Releases · antmicro/kenning

For installation instructions, see the README or Gallery of use cases

Changes:

Introduced support for tinygrad models
Introduced support for ExecuTorch runtime (including Kenning Zephyr Runtime support)
Introduced Converter and conversion registry for model formats
Improved integration with ROS 2 - added possiblity to serve models as ROS 2 services from Kenning level (see ROS 2 GUI Node demos)
Added command for downloading remote resources
Improved AutoML flow deploying models using microTVM runtime
Introduced possibility to customize target metrics for AutoML tasks
Introduced two-way communication with Kenning Zephyr Runtime's evaluation app, allowing passing logs and Kenning-specific data using single UART connection
Introduced examples for models running instance segmentation, depth estimation, pose estimation
Added possibility to define custom report flows
Introduced new quality metrics for anomaly detection tasks (regarding speed of detection)

For installation instructions, see the README or Gallery of use cases

Changes:

Introduced improvements when passing IO and model data between Optimizer and Runtime blocks
Improved and sped up parsing of Renode profiler results
Introduced SparseGPT, GPTQ and SparseGPT+GPTQ optimizers for LLMs
Added ModelWrappers for Phi-2 and Mistral 7B models
Added example of summarization dataset
Introduced support for LLEXT module building for Kenning Zephyr Runtime
Introduced anomaly detection dataset handler for CSV files (on example of CATS dataset)
Introduced VAE model
Introduced Platform class for holding data necessary for Optimizers and HW-aware optimizations
Introduced AutoML class for running automatic model architecture search
Introduced AutoPyTorch-based AutoML implementation
Introduced gallery examples:
- Anomaly detection model deployment on MCU
- AutoML-based search for anomaly detection models
Added support for YAML-based scenarios
Minor fixes in drawing and reporting modules

Provide feedback