人工智能开发系列(3) YOLOV3开发与实现

rknn_inference 20ms is for 416x416 tiny-yolov3, right? I measured 608x608 yolov3 inference time is about 440ms. Is the number correct? That's only 2.5 FPS.

hisping 发表于 2019-3-1 09:17
是的，yolov3-608耗时比较长，在移动设备端还是建议使用tiny-yolov3对视频进行检查。你可以使用tiny-yolo ...

试了一下608x608的Yolov2，rknn_inference()可以达到80ms。同样大小的Yolov3为什么会慢5倍这么多呢？是不是有些Yolov3的算子还没有优化好啊？

someone9388 发表于 2019-3-1 15:47
同问，
如果是安装在板子上，是否因该安装opencv-python-aarch64，但是pip 中的whl 只有3.3.0.1 的。这个 ...

sudo dnf install -y python3-devel python3-opencv python3-numpy-f2py python3-scipy python3-h5py python3-lmdb opencv-devel

Use dnf instead of pip. See the original post.

Tried Yolov3 416. All my results are as follow. Do they algin with RK's internal testing?

Yolov3 608: 440ms
Yolov3 416: 210ms
Yolov2 608: 80ms
Yolov3 tiny: 20ms

BTW, I met below warnings during running rknn_tranform.py. Does it mean the rknn model is not best optimized (i.e. some high efficiency OP's are ignored due to accuracy loss)?

--> Building model
W extend add_13 to add will cause accuracy loss, do not extend.
W extend add_23 to add will cause accuracy loss, do not extend.
W extend add_30 to add will cause accuracy loss, do not extend.
W extend add_40 to add will cause accuracy loss, do not extend.
W extend add_47 to add will cause accuracy loss, do not extend.
W extend add_54 to add will cause accuracy loss, do not extend.
W extend add_61 to add will cause accuracy loss, do not extend.
W extend add_68 to add will cause accuracy loss, do not extend.
W extend add_75 to add will cause accuracy loss, do not extend.
W extend add_82 to add will cause accuracy loss, do not extend.
W extend add_89 to add will cause accuracy loss, do not extend.
W extend add_99 to add will cause accuracy loss, do not extend.
W extend add_106 to add will cause accuracy loss, do not extend.
W extend add_113 to add will cause accuracy loss, do not extend.
W extend add_120 to add will cause accuracy loss, do not extend.
W extend add_127 to add will cause accuracy loss, do not extend.
W extend add_134 to add will cause accuracy loss, do not extend.
W extend add_141 to add will cause accuracy loss, do not extend.
W extend add_148 to add will cause accuracy loss, do not extend.
W extend add_158 to add will cause accuracy loss, do not extend.
W extend add_165 to add will cause accuracy loss, do not extend.
W extend add_172 to add will cause accuracy loss, do not extend.
W extend add_179 to add will cause accuracy loss, do not extend.
done

程子 发表于 2019-3-12 15:10
输出节点只有一个，而你试图把它赋值给两个变量。

Yeah, yolov3 has 3 dimensions output while yolov2 has only one.