dme-compunet/YoloV8

[discussion] optimize code logic to improve operational efficiency

itbencn opened this issue · 2 comments

  • reduce memory allocation
  • avoid double counting
  • reduce access overhead with new features like Span
  • .......

make a start #36

new optimization direction
The onnx output format before yolov8 was {1,8400,84}, while yolov8 used {1,84,8400}
When it is unfolded, this will be incoherent memory, and memory addressing will cause a significant loss of time, resulting in a performance loss of approximately 30% on read speed
We can use third-party tool code to convert YOLOv8's ONNX model to {1,8400,84} format, and our project(YOLOv8) only needs to be compatible with this format at the same time