Question/Issue:
Looking to export a model**_vela.tflite** model directly from Edge Impulse without having to use a vela conversion tool.
Ideally, I’ll configure my model for the Ethos-U55-64 library as it states “vela default configuration”. However I want to export the model as a model_vela.tflite file instead of a C++ (.cpp & .h files) library. Is this possible?
The method I’m using in my project for uploading models is via XIP flash instead of embedding it inside the firmware, so I do not want a C++ library as the export.
I also want to understand the difference between the two different export regions within Edge Impulse (the usual “Deployment” region versus the “Download block output” from within Dashboard.
Within the dashboard, I can download the model specifically as a int8 quantized .lite (why is this not a .tflite file anymore?) file. If I download a model via the dashboard, has this model been formatted relative to my configurations from the Deployment page (i.e. is it vela optimized)?
Two images below are both export locations:
Project ID:
422782
Context/Use case:
I am developing a project using a Himax WiseEye 2 (HX6538) board. I am utilizing the onboard NPU so therefore need a vela model.
There are multiple methods that are compatible to format the model via Edge Impulse (C++ library, Grove Vision Ai V2, Ethos-U55-64 library) however they all get built and exported as a C++ library (i.e. gets embedded in the firmware image via .cpp & .h files) instead of exported as a .tflite (or better, _vela.tflite).
Expected Outcome:
model_vela.tflite
Actual Outcome:
Exported model is either a C++ library (.cpp & .h files) or a .lite file.
C++ isn’t useful as I’m enabling the model to get flashed via XIP to the device, not embedded in the firmware. And the .lite file requires post-processing to first convert it to a .tflite model and then to a vela model
Environment:
- Platform: Himax WiseEye 2 (HX6538)
- Build Environment Details: Ethos-U55-64 library
- OS Version: Windows 11
- Edge Impulse Version (Firmware): v1.66.12
- if you have a library deployment:
#define EI_STUDIO_VERSION_MAJOR 1
#define EI_STUDIO_VERSION_MINOR 71
#define EI_STUDIO_VERSION_PATCH 41 - Project Version: v20
- Custom Blocks / Impulse Configuration: Targeted for - Himax WiseEye2 (Corex-M55 400 MHz) + Ethos-U55-64