search

rockchip-linux / rknn-toolkit2

BSD 3-Clause "New" or "Revised" License
875 stars 154 forks source link

Yolov5s conversion works with rknn_toolkit2, but not with default ONNX export parameters #142

Open neocoretechs opened 1 year ago

neocoretechs commented 1 year ago

Using the yolov5s.onnx supplied in the toolkit demo conversion results in a working rknn model, but using default export.py from ultralytics results in the one layer model that selects all possible bounding boxes.

Mitchelldscott commented 1 year ago

I am currently struggling with this issue as well. @neocoretechs did you find a work around? I was hoping it was just setting the right opset but it looks like the .onnx in this repository was not made with that export (or maybe an older version). I'm using YOLOv5 🚀 v7.0-145-g94714fe Python-3.8.10 torch-1.10.1+cu111 Ubuntu 20.04 and onnx 1.9.0.

Mitchelldscott commented 1 year ago

Update: Khadas forked the ultralytics/yolov5 repo and that script worked. https://github.com/rockchip-linux/rknpu2/issues/57