直接使用yolov5 detect.py
量化后的int8模型
weight:'/mnt/bb/yc/yolov5/runs/train/exp63/weights/last_openvino_model'
soruce:'/mnt/bb/yc/datasets/pro_data4/images/train/82fb466dc9a61953bc4e1267ab3691aa.jpg'
method:yolov5-detect.py
未量化的f32模型
参数如下
weight:'/mnt/bb/yc/yolov5/runs/train/exp63/weights/best_openvino_model'
soruce:'/mnt/bb/yc/datasets/pro_data4/images/train/82fb466dc9a61953bc4e1267ab3691aa.jpg'
method:yolov5-detect.py
推理效果
未量化的f32的pt模型
参数如下
weight:'/mnt/bb/yc/yolov5/runs/train/exp63/weights/best.pt'
soruce:'/mnt/bb/yc/datasets/pro_data4/images/train/82fb466dc9a61953bc4e1267ab3691aa.jpg'
method:yolov5-detect.py
推理效果
总结
weight | pre-process | inference | nms |
---|---|---|---|
openvino_int8 | 8.4ms | 9.7ms | 2.5ms |
openvino_f32 | 7.0ms | 23.1ms | 3.1ms |
pytorch_f32 | 6.9ms | 63.4ms | 3.4ms |
检测效果
openvino_int8
openvino_f32
pytorch_f32