations:
    • Added support for multiple SwinV2 model input sizes (192x192, 224x224, 256x256, 384x384)
    • Implemented dynamic resize and crop calculations based on model architecture
  • Added classification model export to GAR:
    • Added automatic Docker image creation and versioning in Google Artifact Registry (GAR)
    • Implemented model packaging as a Docker container for easy deployment
    • Configured versioned image tags for model tracking and rollback
  • Enhanced configuration validation:
    • Implemented comprehensive TOML parser tests for classification configurations
    • Added validation for all required fields and their correct formats
    • Ensured proper error handling for invalid configurations

v0.8.3 - 2025-06-24​

Added​

  • Added child components for the parent run
  • Added separate Oleander RUNNING events for each epoch level metrics

v0.8.2 - 2025-06-24​

Added​

  • Added support for encord image groups
  • Added support to train models from gcs as source for labelled data
  • Added build and publish commands to makefile

v0.8.1 - 2025-06-24​

Added​

  • Allow automatically converting polygon annotations to keypoints

Updated​

  • protege-engine from 0.14.6 to 0.15.1 in order to support automatically converting polygon annotations to keypoints

v0.8.0 - 2025-06-18​

Added​

  • OpenLineage integration in protege-pipelines

v0.7.0 - 2025-06-16​

Added​

  • phoenix-kube into this repo, so there is one less external dependency
  • cloud provider service account configuration option

v0.6.5 - 2025-05-30​

  • Updated dependencies

v0.6.4 - 2025-05-30​

Changed​

  • Fetch secrets from .env file and bake into vertex piepline components as input parameters.

v0.6.3 - 2025-05-21​

Changed​

  • Pulled secrets, api endpoint and base image urls from env file instead of hard‑coding.
  • Pulled configuration from toml file instead of hard-coding.

v0.6.2 - 2024-04-17​

Added​

  • Internal improvements

v0.6.0​

Added​

  • Initial externalization of protege-pipelines from protege-ml, providing end-to-end orchestration for model training workflows.

  • Job Spec-Driven Training

    • Declarative TOML configuration to define training runs.

  • Multi-Task Support

    • Built-in logic for classification, detection, segmentation, and keypoint tasks.

  • Cloud-Oriented Design

    • Integration with Vertex AI Pipelines and GCS for training and export.

  • CLI Entry Point

    • Simple CLI to trigger the entire pipeline with caching and validation support.