From 08fd79f3bbb8457ad38da7e9ac7db410711d3dc3 Mon Sep 17 00:00:00 2001 From: ChrisKong <609027949@qq.com> Date: Tue, 7 Jan 2020 15:05:00 +0800 Subject: [PATCH] =?UTF-8?q?=E6=9B=B4=E6=94=B9=E6=A8=A1=E5=9E=8B=E5=8F=82?= =?UTF-8?q?=E6=95=B0=E7=BB=93=E6=9E=84?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 更改模型参数结构 --- .gitignore | 3 +- src/facenet.cpp | 368 ++++++++++++++++++++++++++++++++---------------- src/network.cpp | 364 ++++++++++++++++++++++++++++++++++++++++++++++- src/network.h | 20 ++- 4 files changed, 624 insertions(+), 131 deletions(-) mode change 100644 => 100755 src/facenet.cpp mode change 100644 => 100755 src/network.cpp mode change 100644 => 100755 src/network.h diff --git a/.gitignore b/.gitignore index 2214cc9..c8a5d5c 100644 --- a/.gitignore +++ b/.gitignore @@ -35,4 +35,5 @@ .idea/ cmake-build-debug/ cmake-build-release/ -model_128/ \ No newline at end of file +model_128/ +model_512/ diff --git a/src/facenet.cpp b/src/facenet.cpp old mode 100644 new mode 100755 index 4aee9c4..0c84211 --- a/src/facenet.cpp +++ b/src/facenet.cpp @@ -4,7 +4,11 @@ #include "facenet.h" - +/** + * stem网络 + * @param image 输入图片 + * @param output 输出featuremap 指针形式 + */ void facenet::Stem(Mat &image, pBox *output) { pBox *rgb = new pBox; pBox *conv1_out = new pBox; @@ -47,27 +51,35 @@ void facenet::Stem(Mat &image, pBox *output) { struct BN *conv6_beta = new BN; long conv1 = ConvAndFcInit(conv1_wb, 32, 3, 3, 2, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, 32); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, 32); long conv2 = ConvAndFcInit(conv2_wb, 32, 32, 3, 1, 0); - BatchNormInit(conv2_var, conv2_mean, conv2_beta, 32); + BatchNormInit(conv2_beta, conv2_mean, conv2_var, 32); long conv3 = ConvAndFcInit(conv3_wb, 64, 32, 3, 1, 1); - BatchNormInit(conv3_var, conv3_mean, conv3_beta, 64); + BatchNormInit(conv3_beta, conv3_mean, conv3_var, 64); long conv4 = ConvAndFcInit(conv4_wb, 80, 64, 1, 1, 0); - BatchNormInit(conv4_var, conv4_mean, conv4_beta, 80); + BatchNormInit(conv4_beta, conv4_mean, conv4_var, 80); long conv5 = ConvAndFcInit(conv5_wb, 192, 80, 3, 1, 0); - BatchNormInit(conv5_var, conv5_mean, conv5_beta, 192); + BatchNormInit(conv5_beta, conv5_mean, conv5_var, 192); long conv6 = ConvAndFcInit(conv6_wb, 256, 192, 3, 2, 0); - BatchNormInit(conv6_var, conv6_mean, conv6_beta, 256); + BatchNormInit(conv6_beta, conv6_mean, conv6_var, 256); long dataNumber[24] = {conv1, 32, 32, 32, conv2, 32, 32, 32, conv3, 64, 64, 64, conv4, 80, 80, 80, conv5, 192, 192, 192, conv6, 256, 256, 256}; - mydataFmt *pointTeam[24] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ - conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ - conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ - conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ - conv5_wb->pdata, conv5_var->pdata, conv5_mean->pdata, conv5_beta->pdata, \ - conv6_wb->pdata, conv6_var->pdata, conv6_mean->pdata, conv6_beta->pdata}; +// mydataFmt *pointTeam[24] = { +// conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ +// conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ +// conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ +// conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// conv5_wb->pdata, conv5_var->pdata, conv5_mean->pdata, conv5_beta->pdata, \ +// conv6_wb->pdata, conv6_var->pdata, conv6_mean->pdata, conv6_beta->pdata}; + mydataFmt *pointTeam[24] = { + conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata, \ + conv2_wb->pdata, conv2_beta->pdata, conv2_mean->pdata, conv2_var->pdata, \ + conv3_wb->pdata, conv3_beta->pdata, conv3_mean->pdata, conv3_var->pdata, \ + conv4_wb->pdata, conv4_beta->pdata, conv4_mean->pdata, conv4_var->pdata, \ + conv5_wb->pdata, conv5_beta->pdata, conv5_mean->pdata, conv5_var->pdata, \ + conv6_wb->pdata, conv6_beta->pdata, conv6_mean->pdata, conv6_var->pdata}; string filename = "../model_" + to_string(Num) + "/stem_list.txt"; readData(filename, dataNumber, pointTeam, 24); @@ -78,21 +90,19 @@ void facenet::Stem(Mat &image, pBox *output) { convolutionInit(conv1_wb, rgb, conv1_out); //conv1 149 x 149 x 32 convolution(conv1_wb, rgb, conv1_out); -// printData(conv1_out); - BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); -// printData(conv1_out); + BatchNorm(conv1_out, conv1_beta, conv1_mean, conv1_var); relu(conv1_out, conv1_wb->pbias); convolutionInit(conv2_wb, conv1_out, conv2_out); //conv2 147 x 147 x 32 convolution(conv2_wb, conv1_out, conv2_out); - BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); + BatchNorm(conv2_out, conv2_beta, conv2_mean, conv2_var); relu(conv2_out, conv2_wb->pbias); convolutionInit(conv3_wb, conv2_out, conv3_out); //conv3 147 x 147 x 64 convolution(conv3_wb, conv2_out, conv3_out); - BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); + BatchNorm(conv3_out, conv3_beta, conv3_mean, conv3_var); relu(conv3_out, conv3_wb->pbias); maxPoolingInit(conv3_out, pooling1_out, 3, 2); @@ -102,20 +112,23 @@ void facenet::Stem(Mat &image, pBox *output) { convolutionInit(conv4_wb, pooling1_out, conv4_out); //conv4 73 x 73 x 80 convolution(conv4_wb, pooling1_out, conv4_out); - BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); + BatchNorm(conv4_out, conv4_beta, conv4_mean, conv4_var); +// BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); relu(conv4_out, conv4_wb->pbias); convolutionInit(conv5_wb, conv4_out, conv5_out); //conv5 71 x 71 x 192 convolution(conv5_wb, conv4_out, conv5_out); - BatchNorm(conv5_out, conv5_var, conv5_mean, conv5_beta); + BatchNorm(conv5_out, conv5_beta, conv5_mean, conv5_var); +// BatchNorm(conv5_out, conv5_var, conv5_mean, conv5_beta); relu(conv5_out, conv5_wb->pbias); convolutionInit(conv6_wb, conv5_out, output); //conv6 35 x 35 x 256 convolution(conv6_wb, conv5_out, output); - BatchNorm(output, conv6_var, conv6_mean, conv6_beta); + BatchNorm(output, conv6_beta, conv6_mean, conv6_var); +// BatchNorm(output, conv6_var, conv6_mean, conv6_beta); relu(output, conv6_wb->pbias); // firstFlag = false; // } @@ -161,6 +174,13 @@ void facenet::Stem(Mat &image, pBox *output) { freeBN(conv6_beta); } +/** + * Inception_resnet_A网络 + * @param input 输入featuremap + * @param output 输出featuremap + * @param filepath 模型文件路径 + * @param scale 比例系数 + */ void facenet::Inception_resnet_A(pBox *input, pBox *output, string filepath, float scale) { pBox *conv1_out = new pBox; pBox *conv2_out = new pBox; @@ -206,19 +226,19 @@ void facenet::Inception_resnet_A(pBox *input, pBox *output, string filepath, flo long conv1 = ConvAndFcInit(conv1_wb, 32, 256, 1, 1, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, 32); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, 32); long conv2 = ConvAndFcInit(conv2_wb, 32, 256, 1, 1, 0); - BatchNormInit(conv2_var, conv2_mean, conv2_beta, 32); + BatchNormInit(conv2_beta, conv2_mean, conv2_var, 32); long conv3 = ConvAndFcInit(conv3_wb, 32, 32, 3, 1, 1); - BatchNormInit(conv3_var, conv3_mean, conv3_beta, 32); + BatchNormInit(conv3_beta, conv3_mean, conv3_var, 32); long conv4 = ConvAndFcInit(conv4_wb, 32, 256, 1, 1, 0); - BatchNormInit(conv4_var, conv4_mean, conv4_beta, 32); + BatchNormInit(conv4_beta, conv4_mean, conv4_var, 32); long conv5 = ConvAndFcInit(conv5_wb, 32, 32, 3, 1, 1); - BatchNormInit(conv5_var, conv5_mean, conv5_beta, 32); + BatchNormInit(conv5_beta, conv5_mean, conv5_var, 32); long conv6 = ConvAndFcInit(conv6_wb, 32, 32, 3, 1, 1); - BatchNormInit(conv6_var, conv6_mean, conv6_beta, 32); + BatchNormInit(conv6_beta, conv6_mean, conv6_var, 32); long conv7 = ConvAndFcInit(conv7_wb, 256, 96, 1, 1, 0); @@ -227,12 +247,22 @@ void facenet::Inception_resnet_A(pBox *input, pBox *output, string filepath, flo long dataNumber[28] = {conv1, 32, 32, 32, conv2, 32, 32, 32, conv3, 32, 32, 32, conv4, 32, 32, 32, conv5, 32, 32, 32, conv6, 32, 32, 32, conv7, 256, conv8, 0}; - mydataFmt *pointTeam[28] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ - conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ - conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ - conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ - conv5_wb->pdata, conv5_var->pdata, conv5_mean->pdata, conv5_beta->pdata, \ - conv6_wb->pdata, conv6_var->pdata, conv6_mean->pdata, conv6_beta->pdata, \ +// mydataFmt *pointTeam[28] = { +// conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ +// conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ +// conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ +// conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// conv5_wb->pdata, conv5_var->pdata, conv5_mean->pdata, conv5_beta->pdata, \ +// conv6_wb->pdata, conv6_var->pdata, conv6_mean->pdata, conv6_beta->pdata, \ +// conv7_wb->pdata, conv7_wb->pbias, \ +// conv8_wb->pdata, conv8_wb->pbias}; + mydataFmt *pointTeam[28] = { + conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata, \ + conv2_wb->pdata, conv2_beta->pdata, conv2_mean->pdata, conv2_var->pdata, \ + conv3_wb->pdata, conv3_beta->pdata, conv3_mean->pdata, conv3_var->pdata, \ + conv4_wb->pdata, conv4_beta->pdata, conv4_mean->pdata, conv4_var->pdata, \ + conv5_wb->pdata, conv5_beta->pdata, conv5_mean->pdata, conv5_var->pdata, \ + conv6_wb->pdata, conv6_beta->pdata, conv6_mean->pdata, conv6_var->pdata, \ conv7_wb->pdata, conv7_wb->pbias, \ conv8_wb->pdata, conv8_wb->pbias}; @@ -241,34 +271,34 @@ void facenet::Inception_resnet_A(pBox *input, pBox *output, string filepath, flo convolutionInit(conv1_wb, input, conv1_out); //conv1 35 x 35 x 32 convolution(conv1_wb, input, conv1_out); - BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); + BatchNorm(conv1_out, conv1_beta, conv1_mean, conv1_var); relu(conv1_out, conv1_wb->pbias); convolutionInit(conv2_wb, input, conv2_out); //conv2 35 x 35 x 32 convolution(conv2_wb, input, conv2_out); - BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); + BatchNorm(conv2_out, conv2_beta, conv2_mean, conv2_var); relu(conv2_out, conv2_wb->pbias); convolutionInit(conv3_wb, conv2_out, conv3_out); //conv3 35 x 35 x 32 convolution(conv3_wb, conv2_out, conv3_out); - BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); + BatchNorm(conv3_out, conv3_beta, conv3_mean, conv3_var); relu(conv3_out, conv3_wb->pbias); convolutionInit(conv4_wb, input, conv4_out); //conv4 35 x 35 x 32 convolution(conv4_wb, input, conv4_out); - BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); + BatchNorm(conv4_out, conv4_beta, conv4_mean, conv4_var); relu(conv4_out, conv4_wb->pbias); convolutionInit(conv5_wb, conv4_out, conv5_out); //conv5 35 x 35 x 32 convolution(conv5_wb, conv4_out, conv5_out); - BatchNorm(conv5_out, conv5_var, conv5_mean, conv5_beta); + BatchNorm(conv5_out, conv5_beta, conv5_mean, conv5_var); relu(conv5_out, conv5_wb->pbias); convolutionInit(conv6_wb, conv5_out, conv6_out); //conv6 35 x 35 x 32 convolution(conv6_wb, conv5_out, conv6_out); - BatchNorm(conv6_out, conv6_var, conv6_mean, conv6_beta); + BatchNorm(conv6_out, conv6_beta, conv6_mean, conv6_var); relu(conv6_out, conv6_wb->pbias); conv_mergeInit(conv7_out, conv1_out, conv3_out, conv6_out); @@ -280,7 +310,7 @@ void facenet::Inception_resnet_A(pBox *input, pBox *output, string filepath, flo convolution(conv7_wb, conv7_out, conv8_out); addbias(conv8_out, conv7_wb->pbias); - mulandaddInit(input, conv8_out, output, scale); + mulandaddInit(input, conv8_out, output); mulandadd(input, conv8_out, output, scale); relu(output, conv8_wb->pbias); @@ -327,6 +357,11 @@ void facenet::Inception_resnet_A(pBox *input, pBox *output, string filepath, flo freeBN(conv6_beta); } +/** + * Reduction_A + * @param input 输入featuremap + * @param output 输出featuremap + */ void facenet::Reduction_A(pBox *input, pBox *output) { pBox *conv1_out = new pBox; pBox *conv2_out = new pBox; @@ -355,20 +390,26 @@ void facenet::Reduction_A(pBox *input, pBox *output) { long conv1 = ConvAndFcInit(conv1_wb, 384, 256, 3, 2, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, 384); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, 384); long conv2 = ConvAndFcInit(conv2_wb, 192, 256, 1, 1, 0); - BatchNormInit(conv2_var, conv2_mean, conv2_beta, 192); + BatchNormInit(conv2_beta, conv2_mean, conv2_var, 192); long conv3 = ConvAndFcInit(conv3_wb, 192, 192, 3, 1, 0); - BatchNormInit(conv3_var, conv3_mean, conv3_beta, 192); + BatchNormInit(conv3_beta, conv3_mean, conv3_var, 192); long conv4 = ConvAndFcInit(conv4_wb, 256, 192, 3, 2, 0); - BatchNormInit(conv4_var, conv4_mean, conv4_beta, 256); + BatchNormInit(conv4_beta, conv4_mean, conv4_var, 256); long dataNumber[16] = {conv1, 384, 384, 384, conv2, 192, 192, 192, conv3, 192, 192, 192, conv4, 256, 256, 256}; - mydataFmt *pointTeam[16] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ - conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ - conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ - conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata}; +// mydataFmt *pointTeam[16] = { +// conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ +// conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ +// conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ +// conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata}; + mydataFmt *pointTeam[16] = { + conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata, \ + conv2_wb->pdata, conv2_beta->pdata, conv2_mean->pdata, conv2_var->pdata, \ + conv3_wb->pdata, conv3_beta->pdata, conv3_mean->pdata, conv3_var->pdata, \ + conv4_wb->pdata, conv4_beta->pdata, conv4_mean->pdata, conv4_var->pdata}; string filename = "../model_" + to_string(Num) + "/Mixed_6a_list.txt"; readData(filename, dataNumber, pointTeam, 16); @@ -379,25 +420,25 @@ void facenet::Reduction_A(pBox *input, pBox *output) { convolutionInit(conv1_wb, input, conv1_out); //conv1 17 x 17 x 384 convolution(conv1_wb, input, conv1_out); - BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); + BatchNorm(conv1_out, conv1_beta, conv1_mean, conv1_var); relu(conv1_out, conv1_wb->pbias); convolutionInit(conv2_wb, input, conv2_out); //conv2 35 x 35 x 192 convolution(conv2_wb, input, conv2_out); - BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); + BatchNorm(conv2_out, conv2_beta, conv2_mean, conv2_var); relu(conv2_out, conv2_wb->pbias); convolutionInit(conv3_wb, conv2_out, conv3_out); //conv3 35 x 35 x 192 convolution(conv3_wb, conv2_out, conv3_out); - BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); + BatchNorm(conv3_out, conv3_beta, conv3_mean, conv3_var); relu(conv3_out, conv3_wb->pbias); convolutionInit(conv4_wb, conv3_out, conv4_out); //conv4 17 x 17 x 256 convolution(conv4_wb, conv3_out, conv4_out); - BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); + BatchNorm(conv4_out, conv4_beta, conv4_mean, conv4_var); relu(conv4_out, conv4_wb->pbias); conv_mergeInit(output, pooling1_out, conv1_out, conv4_out); //17×17×896 @@ -429,6 +470,13 @@ void facenet::Reduction_A(pBox *input, pBox *output) { freeBN(conv4_beta); } +/** + * Inception_resnet_B网络 + * @param input 输入featuremap + * @param output 输出featuremap + * @param filepath 模型文件路径 + * @param scale 比例系数 + */ void facenet::Inception_resnet_B(pBox *input, pBox *output, string filepath, float scale) { pBox *conv1_out = new pBox; pBox *conv2_out = new pBox; @@ -459,14 +507,14 @@ void facenet::Inception_resnet_B(pBox *input, pBox *output, string filepath, flo long conv1 = ConvAndFcInit(conv1_wb, 128, 896, 1, 1, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, 128); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, 128); long conv2 = ConvAndFcInit(conv2_wb, 128, 896, 1, 1, 0); - BatchNormInit(conv2_var, conv2_mean, conv2_beta, 128); + BatchNormInit(conv2_beta, conv2_mean, conv2_var, 128); long conv3 = ConvAndFcInit(conv3_wb, 128, 128, 0, 1, -1, 7, 1, 3, 0);//[1,7] - BatchNormInit(conv3_var, conv3_mean, conv3_beta, 128); + BatchNormInit(conv3_beta, conv3_mean, conv3_var, 128); long conv4 = ConvAndFcInit(conv4_wb, 128, 128, 0, 1, -1, 1, 7, 0, 3);//[7,1] - BatchNormInit(conv4_var, conv4_mean, conv4_beta, 128); + BatchNormInit(conv4_beta, conv4_mean, conv4_var, 128); long conv5 = ConvAndFcInit(conv5_wb, 896, 256, 1, 1, 0); @@ -475,10 +523,18 @@ void facenet::Inception_resnet_B(pBox *input, pBox *output, string filepath, flo long dataNumber[20] = {conv1, 128, 128, 128, conv2, 128, 128, 128, conv3, 128, 128, 128, conv4, 128, 128, 128, conv5, 896, conv6, 0}; - mydataFmt *pointTeam[20] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ - conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ - conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ - conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// mydataFmt *pointTeam[20] = { +// conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ +// conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ +// conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ +// conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// conv5_wb->pdata, conv5_wb->pbias, \ +// conv6_wb->pdata, conv6_wb->pbias}; + mydataFmt *pointTeam[20] = { + conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata, \ + conv2_wb->pdata, conv2_beta->pdata, conv2_mean->pdata, conv2_var->pdata, \ + conv3_wb->pdata, conv3_beta->pdata, conv3_mean->pdata, conv3_var->pdata, \ + conv4_wb->pdata, conv4_beta->pdata, conv4_mean->pdata, conv4_var->pdata, \ conv5_wb->pdata, conv5_wb->pbias, \ conv6_wb->pdata, conv6_wb->pbias}; @@ -489,24 +545,24 @@ void facenet::Inception_resnet_B(pBox *input, pBox *output, string filepath, flo convolutionInit(conv1_wb, input, conv1_out); //conv1 17*17*128 convolution(conv1_wb, input, conv1_out); - BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); + BatchNorm(conv1_out, conv1_beta, conv1_mean, conv1_var); relu(conv1_out, conv1_wb->pbias); convolutionInit(conv2_wb, input, conv2_out); //conv2 17*17*128 convolution(conv2_wb, input, conv2_out); - BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); + BatchNorm(conv2_out, conv2_beta, conv2_mean, conv2_var); relu(conv2_out, conv2_wb->pbias); convolutionInit(conv3_wb, conv2_out, conv3_out); //conv3 17*17*128 convolution(conv3_wb, conv2_out, conv3_out); - BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); + BatchNorm(conv3_out, conv3_beta, conv3_mean, conv3_var); relu(conv3_out, conv3_wb->pbias); convolutionInit(conv4_wb, conv3_out, conv4_out); //conv4 17*17*128 convolution(conv4_wb, conv3_out, conv4_out); - BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); + BatchNorm(conv4_out, conv4_beta, conv4_mean, conv4_var); relu(conv4_out, conv4_wb->pbias); conv_mergeInit(conv5_out, conv1_out, conv4_out); @@ -518,7 +574,7 @@ void facenet::Inception_resnet_B(pBox *input, pBox *output, string filepath, flo convolution(conv5_wb, conv5_out, conv6_out); addbias(conv6_out, conv5_wb->pbias); - mulandaddInit(input, conv6_out, output, scale); + mulandaddInit(input, conv6_out, output); mulandadd(input, conv6_out, output, scale); relu(output, conv6_wb->pbias); @@ -550,6 +606,11 @@ void facenet::Inception_resnet_B(pBox *input, pBox *output, string filepath, flo freeBN(conv4_beta); } +/** + * Reduction_B + * @param input 输入featuremap + * @param output 输出featuremap + */ void facenet::Reduction_B(pBox *input, pBox *output) { pBox *conv1_out = new pBox; pBox *conv2_out = new pBox; @@ -593,32 +654,41 @@ void facenet::Reduction_B(pBox *input, pBox *output) { long conv1 = ConvAndFcInit(conv1_wb, 256, 896, 1, 1, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, 256); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, 256); long conv2 = ConvAndFcInit(conv2_wb, 384, 256, 3, 2, 0); - BatchNormInit(conv2_var, conv2_mean, conv2_beta, 384); + BatchNormInit(conv2_beta, conv2_mean, conv2_var, 384); long conv3 = ConvAndFcInit(conv3_wb, 256, 896, 1, 1, 0); - BatchNormInit(conv3_var, conv3_mean, conv3_beta, 256); + BatchNormInit(conv3_beta, conv3_mean, conv3_var, 256); long conv4 = ConvAndFcInit(conv4_wb, 256, 256, 3, 2, 0); - BatchNormInit(conv4_var, conv4_mean, conv4_beta, 256); + BatchNormInit(conv4_beta, conv4_mean, conv4_var, 256); long conv5 = ConvAndFcInit(conv5_wb, 256, 896, 1, 1, 0); - BatchNormInit(conv5_var, conv5_mean, conv5_beta, 256); + BatchNormInit(conv5_beta, conv5_mean, conv5_var, 256); long conv6 = ConvAndFcInit(conv6_wb, 256, 256, 3, 1, 1); - BatchNormInit(conv6_var, conv6_mean, conv6_beta, 256); + BatchNormInit(conv6_beta, conv6_mean, conv6_var, 256); long conv7 = ConvAndFcInit(conv7_wb, 256, 256, 3, 2, 0); - BatchNormInit(conv7_var, conv7_mean, conv7_beta, 256); + BatchNormInit(conv7_beta, conv7_mean, conv7_var, 256); long dataNumber[28] = {conv1, 256, 256, 256, conv2, 384, 384, 384, conv3, 256, 256, 256, conv4, 256, 256, 256, conv5, 256, 256, 256, conv6, 256, 256, 256, conv7, 256, 256, 256}; - mydataFmt *pointTeam[28] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ - conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ - conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ - conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ - conv5_wb->pdata, conv5_var->pdata, conv5_mean->pdata, conv5_beta->pdata, \ - conv6_wb->pdata, conv6_var->pdata, conv6_mean->pdata, conv6_beta->pdata, \ - conv7_wb->pdata, conv7_var->pdata, conv7_mean->pdata, conv7_beta->pdata}; +// mydataFmt *pointTeam[28] = { +// conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ +// conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ +// conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ +// conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// conv5_wb->pdata, conv5_var->pdata, conv5_mean->pdata, conv5_beta->pdata, \ +// conv6_wb->pdata, conv6_var->pdata, conv6_mean->pdata, conv6_beta->pdata, \ +// conv7_wb->pdata, conv7_var->pdata, conv7_mean->pdata, conv7_beta->pdata}; + mydataFmt *pointTeam[28] = { + conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata, \ + conv2_wb->pdata, conv2_beta->pdata, conv2_mean->pdata, conv2_var->pdata, \ + conv3_wb->pdata, conv3_beta->pdata, conv3_mean->pdata, conv3_var->pdata, \ + conv4_wb->pdata, conv4_beta->pdata, conv4_mean->pdata, conv4_var->pdata, \ + conv5_wb->pdata, conv5_beta->pdata, conv5_mean->pdata, conv5_var->pdata, \ + conv6_wb->pdata, conv6_beta->pdata, conv6_mean->pdata, conv6_var->pdata, \ + conv7_wb->pdata, conv7_beta->pdata, conv7_mean->pdata, conv7_var->pdata}; string filename = "../model_" + to_string(Num) + "/Mixed_7a_list.txt"; readData(filename, dataNumber, pointTeam, 28); @@ -630,43 +700,46 @@ void facenet::Reduction_B(pBox *input, pBox *output) { convolutionInit(conv1_wb, input, conv1_out); //conv1 17 x 17 x 256 convolution(conv1_wb, input, conv1_out); - BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); + BatchNorm(conv1_out, conv1_beta, conv1_mean, conv1_var); +// BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); relu(conv1_out, conv1_wb->pbias); convolutionInit(conv2_wb, conv1_out, conv2_out); //conv2 8 x 8 x 384 convolution(conv2_wb, conv1_out, conv2_out); - BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); + BatchNorm(conv2_out, conv2_beta, conv2_mean, conv2_var); +// BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); relu(conv2_out, conv2_wb->pbias); convolutionInit(conv3_wb, input, conv3_out); //conv3 17 x 17 x 256 convolution(conv3_wb, input, conv3_out); - BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); + BatchNorm(conv3_out, conv3_beta, conv3_mean, conv3_var); +// BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); relu(conv3_out, conv3_wb->pbias); convolutionInit(conv4_wb, conv3_out, conv4_out); //conv4 8 x 8 x 256 convolution(conv4_wb, conv3_out, conv4_out); - BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); + BatchNorm(conv4_out, conv4_beta, conv4_mean, conv4_var); relu(conv4_out, conv4_wb->pbias); convolutionInit(conv5_wb, input, conv5_out); //conv5 17 x 17 x 256 convolution(conv5_wb, input, conv5_out); - BatchNorm(conv5_out, conv5_var, conv5_mean, conv5_beta); + BatchNorm(conv5_out, conv5_beta, conv5_mean, conv5_var); relu(conv5_out, conv5_wb->pbias); convolutionInit(conv6_wb, conv5_out, conv6_out); //conv6 17 x 17 x 256 convolution(conv6_wb, conv5_out, conv6_out); - BatchNorm(conv6_out, conv6_var, conv6_mean, conv6_beta); + BatchNorm(conv6_out, conv6_beta, conv6_mean, conv6_var); relu(conv6_out, conv6_wb->pbias); convolutionInit(conv7_wb, conv6_out, conv7_out); //conv6 8 x 8 x 256 convolution(conv7_wb, conv6_out, conv7_out); - BatchNorm(conv7_out, conv7_var, conv7_mean, conv7_beta); + BatchNorm(conv7_out, conv7_beta, conv7_mean, conv7_var); relu(conv7_out, conv7_wb->pbias); conv_mergeInit(output, conv2_out, conv4_out, conv7_out, pooling1_out); @@ -714,6 +787,13 @@ void facenet::Reduction_B(pBox *input, pBox *output) { freeBN(conv7_beta); } +/** + * Inception_resnet_C网络 + * @param input 输入featuremap + * @param output 输出featuremap + * @param filepath 模型文件路径 + * @param scale 比例系数 + */ void facenet::Inception_resnet_C(pBox *input, pBox *output, string filepath, float scale) { pBox *conv1_out = new pBox; pBox *conv2_out = new pBox; @@ -744,13 +824,13 @@ void facenet::Inception_resnet_C(pBox *input, pBox *output, string filepath, flo long conv1 = ConvAndFcInit(conv1_wb, 192, 1792, 1, 1, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, 192); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, 192); long conv2 = ConvAndFcInit(conv2_wb, 192, 1792, 1, 1, 0); - BatchNormInit(conv2_var, conv2_mean, conv2_beta, 192); + BatchNormInit(conv2_beta, conv2_mean, conv2_var, 192); long conv3 = ConvAndFcInit(conv3_wb, 192, 192, 0, 1, -1, 3, 1, 1, 0); - BatchNormInit(conv3_var, conv3_mean, conv3_beta, 192); + BatchNormInit(conv3_beta, conv3_mean, conv3_var, 192); long conv4 = ConvAndFcInit(conv4_wb, 192, 192, 0, 1, -1, 1, 3, 0, 1); - BatchNormInit(conv4_var, conv4_mean, conv4_beta, 192); + BatchNormInit(conv4_beta, conv4_mean, conv4_var, 192); long conv5 = ConvAndFcInit(conv5_wb, 1792, 384, 1, 1, 0); @@ -760,10 +840,18 @@ void facenet::Inception_resnet_C(pBox *input, pBox *output, string filepath, flo conv5, 1792, conv6, 0}; - mydataFmt *pointTeam[20] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ - conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ - conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ - conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// mydataFmt *pointTeam[20] = { +// conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ +// conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ +// conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ +// conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// conv5_wb->pdata, conv5_wb->pbias, \ +// conv6_wb->pdata, conv6_wb->pbias}; + mydataFmt *pointTeam[20] = { + conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata, \ + conv2_wb->pdata, conv2_beta->pdata, conv2_mean->pdata, conv2_var->pdata, \ + conv3_wb->pdata, conv3_beta->pdata, conv3_mean->pdata, conv3_var->pdata, \ + conv4_wb->pdata, conv4_beta->pdata, conv4_mean->pdata, conv4_var->pdata, \ conv5_wb->pdata, conv5_wb->pbias, \ conv6_wb->pdata, conv6_wb->pbias}; @@ -774,25 +862,25 @@ void facenet::Inception_resnet_C(pBox *input, pBox *output, string filepath, flo convolutionInit(conv1_wb, input, conv1_out); //conv1 8 x 8 x 192 convolution(conv1_wb, input, conv1_out); - BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); + BatchNorm(conv1_out, conv1_beta, conv1_mean, conv1_var); relu(conv1_out, conv1_wb->pbias); convolutionInit(conv2_wb, input, conv2_out); //conv2 8 x 8 x 192 convolution(conv2_wb, input, conv2_out); - BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); + BatchNorm(conv2_out, conv2_beta, conv2_mean, conv2_var); relu(conv2_out, conv2_wb->pbias); convolutionInit(conv3_wb, conv2_out, conv3_out); //conv3 8 x 8 x 192 convolution(conv3_wb, conv2_out, conv3_out); - BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); + BatchNorm(conv3_out, conv3_beta, conv3_mean, conv3_var); relu(conv3_out, conv3_wb->pbias); convolutionInit(conv4_wb, conv3_out, conv4_out); //conv4 8 x 8 x 192 convolution(conv4_wb, conv3_out, conv4_out); - BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); + BatchNorm(conv4_out, conv4_beta, conv4_mean, conv4_var); relu(conv4_out, conv4_wb->pbias); conv_mergeInit(conv5_out, conv1_out, conv4_out); @@ -804,7 +892,7 @@ void facenet::Inception_resnet_C(pBox *input, pBox *output, string filepath, flo convolution(conv5_wb, conv5_out, conv6_out); addbias(conv6_out, conv5_wb->pbias); - mulandaddInit(input, conv6_out, output, scale); + mulandaddInit(input, conv6_out, output); mulandadd(input, conv6_out, output, scale); relu(output, conv6_wb->pbias); @@ -836,6 +924,13 @@ void facenet::Inception_resnet_C(pBox *input, pBox *output, string filepath, flo freeBN(conv4_beta); } +/** + * Inception_resnet_C网络 最后无激活函数 + * @param input 输入featuremap + * @param output 输出featuremap + * @param filepath 模型文件路径 + * @param scale 比例系数 + */ void facenet::Inception_resnet_C_None(pBox *input, pBox *output, string filepath) { pBox *conv1_out = new pBox; pBox *conv2_out = new pBox; @@ -864,23 +959,30 @@ void facenet::Inception_resnet_C_None(pBox *input, pBox *output, string filepath struct BN *conv4_beta = new BN; long conv1 = ConvAndFcInit(conv1_wb, 192, 1792, 1, 1, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, 192); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, 192); long conv2 = ConvAndFcInit(conv2_wb, 192, 1792, 1, 1, 0); - BatchNormInit(conv2_var, conv2_mean, conv2_beta, 192); + BatchNormInit(conv2_beta, conv2_mean, conv2_var, 192); long conv3 = ConvAndFcInit(conv3_wb, 192, 192, 0, 1, -1, 3, 1, 1, 0); - BatchNormInit(conv3_var, conv3_mean, conv3_beta, 192); + BatchNormInit(conv3_beta, conv3_mean, conv3_var, 192); long conv4 = ConvAndFcInit(conv4_wb, 192, 192, 0, 1, -1, 1, 3, 0, 1); - BatchNormInit(conv4_var, conv4_mean, conv4_beta, 192); + BatchNormInit(conv4_beta, conv4_mean, conv4_var, 192); long conv5 = ConvAndFcInit(conv5_wb, 1792, 384, 1, 1, 0); long dataNumber[18] = {conv1, 192, 192, 192, conv2, 192, 192, 192, conv3, 192, 192, 192, conv4, 192, 192, 192, conv5, 1792}; - mydataFmt *pointTeam[18] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ - conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ - conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ - conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// mydataFmt *pointTeam[18] = { +// conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata, \ +// conv2_wb->pdata, conv2_var->pdata, conv2_mean->pdata, conv2_beta->pdata, \ +// conv3_wb->pdata, conv3_var->pdata, conv3_mean->pdata, conv3_beta->pdata, \ +// conv4_wb->pdata, conv4_var->pdata, conv4_mean->pdata, conv4_beta->pdata, \ +// conv5_wb->pdata, conv5_wb->pbias}; + mydataFmt *pointTeam[18] = { + conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata, \ + conv2_wb->pdata, conv2_beta->pdata, conv2_mean->pdata, conv2_var->pdata, \ + conv3_wb->pdata, conv3_beta->pdata, conv3_mean->pdata, conv3_var->pdata, \ + conv4_wb->pdata, conv4_beta->pdata, conv4_mean->pdata, conv4_var->pdata, \ conv5_wb->pdata, conv5_wb->pbias}; // string filename = "../model_128/Repeat_2_list.txt"; @@ -890,25 +992,25 @@ void facenet::Inception_resnet_C_None(pBox *input, pBox *output, string filepath convolutionInit(conv1_wb, input, conv1_out); //conv1 8 x 8 x 192 convolution(conv1_wb, input, conv1_out); - BatchNorm(conv1_out, conv1_var, conv1_mean, conv1_beta); + BatchNorm(conv1_out, conv1_beta, conv1_mean, conv1_var); relu(conv1_out, conv1_wb->pbias); convolutionInit(conv2_wb, input, conv2_out); //conv2 8 x 8 x 192 convolution(conv2_wb, input, conv2_out); - BatchNorm(conv2_out, conv2_var, conv2_mean, conv2_beta); + BatchNorm(conv2_out, conv2_beta, conv2_mean, conv2_var); relu(conv2_out, conv2_wb->pbias); convolutionInit(conv3_wb, conv2_out, conv3_out); //conv3 8 x 8 x 192 convolution(conv3_wb, conv2_out, conv3_out); - BatchNorm(conv3_out, conv3_var, conv3_mean, conv3_beta); + BatchNorm(conv3_out, conv3_beta, conv3_mean, conv3_var); relu(conv3_out, conv3_wb->pbias); convolutionInit(conv4_wb, conv3_out, conv4_out); //conv4 8 x 8 x 192 convolution(conv4_wb, conv3_out, conv4_out); - BatchNorm(conv4_out, conv4_var, conv4_mean, conv4_beta); + BatchNorm(conv4_out, conv4_beta, conv4_mean, conv4_var); relu(conv4_out, conv4_wb->pbias); conv_mergeInit(conv5_out, conv1_out, conv4_out); @@ -920,7 +1022,7 @@ void facenet::Inception_resnet_C_None(pBox *input, pBox *output, string filepath convolution(conv5_wb, conv5_out, conv6_out); addbias(conv6_out, conv5_wb->pbias); - mulandaddInit(input, conv6_out, output, 1); + mulandaddInit(input, conv6_out, output); mulandadd(input, conv6_out, output); freepBox(conv1_out); @@ -950,12 +1052,22 @@ void facenet::Inception_resnet_C_None(pBox *input, pBox *output, string filepath freeBN(conv4_beta); } +/** + * 平均池化 + * @param input 输入featuremap + * @param output 输出featuremap + */ void facenet::AveragePooling(pBox *input, pBox *output) { // cout << "size:" << input->height << endl; avePoolingInit(input, output, input->height, 2); avePooling(input, output, input->height, 2); } +/** + * flatten 多维转换到一维 + * @param input + * @param output + */ void facenet::Flatten(pBox *input, pBox *output) { output->width = input->channel; output->height = 1; @@ -965,18 +1077,25 @@ void facenet::Flatten(pBox *input, pBox *output) { memcpy(output->pdata, input->pdata, output->channel * output->width * output->height * sizeof(mydataFmt)); } +/** + * 全连接网络 + * @param input 输入featuremap + * @param output 输出featuremap + * @param filepath 网络模型参数文件路径 + */ //参数还未设置 void facenet::fully_connect(pBox *input, pBox *output, string filepath) { struct Weight *conv1_wb = new Weight; - struct BN *conv1_var = new BN; - struct BN *conv1_mean = new BN; struct BN *conv1_beta = new BN; + struct BN *conv1_mean = new BN; + struct BN *conv1_var = new BN; long conv1 = ConvAndFcInit(conv1_wb, Num, 1792, input->height, 1, 0); - BatchNormInit(conv1_var, conv1_mean, conv1_beta, Num); + BatchNormInit(conv1_beta, conv1_mean, conv1_var, Num); long dataNumber[4] = {conv1, Num, Num, Num}; // cout << to_string(sum) << endl; - mydataFmt *pointTeam[4] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata}; +// mydataFmt *pointTeam[4] = {conv1_wb->pdata, conv1_var->pdata, conv1_mean->pdata, conv1_beta->pdata}; + mydataFmt *pointTeam[4] = {conv1_wb->pdata, conv1_beta->pdata, conv1_mean->pdata, conv1_var->pdata}; // string filename = "../model_128/Bottleneck_list.txt"; // int length = sizeof(dataNumber) / sizeof(*dataNumber); readData(filepath, dataNumber, pointTeam, 4); @@ -985,7 +1104,7 @@ void facenet::fully_connect(pBox *input, pBox *output, string filepath) { //conv1 8 x 8 x 192 fullconnect(conv1_wb, input, output); - BatchNorm(output, conv1_var, conv1_mean, conv1_beta); + BatchNorm(output, conv1_beta, conv1_mean, conv1_var); freeWeight(conv1_wb); freeBN(conv1_var); @@ -1009,13 +1128,17 @@ void facenet::printData(pBox *in) { cout << "printData" << endl; } +/** + * facenet网络运行入口 + * @param image + * @param o + * @param count + */ void facenet::run(Mat &image, vector &o, int count) { cout << "=====This is No." + to_string(count) + " Picture=====" << endl; pBox *output = new pBox; pBox *input; Stem(image, output); -// printData(output); -// return; cout << "Stem Finally" << endl; input = output; output = new pBox; @@ -1030,7 +1153,6 @@ void facenet::run(Mat &image, vector &o, int count) { Reduction_A(input, output); cout << "Reduction_A Finally" << endl; input = output; -// freepBox(output); output = new pBox; for (int j = 0; j < 10; ++j) { // model_128/block17_1_list.txt @@ -1048,10 +1170,8 @@ void facenet::run(Mat &image, vector &o, int count) { for (int k = 0; k < 5; ++k) { // model_128/block8_1_list.txt string filepath = "../model_" + to_string(Num) + "/block8_" + to_string((k + 1)) + "_list.txt"; -// cout << filepath << endl; Inception_resnet_C(input, output, filepath, 0.2); input = output; -// freepBox(output); output = new pBox; } cout << "Inception_resnet_C Finally" << endl; @@ -1070,6 +1190,10 @@ void facenet::run(Mat &image, vector &o, int count) { output = new pBox; fully_connect(input, output, "../model_" + to_string(Num) + "/Bottleneck_list.txt"); cout << "Fully_Connect Finally" << endl; + + /** + * L2归一化 + */ mydataFmt sq = 0, sum = 0; for (int i = 0; i < Num; ++i) { sq = pow(output->pdata[i], 2); diff --git a/src/network.cpp b/src/network.cpp old mode 100644 new mode 100755 index b21e1d6..c76f364 --- a/src/network.cpp +++ b/src/network.cpp @@ -1,5 +1,10 @@ #include "network.h" +/** + * 卷积以后偏移 + * @param pbox　feature map + * @param pbias　偏移量 + */ void addbias(struct pBox *pbox, mydataFmt *pbias) { if (pbox->pdata == NULL) { cout << "Relu feature is NULL!!" << endl; @@ -22,6 +27,11 @@ void addbias(struct pBox *pbox, mydataFmt *pbias) { } } +/** + * mat图片转成pbox结构体初始化 + * @param image　mat格式的图片 + * @param pbox　结构体pbox + */ void image2MatrixInit(Mat &image, struct pBox *pbox) { if ((image.data == NULL) || (image.type() != CV_8UC3)) { cout << "image's type is wrong!!Please set CV_8UC3" << endl; @@ -36,6 +46,12 @@ void image2MatrixInit(Mat &image, struct pBox *pbox) { memset(pbox->pdata, 0, pbox->channel * pbox->height * pbox->width * sizeof(mydataFmt)); } +/** + * mat图片转成pbox结构体 + * @param image　mat格式的图片 + * @param pbox　结构体pbox + * @param num　选择mtcnn还是facenet 0-mtcnn 非0-facenet 缺省为０ + */ void image2Matrix(const Mat &image, const struct pBox *pbox, int num) { if ((image.data == NULL) || (image.type() != CV_8UC3)) { cout << "image's type is wrong!!Please set CV_8UC3" << endl; @@ -78,6 +94,12 @@ void image2Matrix(const Mat &image, const struct pBox *pbox, int num) { } } +/** + * 求图片像素的平均值和标准差 + * @param image 图片 + * @param p 平均值 + * @param q 标准差 + */ void MeanAndDev(const Mat &image, mydataFmt &p, mydataFmt &q) { mydataFmt meansum = 0, stdsum = 0; for (int rowI = 0; rowI < image.rows; rowI++) { @@ -96,6 +118,14 @@ void MeanAndDev(const Mat &image, mydataFmt &p, mydataFmt &q) { q = sqrt(stdsum / (image.cols * image.rows * image.channels())); } +/** + * 卷积补偿初始化 + * @param pbox　输入feature map + * @param outpBox　输出feature map + * @param pad　补偿 正方形算子(-1为不规则补偿，0为不需要补偿) + * @param padw 补偿　不规则算子的宽度 + * @param padh　补偿　不规则算子的高度 + */ void featurePadInit(const pBox *pbox, pBox *outpBox, const int pad, const int padw, const int padh) { if (pad < -1) { cout << "the data needn't to pad,please check you network!" << endl; @@ -115,6 +145,14 @@ void featurePadInit(const pBox *pbox, pBox *outpBox, const int pad, const int pa memset(outpBox->pdata, 0, outpBox->channel * outpBox->height * RowByteNum); } +/** + * 卷积补偿 + * @param pbox　输入feature map + * @param outpBox 输出feature map + * @param pad　补偿 正方形算子(-1为不规则补偿，0为不需要补偿) + * @param padw 补偿　不规则算子的宽度 + * @param padh　补偿　不规则算子的高度 + */ void featurePad(const pBox *pbox, pBox *outpBox, const int pad, const int padw, const int padh) { mydataFmt *p = outpBox->pdata; mydataFmt *pIn = pbox->pdata; @@ -143,6 +181,12 @@ void featurePad(const pBox *pbox, pBox *outpBox, const int pad, const int padw, } } +/** + * 卷积初始化 + * @param weight 卷积权重 + * @param pbox　输入feature map + * @param outpBox 输出feature map + */ void convolutionInit(const Weight *weight, pBox *pbox, pBox *outpBox) { outpBox->channel = weight->selfChannel; // ((imginputh - ckh + 2 * ckpad) / stride) + 1; @@ -168,6 +212,12 @@ void convolutionInit(const Weight *weight, pBox *pbox, pBox *outpBox) { } } +/** + * 卷积 + * @param weight 卷积权重 + * @param pbox　输入feature map + * @param outpBox 输出feature map + */ void convolution(const Weight *weight, const pBox *pbox, pBox *outpBox) { int ckh, ckw, ckd, stride, cknum, ckpad, imginputh, imginputw, imginputd, Nh, Nw; mydataFmt *ck, *imginput; @@ -215,6 +265,14 @@ void convolution(const Weight *weight, const pBox *pbox, pBox *outpBox) { } } +/** + * 最大值池化初始化 + * @param pbox 输入feature map + * @param Matrix 输出feature map + * @param kernelSize 池化算子大小 + * @param stride　步长 + * @param flag　标志位 + */ void maxPoolingInit(const pBox *pbox, pBox *Matrix, int kernelSize, int stride, int flag) { if (flag == 1) { Matrix->width = floor((float) (pbox->width - kernelSize) / stride + 1); @@ -229,6 +287,13 @@ void maxPoolingInit(const pBox *pbox, pBox *Matrix, int kernelSize, int stride, memset(Matrix->pdata, 0, Matrix->channel * Matrix->width * Matrix->height * sizeof(mydataFmt)); } +/** + * 最大值池化 + * @param pbox 输入feature map + * @param Matrix 输出feature map + * @param kernelSize 池化算子大小 + * @param stride　步长 + */ void maxPooling(const pBox *pbox, pBox *Matrix, int kernelSize, int stride) { if (pbox->pdata == NULL) { cout << "the feature2Matrix pbox is NULL!!" << endl; @@ -281,6 +346,13 @@ void maxPooling(const pBox *pbox, pBox *Matrix, int kernelSize, int stride) { } } +/** + * 平均值池化初始化 + * @param pbox 输入feature map + * @param Matrix 输出feature map + * @param kernelSize 池化算子大小 + * @param stride　步长 + */ void avePoolingInit(const pBox *pbox, pBox *Matrix, int kernelSize, int stride) { Matrix->width = ceil((float) (pbox->width - kernelSize) / stride + 1); Matrix->height = ceil((float) (pbox->height - kernelSize) / stride + 1); @@ -290,6 +362,13 @@ void avePoolingInit(const pBox *pbox, pBox *Matrix, int kernelSize, int stride) memset(Matrix->pdata, 0, Matrix->channel * Matrix->width * Matrix->height * sizeof(mydataFmt)); } +/** + * 平均值池化 + * @param pbox 输入feature map + * @param Matrix 输出feature map + * @param kernelSize 池化算子大小 + * @param stride　步长 + */ void avePooling(const pBox *pbox, pBox *Matrix, int kernelSize, int stride) { if (pbox->pdata == NULL) { cout << "the feature2Matrix pbox is NULL!!" << endl; @@ -321,10 +400,53 @@ void avePooling(const pBox *pbox, pBox *Matrix, int kernelSize, int stride) { } } +/** + * 激活函数 有系数 初始化 + * @param prelu　激活函数权重 + * @param width　长度 + */ +void pReluInit(struct pRelu *prelu, int width) { + prelu->width = width; + prelu->pdata = (mydataFmt *) malloc(width * sizeof(mydataFmt)); + if (prelu->pdata == NULL)cout << "prelu apply for memory failed!!!!"; + memset(prelu->pdata, 0, width * sizeof(mydataFmt)); +} + +/** + * 激活函数 有系数 + * @param pbox　输入feature + * @param pbias 偏移 + * @param prelu_gmma　激活函数权重 + */ +void prelu(struct pBox *pbox, mydataFmt *pbias, mydataFmt *prelu_gmma) { + if (pbox->pdata == NULL) { + cout << "the pRelu feature is NULL!!" << endl; + return; + } + if (pbias == NULL) { + cout << "the pRelu bias is NULL!!" << endl; + return; + } + mydataFmt *op = pbox->pdata; + mydataFmt *pb = pbias; + mydataFmt *pg = prelu_gmma; + + long dis = pbox->width * pbox->height; + for (int channel = 0; channel < pbox->channel; channel++) { + for (int col = 0; col < dis; col++) { + *op = *op + *pb; + *op = (*op > 0) ? (*op) : ((*op) * (*pg)); + op++; + } + pb++; + pg++; + } +} + /** * 激活函数 没有系数 - * @param pbox - * @param pbias + * @param pbox　输入feature + * @param pbias 偏移 */ void relu(struct pBox *pbox, mydataFmt *pbias) { if (pbox->pdata == NULL) { @@ -349,6 +471,11 @@ void relu(struct pBox *pbox, mydataFmt *pbias) { } } +/** + * 全连接初始化 + * @param weight　权重参数 + * @param outpBox　输出feature map + */ void fullconnectInit(const Weight *weight, pBox *outpBox) { outpBox->channel = weight->selfChannel; outpBox->width = 1; @@ -358,6 +485,12 @@ void fullconnectInit(const Weight *weight, pBox *outpBox) { memset(outpBox->pdata, 0, weight->selfChannel * sizeof(mydataFmt)); } +/** + * 全连接 + * @param weight　权重参数 + * @param pbox 　输入feature map + * @param outpBox　输出feature map + */ void fullconnect(const Weight *weight, const pBox *pbox, pBox *outpBox) { if (pbox->pdata == NULL) { cout << "the fc feature is NULL!!" << endl; @@ -376,6 +509,14 @@ void fullconnect(const Weight *weight, const pBox *pbox, pBox *outpBox) { outpBox->pdata); } +/** + * 一维数组与二位矩阵相乘 + * @param matrix　输入feature map + * @param v 权重 + * @param v_w　权重矩阵的宽度 + * @param v_h　权重矩阵的高度 + * @param p　输出feature map + */ void vectorXmatrix(mydataFmt *matrix, mydataFmt *v, int v_w, int v_h, mydataFmt *p) { for (int i = 0; i < v_h; i++) { p[i] = 0; @@ -385,6 +526,13 @@ void vectorXmatrix(mydataFmt *matrix, mydataFmt *v, int v_w, int v_h, mydataFmt } } +/** + * 读取模型文件 + * @param filename　文件路径 + * @param dataNumber　参数个数数组 + * @param pTeam　变量数组 + * @param length + */ void readData(string filename, long dataNumber[], mydataFmt *pTeam[], int length) { ifstream in(filename.data()); string line; @@ -434,6 +582,20 @@ void readData(string filename, long dataNumber[], mydataFmt *pTeam[], int length } } +/** + * 卷积和全连接初始化 + * @param weight　权重 + * @param schannel　卷积核个数 + * @param lchannel　上一层feature map个数 + * @param kersize　卷积核大小 + * @param stride　卷积步长 + * @param pad　卷积是否补偿 + * @param w　卷积核宽度 + * @param h　卷积核高度 + * @param padw　补偿宽度 + * @param padh　补偿高度 + * @return　参数长度 + */ // w sc lc ks s p kw kh long ConvAndFcInit(struct Weight *weight, int schannel, int lchannel, int kersize, int stride, int pad, int w, int h, int padw, int padh) { @@ -461,6 +623,159 @@ long ConvAndFcInit(struct Weight *weight, int schannel, int lchannel, int kersiz return byteLenght; } +/** + * softmax + * @param pbox feature map + */ +void softmax(const struct pBox *pbox) { + if (pbox->pdata == NULL) { + cout << "the softmax's pdata is NULL , Please check !" << endl; + return; + } + mydataFmt *p2D = pbox->pdata; + mydataFmt *p3D = NULL; + long mapSize = pbox->width * pbox->height; + mydataFmt eleSum = 0; + for (int row = 0; row < pbox->height; row++) { + for (int col = 0; col < pbox->width; col++) { + eleSum = 0; + for (int channel = 0; channel < pbox->channel; channel++) { + p3D = p2D + channel * mapSize; + *p3D = exp(*p3D); + eleSum += *p3D; + } + for (int channel = 0; channel < pbox->channel; channel++) { + p3D = p2D + channel * mapSize; + *p3D = (*p3D) / eleSum; + } + p2D++; + } + } +} + +bool cmpScore(struct orderScore lsh, struct orderScore rsh) { + if (lsh.score < rsh.score) + return true; + else + return false; +} + +/** + * 非极大值抑制 + * @param boundingBox_ + * @param bboxScore_ + * @param overlap_threshold + * @param modelname + */ +void nms(vector &boundingBox_, vector &bboxScore_, const mydataFmt overlap_threshold, + string modelname) { + if (boundingBox_.empty()) { + return; + } + std::vector heros; + //sort the score + sort(bboxScore_.begin(), bboxScore_.end(), cmpScore); + + int order = 0; + float IOU = 0; + float maxX = 0; + float maxY = 0; + float minX = 0; + float minY = 0; + while (bboxScore_.size() > 0) { + order = bboxScore_.back().oriOrder; + bboxScore_.pop_back(); + if (order < 0)continue; + heros.push_back(order); + boundingBox_.at(order).exist = false;//delete it + + for (int num = 0; num < boundingBox_.size(); num++) { + if (boundingBox_.at(num).exist) { + //the iou + maxX = (boundingBox_.at(num).x1 > boundingBox_.at(order).x1) ? boundingBox_.at(num).x1 + : boundingBox_.at(order).x1; + maxY = (boundingBox_.at(num).y1 > boundingBox_.at(order).y1) ? boundingBox_.at(num).y1 + : boundingBox_.at(order).y1; + minX = (boundingBox_.at(num).x2 < boundingBox_.at(order).x2) ? boundingBox_.at(num).x2 + : boundingBox_.at(order).x2; + minY = (boundingBox_.at(num).y2 < boundingBox_.at(order).y2) ? boundingBox_.at(num).y2 + : boundingBox_.at(order).y2; + //maxX1 and maxY1 reuse + maxX = ((minX - maxX + 1) > 0) ? (minX - maxX + 1) : 0; + maxY = ((minY - maxY + 1) > 0) ? (minY - maxY + 1) : 0; + //IOU reuse for the area of two bbox + IOU = maxX * maxY; + if (!modelname.compare("Union")) + IOU = IOU / (boundingBox_.at(num).area + boundingBox_.at(order).area - IOU); + else if (!modelname.compare("Min")) { + IOU = IOU / + ((boundingBox_.at(num).area < boundingBox_.at(order).area) ? boundingBox_.at(num).area + : boundingBox_.at( + order).area); + } + if (IOU > overlap_threshold) { + boundingBox_.at(num).exist = false; + for (vector::iterator it = bboxScore_.begin(); it != bboxScore_.end(); it++) { + if ((*it).oriOrder == num) { + (*it).oriOrder = -1; + break; + } + } + } + } + } + } + for (int i = 0; i < heros.size(); i++) + boundingBox_.at(heros.at(i)).exist = true; +} + +void refineAndSquareBbox(vector &vecBbox, const int &height, const int &width) { + if (vecBbox.empty()) { + cout << "Bbox is empty!!" << endl; + return; + } + float bbw = 0, bbh = 0, maxSide = 0; + float h = 0, w = 0; + float x1 = 0, y1 = 0, x2 = 0, y2 = 0; + for (vector::iterator it = vecBbox.begin(); it != vecBbox.end(); it++) { + if ((*it).exist) { + bbh = (*it).x2 - (*it).x1 + 1; + bbw = (*it).y2 - (*it).y1 + 1; + x1 = (*it).x1 + (*it).regreCoord[1] * bbh; + y1 = (*it).y1 + (*it).regreCoord[0] * bbw; + x2 = (*it).x2 + (*it).regreCoord[3] * bbh; + y2 = (*it).y2 + (*it).regreCoord[2] * bbw; + + h = x2 - x1 + 1; + w = y2 - y1 + 1; + + maxSide = (h > w) ? h : w; + x1 = x1 + h * 0.5 - maxSide * 0.5; + y1 = y1 + w * 0.5 - maxSide * 0.5; + (*it).x2 = round(x1 + maxSide - 1); + (*it).y2 = round(y1 + maxSide - 1); + (*it).x1 = round(x1); + (*it).y1 = round(y1); + + //boundary check + if ((*it).x1 < 0)(*it).x1 = 0; + if ((*it).y1 < 0)(*it).y1 = 0; + if ((*it).x2 > height)(*it).x2 = height - 1; + if ((*it).y2 > width)(*it).y2 = width - 1; + + it->area = (it->x2 - it->x1) * (it->y2 - it->y1); + } + } +} + +/** + * 残差融合初始化 + * @param output 输出feature map + * @param c1 输入feature map + * @param c2 输入feature map + * @param c3 输入feature map + * @param c4 输入feature map + */ void conv_mergeInit(pBox *output, pBox *c1, pBox *c2, pBox *c3, pBox *c4) { output->channel = 0; output->height = c1->height; @@ -482,6 +797,14 @@ void conv_mergeInit(pBox *output, pBox *c1, pBox *c2, pBox *c3, pBox *c4) { memset(output->pdata, 0, output->width * output->height * output->channel * sizeof(mydataFmt)); } +/** + * 残差网络融合 + * @param output 输出feature map + * @param c1 输入feature map + * @param c2 输入feature map + * @param c3 输入feature map + * @param c4 输入feature map + */ void conv_merge(pBox *output, pBox *c1, pBox *c2, pBox *c3, pBox *c4) { // cout << "output->channel:" << output->channel << endl; if (c1 != 0) { @@ -511,7 +834,16 @@ void conv_merge(pBox *output, pBox *c1, pBox *c2, pBox *c3, pBox *c4) { } else { cout << "conv_mergeInit" << endl; } } -void mulandaddInit(const pBox *inpbox, const pBox *temppbox, pBox *outpBox, float scale) { +/** + * 残差网络做多次按比例相加初始化 + * @param inpbox 输入feature map + * @param temppbox 输入feature map + * @param outpBox 输出feature map + */ +void mulandaddInit(const pBox *inpbox, const pBox *temppbox, pBox *outpBox) { + if (!((inpbox->width == temppbox->width) && (inpbox->height == temppbox->height) && + (inpbox->channel == temppbox->channel))) + cout << "the mulandaddInit is failed!!" << endl; outpBox->channel = temppbox->channel; outpBox->width = temppbox->width; outpBox->height = temppbox->height; @@ -520,6 +852,13 @@ void mulandaddInit(const pBox *inpbox, const pBox *temppbox, pBox *outpBox, floa memset(outpBox->pdata, 0, outpBox->width * outpBox->height * outpBox->channel * sizeof(mydataFmt)); } +/** + * 残差网络做多次按比例相加 + * @param inpbox 输入feature map + * @param temppbox 输入feature map + * @param outpBox 输出feature map + * @param scale 比例系数 + */ void mulandadd(const pBox *inpbox, const pBox *temppbox, pBox *outpBox, float scale) { mydataFmt *ip = inpbox->pdata; mydataFmt *tp = temppbox->pdata; @@ -530,7 +869,15 @@ void mulandadd(const pBox *inpbox, const pBox *temppbox, pBox *outpBox, float sc } } -void BatchNormInit(struct BN *var, struct BN *mean, struct BN *beta, int width) { + +/** + * BN初始化 + * @param beta beta + * @param mean 平均值 + * @param var 方差 + * @param width 参数个数 + */ +void BatchNormInit(struct BN *beta, struct BN *mean, struct BN *var, int width) { var->width = width; var->pdata = (mydataFmt *) malloc(width * sizeof(mydataFmt)); if (var->pdata == NULL)cout << "prelu apply for memory failed!!!!"; @@ -547,7 +894,14 @@ void BatchNormInit(struct BN *var, struct BN *mean, struct BN *beta, int width) memset(beta->pdata, 0, width * sizeof(mydataFmt)); } -void BatchNorm(struct pBox *pbox, struct BN *var, struct BN *mean, struct BN *beta) { +/** + * BN实现 + * @param pbox　输入feature map + * @param beta beta + * @param mean 平均值 + * @param var 方差 + */ +void BatchNorm(struct pBox *pbox, struct BN *beta, struct BN *mean, struct BN *var) { if (pbox->pdata == NULL) { cout << "Relu feature is NULL!!" << endl; return; diff --git a/src/network.h b/src/network.h old mode 100644 new mode 100755 index d7432f7..c311b23 --- a/src/network.h +++ b/src/network.h @@ -24,6 +24,8 @@ void avePooling(const pBox *pbox, pBox *Matrix, int kernelSize, int stride); void featurePad(const pBox *pbox, pBox *outpBox, const int pad, const int padw = 0, const int padh = 0); +void prelu(struct pBox *pbox, mydataFmt *pbias, mydataFmt *prelu_gmma); + void relu(struct pBox *pbox, mydataFmt *pbias); void fullconnect(const Weight *weight, const pBox *pbox, pBox *outpBox); @@ -33,6 +35,10 @@ void readData(string filename, long dataNumber[], mydataFmt *pTeam[], int length long ConvAndFcInit(struct Weight *weight, int schannel, int lchannel, int kersize, int stride, int pad, int w = 0, int h = 0, int padw = 0, int padh = 0); +void pReluInit(struct pRelu *prelu, int width); + +void softmax(const struct pBox *pbox); + void image2MatrixInit(Mat &image, struct pBox *pbox); void featurePadInit(const pBox *pbox, pBox *outpBox, const int pad, const int padw = 0, const int padh = 0); @@ -45,6 +51,13 @@ void convolutionInit(const Weight *weight, pBox *pbox, pBox *outpBox); void fullconnectInit(const Weight *weight, pBox *outpBox); +bool cmpScore(struct orderScore lsh, struct orderScore rsh); + +void nms(vector &boundingBox_, vector &bboxScore_, const mydataFmt overlap_threshold, + string modelname = "Union"); + +void refineAndSquareBbox(vector &vecBbox, const int &height, const int &width); + void vectorXmatrix(mydataFmt *matrix, mydataFmt *v, int v_w, int v_h, mydataFmt *p); void convolution(const Weight *weight, const pBox *pbox, pBox *outpBox); @@ -55,11 +68,12 @@ void conv_merge(pBox *output, pBox *c1 = 0, pBox *c2 = 0, pBox *c3 = 0, pBox *c4 void conv_mergeInit(pBox *output, pBox *c1 = 0, pBox *c2 = 0, pBox *c3 = 0, pBox *c4 = 0); -void mulandaddInit(const pBox *inpbox, const pBox *temppbox, pBox *outpBox, float scale); +void mulandaddInit(const pBox *inpbox, const pBox *temppbox, pBox *outpBox); void mulandadd(const pBox *inpbox, const pBox *temppbox, pBox *outpBox, float scale = 1); -void BatchNormInit(struct BN *var, struct BN *mean, struct BN *beta, int width); +void BatchNormInit(struct BN *beta, struct BN *mean, struct BN *var, int width); + +void BatchNorm(struct pBox *pbox, struct BN *beta, struct BN *mean, struct BN *var); -void BatchNorm(struct pBox *pbox, struct BN *var, struct BN *mean, struct BN *beta); #endif \ No newline at end of file