Skip to content

Release Notes

April 1, 2025 - LEIP Design 1.3

Updates and Enhancements

  • RT-DETRv2 Support: You can now use RT-DETRv2 in object detection workflows for improved detection capabilities.

  • D-FINE Model Family Integration: The D-FINE model family is now fully integrated, enhancing flexibility and performance.

  • Enhanced ONNX Support: You can now configure ONNX export settings and specify ONNX runtime for improved backend compatibility.

  • FiftyOne Dataset Integration: You can now integrate classification datasets using attach_fiftyone_data_generator for seamless data ingestion.

  • Performance Improvements for YOLOv8 and YOLOv10 models.

Bug Fixes

  • Fixed Unnecessary Archive Extraction: Certain archives will no longer be unzipped unnecessarily.

  • Resolved Task Execution Errors: Fixed an issue causing empty ingredient slot errors.

  • Improved Dataset Ingestion: Unlabeled datasets are now handled more robustly.

January 31, 2025 - LEIP Design 1.2

Updates and Enhancements

  • Full Image Classification Support: You can now leverage the entire LEIP Design workflow for image classification tasks, from data ingestion to training, optimization, and deployment.

  • Classifier GRDB: You can now access a new set of optimized and validated Golden Recipes for image classification, along with an end-to-end how-to guide detailing its usage.

  • Stub Code Generator now supports batch sizes greater than one.

  • Offline Mode API: API reference for offline mode has been expanded and improved.

  • Classification Data Generator: API reference for classification data generator has been expanded and improved.

  • Voxel FiftyOne Data Integration Guide: A new how-to guide shows you how to visualize data, manage data versions, filter noisy samples, and ingest refined datasets into a LEIP Recipe.

December 10, 2024 – LEIP Design 1.1

Updates and Enhancements

  • Offline Mode: You can now pre-download and cache all of the artifacts needed to run recipes without network connectivity.

  • Improved Security: LEIP Design 1.1 passes U.S. Department of Defense (DoD) Impact Level 5 (IL5) security scans.

November 25, 2024 - LEIP Design 1.0

Updates and Enhancements

  • Introducing LEIP Design, a core LEIP Onsite SDK capability that allows you to orchestrate your entire machine learning workflow, from data ingestion and model development to optimization and deployment, all within a unified environment.

  • Flexible Installation: Install LEIP Design via Docker, conda, or pip.

  • Voxel FiftyOne Integration: Directly ingest any dataset supported by the Voxel FiftyOne API into LEIP Design.

  • Advanced Data Augmentations with MixUp: Implement MixUp augmentation for classifiers to enhance model generalization and performance by blending samples during training.

  • Custom Input Channels for Classifiers: You can now train classifiers from scratch with images that have a different number of channels.

  • Support for New Model Architectures: LEIP Design supports YOLOv10 for advanced object detection and RT-DETR for efficient, real-time object detection tasks.

  • Expanded Classification Dataset Support: Access and train on five new classification datasets:

    • CIFAR-100
    • FGVS Aircraft
    • EuroSAT
    • GTSRB (German Traffic Sign Recognition Benchmark)
    • Stanford Cars

  • Enhanced Classification Metrics: Configure and evaluate models using new classification metrics through the Recipe Designer API:

    • AUROC (Area Under the Receiver Operating Characteristic Curve)
    • Precision
    • Recall
    • F1-Score

  • LEIP Stub Code Generator: Simplify deployment to the edge by generating all necessary code during the recipe design process.

  • Golden Recipe Database (GRDB) API Enhancements: The GRDB now suggests optimized, tested, and validated Golden Recipes, with compiler and stub code generation parameters, to further streamline the machine learning workflow through optimization to deployment.