Train the model .pt or configure arquitecture of Yolov8n to run RKNN 2.3 int8 segmentation

[HyAgOsK]

Hello, community. How are you all doing? I need some assistance with updating to RKNN 2.3. I was running an application using RKNN 1.6, but I need to update as it is no longer supported. I need to understand how to convert my YOLO ONNX model to RKNN, but I’m experiencing shape-related issues. I believe it’s not just a preprocessing or conversion issue but rather a problem with the model itself.

I noticed that the YOLOv8n.onnx model provided as an example for conversion in the given link has been modified. I need guidance on how to address this and successfully convert and use the model with RKNN 2.3.

When using RKNN 1.6, the performance was very low, around 10-20% utilization of the NPU on the 3588s platform, resulting in poor performance. I am currently using a YOLOv8n segmentation model and urgently need help to resolve these issues.

I’ve already updated the RKNN SDK to 2.6, trained and exported the model, but I’m encountering shape-related errors. I’m unsure if the issue lies with the model I trained or if it’s related to how the YOLOv8n architecture should be modified for proper ONNX export. I need to understand how to adjust the YOLOv8n architecture, as it was done for the provided example, to export it correctly.

https://github.com/airockchip/rknn_model_zoo/tree/main/examples/yolov8_seg

This is an urgent matter, and I would greatly appreciate any assistance or guidance you can provide. Thank you!