As you have C++ runtime and driver support for Yocto (I’ll publish a pull request to fix the recipes), could you make Python runtime and support modules also available separately? pip install generally brings in main onnxruntime, tflite and supporting dependencies that often don’t run on specific targeted ARM64 cores. Most of these dependencies, for NXP SoCs, have forks.
Could you supply PyPi package or a no-dependency wheel to install runtime and MX3 Python API modules? It would be nice to list the requirements however.
Thank you for raising the PR to the Yocto repository. The public repo we have today is a boilerplate recipe — we are still in the process of cleaning it up and aligning it with our latest SDK release.
As for the Python runtime wheel without dependencies:
We are actively working on a runtime-only pip package, which we’ll make available with the next SDK release, coming in a month or so.
Currently, our runtime library depends on ONNX Runtime, TensorFlow, and TFLite libraries to automatically handle pre/post-processing when a user connects cropped portions of models. (like with set_postprocessing_model )
If you’re comfortable handling the cropped portions yourself, like using numpy or other modules, we have a workaround available:
Workaround: We recently integrated MemryX hardware with the open-source VMS software Frigate. To reduce dependencies in the final Frigate Docker image, we created a lightweight MemryX Python runtime-only repository: mx_accl_frigate repo
You can use this Python runtime code, set up your environment, and benchmark it for now.
We’ll keep you updated as we release the no-dependency Python wheel option soon.