Skip to content

基础 AI 推理示例

YOLOv5 目标检测(官方示例)

bash
# 进入官方示例目录
cd /usr/local/share/aidlite/examples/aidlite_qnn236/cpp
# DSP加速运行(推荐,参数3对应.bin模型,参数4对应.so模型)
LD_PRELOAD=/usr/lib/libstdc++.so.6.0.30 ./qnn_yolov5_multi 3

预期输出

bash
classid: 0, score 0.812817, pos ([115.072533], [236.646347]) ([202.021912], [541.845459])
classid: 0, score 0.789807, pos ([211.840363], [246.274216]) ([283.799805], [514.897522])
classid: 0, score 0.785955, pos ([472.021393], [233.328232]) ([561.091858], [520.786987])

Python 版 YOLOv5 推理

bash
# 由于python测试例子会创建文件,但是此时的/分区是禁用读写的
cp /usr/local/share/aidlite/examples/aidlite_qnn236 -r /data
cd /data/aidlite_qnn236/python
LD_PRELOAD=/usr/lib/libstdc++.so.6.0.30 python3 qnn_yolov5_multi.py 3

验证:查看生成的qnn_yolov5_multi_pil.jpg,确认检测框位置正确。