Question/Issue:
My INT8 model performs correctly in Edge Impulse Model Testing, but after compiling/deploying it with Vela on the Himax WiseEye2 (Cortex-M55 + Ethos-U55-64), I observe a noticeable loss of accuracy.
Project ID: 1084313
Is this behavior expected?
From EI site : " Model optimizations can increase on-device performance but may reduce accuracy. Performance estimate for Himax WiseEye2 (Cortex-M55 400 MHz) + Ethos-U55-64 - [Change target]"
If the INT8 model accuracy is already good before Vela compilation, what could explain the degradation on the device?
Are there recommended settings or debugging steps to ensure that the Vela-compiled model produces results as close as possible to the INT8 Model Testing results?
Thanks !