Prepare Target AI Model

AI Models must be compiled using the Renesas RUHMI Framework AI Compiler before they can be evaluated on the RZ/G3E SMARC EVK.

The RUHMI Framework AI Compiler is a software tool that compiles a TensorFlow Lite model into an optimized version that can be run on the Ethos-U55 NPU using the EdgeCortix MERA runtime library.

The compiled model contains the Ethos-U custom operator (ethos-u) for parts of the model that can be executed on the NPU. Parts of the model that are not supported by the NPU remain in the model and are ran on the Arm Cortex-M33 CPU.

The automated compilation script generate-model-data.py allows for multiple models to be compiled and randomised input data to be generated in one run. This is packaged into a single directory that can be copied to the RZ/G3E SMARC EVK.

Note

The software package supports models where the compiler has allocated up to 4MB for the TensorFlow Lite Micro Tensor Arena.

Any models with an allocation above 4MB are not supported.

The Tensor Arena is used for storing the inputs, outputs and intermediate values during inferences.