????前面直接運(yùn)行測(cè)試.rknn完成。

????現(xiàn)在進(jìn)一步：生成.rknn并pc仿真運(yùn)行測(cè)試

????步驟：examples/pytorch/yolov5/README.md

1. ?git clone https://github.com/ultralytics/yolov5

2. git checkout c5360f6e7009eb4d05f14d1cc9dae0963e949213

3. 需要修改 yolov5/models/yolo.py文件的class Detect(nn.Module) 類的子函數(shù)forward
   

4. ?pip3 install?pandas(2.0.0太高了，我選擇==1.0.0，來降低依賴的numpy版本)

   注意rknn-toolkit 1.7.3 has requirement numpy==1.19.5

5. python export.py --weights yolov5s.pt --img 640 --batch 1 --include torchscript、

6. ln -s yolov5/yolov5s.torchscript.pt yolov5s.torchscript.pt

7. 修改test.py?

   仿真使用ret = rknn.init_runtime()

8. ?python3 test.py

   現(xiàn)在特意看了下系統(tǒng)監(jiān)視器，發(fā)現(xiàn)只有一個(gè)cpu核 100%

   
   

同樣yolov5s，

examples/onnx/yolov5/yolov5.onnx =>?yolov5.rknn 7.4MB

examples/pytorch/yolov5/yolov5s.torchscript.pt =>?yolov5s.rknn 14.6MB