Skip to content

3.6.0
Compare
Choose a tag to compare
@deci-services deci-services released this 25 Jan 14:25
· 82 commits to master since this release
3.6.0
1550a77

Hey @channel
We're pleased to announce the release of Super-Gradients 3.6.0 🙂 .
This update includes several important changes and improvements:
Changes and Enhancements

  • Added segmentation samples and support for albumentation transforms for segmentation, contributed by @shaydeci.
  • Implemented distance-based detection matching in DetectionMetrics as an enhancement by @DimaBir.
  • New training hyperparameter - finetune, and multiple LR assignment by @shaydeci, read about it here
  • Enhanced ImagePermute processing inclusion, by @BloodAxe.
  • Improved dataset plotting and plot functionality, by
    @Louis Dupont
    .
  • Updated prediction notebooks and documentation, thanks to
    @Louis Dupont
    .
  • Supported depth estimation dataset and added depth estimation metrics support, by @shaydeci.
  • Proposed an API for checking model input compatibility, by @BloodAxe.
  • Extended predict() support for segmentation models, by @BloodAxe.
  • Removed deprecated features from version 3.6.0, by @shaydeci.
  • Updated pre-trained models badge URL, contributed by @gasparitiago.
  • Made changes to PPYoloELoss, removing the requirement for a reg_max parameter, by @BloodAxe.
  • Switched to using onnxsim instead of onnx-simplifier for consistency in naming, thanks to @BloodAxe.
    Bugfixes
  • Resolved a bug in OhemLoss thanks to @danielafrimi.
  • Updated conditions to ensure functionality only on rank 0 where [context.sg](http://context.sg/)_logger is available, by @shaydeci.
  • Modified the default set_device value to prevent unintentional launch of DDP, updated by
    @Louis Dupont
    .
  • Addressed a bug where multigpu=None with device=cpu wasn't functioning as expected, thanks to
    @Louis Dupont
    .
  • Adjusted bounding box thickness and text size relative to bbox size, for object detection model's predict() by
    @Louis Dupont
    .
  • Addressed a bug in DetectionMixup that affected YoloXTrainingStageSwitchCallback, by @BloodAxe.
  • Corrected a typo in an exception message variable name, by @BloodAxe.
  • Reintegrated tests and refined CI/CD workflows, thanks to @shaydeci and @Yonatan-Kaplounov.
  • Fixed import issues and improved model flexibility and metrics handling, mainly by @BloodAxe.
  • Ensured class names in DetectionDataset are contained within a trivial container, by @BloodAxe.
  • Fixed ExtremeBatchDetectionVisualizationCallback for multiscale collate function, by @BloodAxe.
  • Several bug fixes and improvements in DistanceBasedDetectionMetrics and DetectionMetrics, by @BloodAxe.
    And various other fixes and improvements across the board to enhance functionality and user experience.
    For a detailed list of changes, refer to the full changelog.
    New Contributors
  • Welcome to the contributors' team, @DimaBir, @gasparitiago, and @Yonatan-Kaplounov, for making their first contributions.Hey @channel
    We're pleased to announce the release of Super-Gradients 3.6.0 🙂 .
    This update includes several important changes and improvements:
    Changes and Enhancements
  • Added segmentation samples and support for albumentation transforms for segmentation, contributed by @shaydeci.
  • Implemented distance-based detection matching in DetectionMetrics as an enhancement by @DimaBir.
  • New training hyperparameter - finetune, and multiple LR assignment by @shaydeci, read about it here
  • Enhanced ImagePermute processing inclusion, by @BloodAxe.
  • Improved dataset plotting and plot functionality, by
    @Louis Dupont
    .
  • Updated prediction notebooks and documentation, thanks to
    @Louis Dupont
    .
  • Supported depth estimation dataset and added depth estimation metrics support, by @shaydeci.
  • Proposed an API for checking model input compatibility, by @BloodAxe.
  • Extended predict() support for segmentation models, by @BloodAxe.
  • Removed deprecated features from version 3.6.0, by @shaydeci.
  • Updated pre-trained models badge URL, contributed by @gasparitiago.
  • Made changes to PPYoloELoss, removing the requirement for a reg_max parameter, by @BloodAxe.
  • Switched to using onnxsim instead of onnx-simplifier for consistency in naming, thanks to @BloodAxe.
    Bugfixes
  • Resolved a bug in OhemLoss thanks to @danielafrimi.
  • Updated conditions to ensure functionality only on rank 0 where [context.sg](http://context.sg/)_logger is available, by @shaydeci.
  • Modified the default set_device value to prevent unintentional launch of DDP, updated by
    @Louis Dupont
    .
  • Addressed a bug where multigpu=None with device=cpu wasn't functioning as expected, thanks to
    @Louis Dupont
    .
  • Adjusted bounding box thickness and text size relative to bbox size, for object detection model's predict() by
    @Louis Dupont
    .
  • Addressed a bug in DetectionMixup that affected YoloXTrainingStageSwitchCallback, by @BloodAxe.
  • Corrected a typo in an exception message variable name, by @BloodAxe.
  • Reintegrated tests and refined CI/CD workflows, thanks to @shaydeci and @Yonatan-Kaplounov.
  • Fixed import issues and improved model flexibility and metrics handling, mainly by @BloodAxe.
  • Ensured class names in DetectionDataset are contained within a trivial container, by @BloodAxe.
  • Fixed ExtremeBatchDetectionVisualizationCallback for multiscale collate function, by @BloodAxe.
  • Several bug fixes and improvements in DistanceBasedDetectionMetrics and DetectionMetrics, by @BloodAxe.
    And various other fixes and improvements across the board to enhance functionality and user experience.
    For a detailed list of changes, refer to the full changelog.
    New Contributors
  • Welcome to the contributors' team, @DimaBir, @gasparitiago, and @Yonatan-Kaplounov, for making their first contributions..

Contributors

BloodAxe, gasparitiago, and 5 other contributors
Assets 2
Loading