Detail:

• rockchip_yolonas_pth_to_onnx.py

  Modify From DeepStream Convert Tool + Cut Post process in YOLO-NAS ONNX Model

• export_yolonas_onnx_to_rknn.py

  Modify From https://github.com/MarcA711/rknn-models

How to use:

• Convert pth model to onnx (yolo_nas_s, input size 320x320, coco_label) ** You can down input size this step.

  Run: python rockchip_yolonas_pth_to_onnx.py -m yolo_nas_s -w last.pth --simplify -n 80 -s 320

• Convert pth model to onnx (yolo_nas_s, input size 640x640 downsize model to support 320x320, coco_label)

  Run: python rockchip_yolonas_pth_to_onnx.py -m yolo_nas_s -w last.pth --simplify -n 80 -s 320

  • m: Base Model (yolo_nas_s, yolo_nas_m, yolo_nas_l)
  • w: Weight Model
  • n: Number Classes
  • s: Input Size

• Convert onnx to rknn

  Run: python export_yolonas_onnx_to_rknn.py -t rk3588 -w rk_last.onnx -n 80 -s 320

  Run: python export_yolonas_onnx_to_rknn.py -t rk3588 -w rk_last.onnx -n 80 -s 320 -q true

  • t: SOC Type (rk3588, rk3566 etc.)
  • w: Weight Model (onnx format)
  • n: Number Classes (COCO = 80 Classes)
  • s: Input Size (320x320)
  • q: Quantize (True of False) If True, It's used dataset from datasets folder.

** If You used custom model on frigate. You need modify frigate file. ** Dont' forget install RKNN toolkit2 https://github.com/airockchip/rknn-toolkit2

Compare Graph before and after modify

Result from yolo_nas_s quantize input 640x640

Result from yolo_nas_s quantize input 320x320

Result from yolo_nas_s quantize input 224x224