Jason Li

@SunshineJZJ train_models/mtcnn_model.py里有loss计算的过程

@MILOTIDE sorry the old link wasn't available, check this new link in your explorer: ftp://ftp.openailab.net.cn/tools/package/

Change ./detect/Makefile may solver your problem. It seems that you don't have caffe-hrt and arm computelibrary installed in /usr/local/AID. See readme for the target platform.

@sejal-gupta You may have the newer version of Proto-buf, use `protoc --version` to check. We recommend libprotoc 2.6.1

@sejal-gupta See the answer to @banqimiao question. As a reminder, install Caffe-HRT first can pretty mush solve all the dependency problems! [https://github.com/OAID/Caffe-HRT](url).

You can check your memory to see if the Pnet output too many boxes when the img is too big.

@1753939775 真正的问题在这儿呢， https://github.com/rockchip-linux/rknn-toolkit/blob/39d282457a5205550ff0ebea1bb0456d2a57b82b/examples/onnx/yolov5/test.py#L294 把transpose的参数从2，3，0，1改成正确的就行啦。这个其实跟你的rknn模型输出有关系，按照你rknn的输出去调这个transpose的参数，比如正常yolov5的onnx输出最后一个featuremap维度是1x3x20x20x85，那么经过它这个你只要调成20x20x3x85就好，但是有时候转出来的rknn模型输出的维度顺序对不上，比如1x85x3x20x20，那么参数就是2,3,1,0（第一个维度已经干掉了前面），网上还有很多1，2，0，3的，他们的RKNN模型输出肯定是1x3x20x20x85。

@rahulsharma11 Yes, see facecaffe/face_demo.cpp: std::string FaceDemo::Recognize(cv::Mat &frame, int face_num) you can see max number is set when it's called, and called here: facecaffe/AlgThread.cpp: line 60 (default set to 3) std::string...

@rahulsharma11 sorry, i forget. you should also change the default setting here: include/AlgThread.h (line 64, Mface m_face[3];) this should work.

这个有landmark训练吗？