nπ

RK3399pro is special, use this https://github.com/airockchip/RK3399Pro_npu

1. 如果浮点模型的结果就存在异常，大概率是 toolkit/ librknn_runtime 的bug。可以尝试更新到1.7.3版本 2. 如果更新到1.7.3仍未解决，可以上传模型，我们复现下问题

Yolov8-seg demo for RKNN is available here https://github.com/airockchip/rknn_model_zoo. The model output struct needs to be changed for quantization, which has been done in rknn_model_zoo demo.

You can get from here baidu cloud: https://eyun.baidu.com/s/3dHiqukh fetch code: rknn

Hello, Could you upload the test model and we can take a close look on how it happen? Thank you.

For yolo model, please refer here. https://github.com/airockchip/rknn_model_zoo/tree/main/models/vision/object_detection/

Please do the RKNN model generate step on PC(x86) but not edge device.

rv1126驱动版本建议更新到1.6.1及以上。 这边测试的参考结果是 yolov5s -> float_rknn，map 的情况是从0.339 -> 0.336； 使用量化的时候， yolov5s -> u8_rknn，map的情况是0.339 -> 0.319。 测试集使用的是 coco2017_val

Thanks for your report. Seems we got a bug here. Could you upload your test_onnx_model and we can reproduce this result and check it closely.

你好， toolkit1目前仅支持onnxruntime1.5.2生成的onnx量化模型，由于onnx量化功能目前不完善，推荐使用pytorch量化后导出为torchscript格式的模型再转rknn，目前pytorch支持到1.9.0（官方最新是1.10）