Skip to content

v2.12.0

Choose a tag to compare

View all tags
@piotrgrubicki piotrgrubicki released this 21 Aug 13:53
· 287 commits to main since this release
v2.12.0
f123949
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Release Notes

Geti keeps evolving - and thanks to your feedback, version 2.12.0 introduces several improvements and new features to elevate your experience.

Summary of major features and improvements

  • Gold Support for Model Fine-Tuning on Intel Arc GPU: Succeeding the Beta Support from Geti™ 2.10.0 pre-release, you can now fine-tune models across the entire Geti model suite with performance on par with top-tier GPUs. Any Arc card will be supported (BMG, Alchemist), but we recommend using cards with 16GB VRAM or more. Even though 12GB cards will also be supported, a few large models may encounter issues during training.

  • Streamlined Advanced Training Settings: The model training interface has been completely updated with a more intuitive and cohesive design to select the model architecture and configure the various advanced training settings. In addition, new advanced parameters have been exposed:

    • Configurable data augmentation as part of ‘Data Management’ for classification tasks (will be applied to the other Geti tasks in upcoming releases), so that users can adjust the augmentation settings based on their needs;
    • Configurable model input size as part of ‘Training’, so that users can control the resolution at which images are processed for model fine-tuning.

  • Improved Media Upload Reliability: Various improvements to the media upload logic to make it faster and more reliable, including a better algorithm to detect corrupted videos and attempt repairing them on the fly.

  • Added support for key point detection dataset import and export in Datumaro format.

Changes to the REST API

For more details, we recommend you refer to the full API specification.

  • Added new endpoints to get and update the project and training configuration.
  • Deprecated the legacy configuration endpoints.
  • Updated the response of the supported algorithms endpoint.
  • Several fields have been added, removed or renamed.
  • Added a new query parameter to the project export endpoint to control which models to include in the exported zip archive (choice between all models, only the active one or none).
  • Removed a deprecated set of endpoints to download the models along with sample scripts (aka code deployment).

Model deprecations

The following model architectures are now deprecated. While it is still possible to use them in 2.12, users are recommended to switch to one of the newer, higher-performance model architectures.

  • Detection
    • ATSS-ResNeXt101
    • RTDetr-R18
    • RTDetr-R101
    • RTMDet-Tiny
  • Rotated detection
    • MaskRCNN-ResNet50-V1
  • Classification
    • EfficientNet-V2-L
    • MobileNet-V3-small
  • Instance Segmentation:
    • MaskRCNN-ResNet50-V1
  • Semantic Segmentation
    • LiteHRNet-X

Download this release

Installer

Download installer.

Container Images and Helm charts

Container images and Helm charts are available at ghcr.io/open-edge-platform/geti/*

Source code for open source components, provided as required by license.

Assets 2
Loading