chenyujing1234 opened a new issue #13111: The same code in MXNet 1.2.0 and MXNet 1.3.0, but the reasoning results are different URL: https://github.com/apache/incubator-mxnet/issues/13111 When using mtcnn model for face detection, it is found that the reasoning result of MXNet 1.2.0 is consistent with that of TensorFlow, which is correct, but The same code in MXNet 1.2.0 and MXNet 1.3.0 inference results are different, but the comparison into the inference data is the same. Code Description: I use c++ API to recommend the main function (c_predict_api.h): MXPredCreate > MXPredSetInput > MXPredForward > MXPredGetOutputShape > MXPredGetOutput ============================== Explain: 1) code address: https://github.com/deepinsight/mxnet-mtcnn/tree/master 2) this phenomenon appears as probability, testing 50 thousand diagrams, and about 10% of the pictures are defective. 3) the algorithm mtcnn used is different in the P stage. For 124.jpg pictures, Pyramid pictures (37, 37) big hours, reasoning results are different. [http://note.youdao.com/noteshare?id=ed4ec6f5ad2fc7cc164b0c28c86f3b84](url) 4) model: Https://github.com/deepinsight/mxnet-mtcnn/tree/master/models/det-0001.params Https://github.com/deepinsight/mxnet-mtcnn/tree/master/models/det-symbol.json `void MxNetMtcnn::RunPNet(const cv::Mat& img, scale_window& win, std::vector<face_box>&box_list) { cv::Mat resized; int scale_h=win.h; int scale_w=win.w; float scale=win.scale; LoadPNet(scale_h,scale_w); cv::resize(img, resized, cv::Size(scale_w, scale_h), 0, 0, cv::INTER_LINEAR); std::vector<float> input(3*scale_h*scale_w); std::vector<cv::Mat> input_channels; set_input_buffer(input_channels, input.data(), scale_h, scale_w); cv::split(resized, input_channels); // 这里对比了输入的数据，都是一样的。 MXPredSetInput(PNet_,"data",input.data(),input.size()); MXPredForward(PNet_); mx_uint *shape = NULL; mx_uint shape_len = 0; MXPredGetOutputShape(PNet_,0,&shape,&shape_len); int reg_size=1; for(unsigned int i=0;i<shape_len;i++) reg_size*=shape[i]; MXPredGetOutputShape(PNet_,1,&shape,&shape_len); int confidence_size=1; for(unsigned int i=0;i<shape_len;i++) confidence_size*=shape[i]; std::vector<float> reg(reg_size); std::vector<float> confidence(confidence_size); // 下面两步出来的结果是错的。 MXPredGetOutput(PNet_,0, reg.data(), reg_size); MXPredGetOutput(PNet_,1, confidence.data(), confidence_size); std::vector<face_box> candidate_boxes; int feature_h=shape[2]; int feature_w=shape[3]; generate_bounding_box(confidence.data(),confidence.size(), reg.data(), scale, pnet_threshold_, feature_h, feature_w, candidate_boxes,false); nms_boxes(candidate_boxes, 0.5, NMS_UNION,box_list); }`
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
