c++编写神经网络（一）MTCNN内存空间的调用

最新推荐文章于 2024-08-29 07:48:56 发布
最新推荐文章于 2024-08-29 07:48:56 发布
阅读量1.4k 收藏 3
点赞数 4
CC 4.0 BY-SA版权
分类专栏： FPGA 目标检测 机器学习
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.youkuaiyun.com/weixin_36474809/article/details/85990687
本文详细解析了MTCNN人脸检测网络在FPGA上的移植过程，包括内存管理和接口设计。探讨了如何在受限的硬件资源下，为网络权重和特征图分配连续的内存空间，以及如何设计高效的硬件接口以加速网络运算。

背景：FPGA端程序需要确定的偏移地址，只能有一个指针。我们需要一次性的在DRAM上开辟内存空间然后传给IPcore来运算。所以我们必须从头编写MTCNN的代码。

目的：查出原MTCNN代码的内存空间的开辟量。

目录

一、旧MTCNN的内存实现

1.1 每个子网络的构建

1.2 mtcnn的构造

1.3 权重的内存实现

1.4 feature的内存实现

1.5 内存实现相关的子函数

二、新MTCNN内存实现的思路

2.1 weight的初始化与权重写入

2.2 feature内存的初始化

  Pnet

  Rnet，Onet

 三、IPcore的接口设计

四、weight的权重大小

  网络尺寸表

    Pnet

    Rnet

    Onet

4.1 Pnet weight的大小

  Pnet

  Rnet

  Onet

五、feature的大小

  5.1 Rnet

5.2 Onet

5.3 Pnet

六、网络的总尺寸

6.1 权重尺寸

6.2 feature大小

一、旧MTCNN的内存实现

旧MTCNN的weight，input，output分别用malloc开辟新的内存空间。

1.1 每个子网络的构建

以Pnet为例

Pnet::Pnet(){
    Pthreshold = 0.9;//0.6;
    firstFlag = true;
	//in order of [weight] [input] [output]
    
//the RGB image in pBox format 
    this->rgb = new pBox;
//conv1,prelu1
    this->conv1_wb = new Weight;
    this->conv1_out = new pBox;//output
    this->conv1_out_pad = new pBox;//output padding
//pool1
    this->pool1_conv1_wb = new Weight;
    this->pool1_conv1_out = new pBox;//output
//conv2,relu2
    this->conv2_wb = new Weight;
    this->conv2_out = new pBox;//output
//conv3,relu3
    this->conv3_wb = new Weight;
    this->conv3_out = new pBox;//output
//post conv layers
    this->conv4c1_wb = new Weight;
    this->conv4c2_wb = new Weight;//weight
//    this->score_matrix = new pBox;
    this->score_ = new pBox;
//    this->location_matrix = new pBox;
    this->location_ = new pBox;    
    
    //weight pointer create,network initialize
    //                                 w           oc  ic  ks s lp  rp 
    long conv1 = initConvAndFc(   this->conv1_wb,  10, 3,  3, 1, 0, 0);
long pool1_conv1=initConvAndFc(this->pool1_conv1_wb,16,10, 3, 2, 0, 0);
    long conv2 = initConvAndFc(   this->conv2_wb,  32, 16, 3, 1, 0, 0);
    long conv3 = initConvAndFc(   this->conv3_wb,  32, 32, 3, 1, 0, 0);
    long conv4c1 = initConvAndFc(this->conv4c1_wb, 2,  32, 1, 1, 0, 0);
    long conv4c2 = initConvAndFc(this->conv4c2_wb, 4,  32, 1, 1, 0, 0);
    long dataNumber[12] = {conv1,10,pool1_conv1,16,conv2,32,conv3,32,conv4c1,2,conv4c2,4};
    mydataFmt *pointTeam[12] = {this->conv1_wb->pdata, this->conv1_wb->pbias, \
                            this->pool1_conv1_wb->pdata, this->pool1_conv1_wb->pbias, \
                            this->conv2_wb->pdata, this->conv2_wb->pbias, \
                            this->conv3_wb->pdata, this->conv3_wb->pbias, \
                            this->conv4c1_wb->pdata, this->conv4c1_wb->pbias, \
                            this->conv4c2_wb->pdata, this->conv4c2_wb->pbias \
                            };
    string filename = "Pnet.bin";
	printf("Create Pnet,Read Pnet.bin ,12 pointers \n");
    readData(filename, dataNumber, pointTeam, 12);
}

c++的类构造函数与析构函数 http://www.runoob.com/cplusplus/cpp-constructor-destructor.html

运用构造函数开辟每层的内存存入指针，每层存在一个结构体之中，用initConvandFc函数初始化weight的大小，用convolutioninit这些函数初始化feature的内存。

运用此方法构造的Pnet，Rnet，Onet

1.2 mtcnn的构造

因为Pnet根据金字塔结构，需要多次构造Pnet；

Rnet，Onet只需要构造一次。

子网络的构建mtcnn的类初始化函数之中实现。

1.3 权重的内存实现

网络初始化过程之中，初始化一系列权重指针，然后从文件读入权重。权重之间独立。

    //weight pointer create,network initialize
    //                                 w           oc  ic  ks s lp  rp 
    long conv1 = initConvAndFc(   this->conv1_wb,  10, 3,  3, 1, 0, 0);
long pool1_conv1=initConvAndFc(this->pool1_conv1_wb,16,10, 3, 2, 0, 0);
    long conv2 = initConvAndFc(   this->conv2_wb,  32, 16, 3, 1, 0, 0);
    long conv3 = initConvAndFc(   this->conv3_wb,  32, 32, 3, 1, 0, 0);
    long conv4c1 = initConvAndFc(this->conv4c1_wb, 2,  32, 1, 1, 0, 0);
    long conv4c2 = initConvAndFc(this->conv4c2_wb, 4,  32, 1, 1, 0, 0);
    long dataNumber[12] = {conv1,10,pool1_conv1,16,conv2,32,conv3,32,conv4c1,2,conv4c2,4};
    mydataFmt *pointTeam[12] = {this->conv1_wb->pdata, this->conv1_wb->pbias, \
                            this->pool1_conv1_wb->pdata, this->pool1_conv1_wb->pbias, \
                            this->conv2_wb->pdata, this->conv2_wb->pbias, \
                            this->conv3_wb->pdata, this->conv3_wb->pbias, \
                            this->conv4c1_wb->pdata, this->conv4c1_wb->pbias, \
                            this->conv4c2_wb->pdata, this->conv4c2_wb->pbias \
                            };
    string filename = "Pnet.bin";
	printf("Create Pnet,Read Pnet.bin ,12 pointers \n");
    readData(filename, dataNumber, pointTeam, 12);

1.4 feature的内存实现

权重固定而featreu不固定，所以每次运行网络之前都会初始化feature。

    if(firstFlag){
        printf("Pnet buffer init\n");
    //change Mat image to pBox format
        image2MatrixInit(image, this->rgb);
		//conv1,prelu1
        convolutionInit(this->conv1_wb, this->rgb, this->conv1_out);
        featurePadInit(this->conv1_out, this->conv1_out_pad, 0,1);
		//pool1
        convolutionInit(this->pool1_conv1_wb, this->conv1_out_pad, this->pool1_conv1_out);
    //conv2,prelu2
        convolutionInit(this->conv2_wb, this->pool1_conv1_out, this->conv2_out);  
    //conv3,prelu3
        convolutionInit(this->conv3_wb, this->conv2_out, this->conv3_out);
    //post conv layers
        convolutionInit(this->conv4c1_wb, this->conv3_out, this->score_);
        convolutionInit(this->conv4c2_wb, this->conv3_out, this->location_);
        firstFlag = false;
    }

以Pnet初始化为例，用相关的Init函数开辟独立的内存以供feature存储。

1.5 内存实现相关的子函数

void convolutionInit(const Weight *weightIn, const pBox *pboxParameter, pBox *outpBox){
	outpBox->channel = weightIn->out_ChannelNum;
    outpBox->width = (pboxParameter->width - weightIn->kernelSize) / weightIn->stride + 1;
    outpBox->height = (pboxParameter->height - weightIn->kernelSize) / weightIn->stride + 1;
	int outpBoxByteSize=weightIn->out_ChannelNum*outpBox->width*outpBox->height*sizeof(mydataFmt);
    outpBox->pdata = (mydataFmt *)malloc(outpBoxByteSize);
    
	if(outpBox->pdata==NULL)cout<<"the convolutionInit is failed!!"<<endl;
    memset(outpBox->pdata , 0, outpBoxByteSize);
}

例如卷积的初始化函数就是独立的开辟出卷积需要的内存大小。

long initConvAndFc(struct Weight *weight, int schannel, int lchannel, int kersize, int stride, int leftPad,int rightPad){
    weight->out_ChannelNum = schannel;
    weight->in_ChannelNum = lchannel;
    weight->kernelSize = kersize;
    weight->stride = stride;
    weight->leftPad = leftPad;
	  weight->rightPad = rightPad;
    weight->pbias = (mydataFmt *)malloc(schannel*sizeof(mydataFmt));
    if(weight->pbias==NULL)cout<<"In initConvAndFc weight buffer malloc failure!";
    memset(weight->pbias, 0, schannel*sizeof(mydataFmt));
    long byteLenght = weight->out_ChannelNum*weight->in_ChannelNum*weight->kernelSize*weight->kernelSize;
    weight->pdata = (mydataFmt *)malloc(byteLenght*sizeof(mydataFmt));
    if(weight->pdata==NULL)cout<<"In initConvAndFc weight buffer malloc failure!";
    memset(weight->pdata, 0, byteLenght*sizeof(mydataFmt));

    return byteLenght;
}

权重的初始化函数就是开辟出权重需要的大小。

二、新MTCNN内存实现的思路

按照weight，feature的顺序存储

其中，按照Pnet，Rnet，Onet的顺序存储。每个子网络之中按层的先后顺序来实现。

 根据具体的卷积算出针对具体卷积的weight，input，output的偏移地址。

weight_offset=net_offset+weight_layer_offset

input_offset=feature_offset+feature_layer_offset

2.1 weight的初始化与权重写入

一次性开辟所有weight的内存，然后根据Pnet，Rnet，Onet的顺序写入

具体子网络之后用层用从前往后的层顺序。写入weight，

然后从weight文件中读入指针。

2.2 feature内存的初始化

  Pnet

因为Pnet按照图像金字塔的结构，所以每层的feature大小都不一样。

按照pyramind的顺序，pyramid之中的feature按照层的顺序。

  Rnet，Onet

按照层的顺序进行循环。一次性开辟内存空间反复使用。

 三、IPcore的接口设计

  • 接口传入该层参数，
  • weight，input,output三个偏移量
  • DRAM地址
//----------------convolution in FPGA-----------------------------------
void convolution_3x3(int inHight,int inWidth,int inChanNum,int outHight,int outWidth,int OutChanNum,int stride,
    volatile float *SHARED_DRAM, 
    int weight_offset,int input_offset,int output offset){
#pragma HLS INTERFACE s_axilite register port=inHight bundle=axilite
#pragma HLS INTERFACE s_axilite register port=inWidth bundle=axilite
#pragma HLS INTERFACE s_axilite register port=inChanNum bundle=axilite
#pragma HLS INTERFACE s_axilite register port=outHight bundle=axilite
#pragma HLS INTERFACE s_axilite register port=outWidth bundle=axilite
#pragma HLS INTERFACE s_axilite register port=OutChanNum bundle=axilite
#pragma HLS INTERFACE s_axilite register port=stride bundle=axilite
#pragma HLS INTERFACE m_axi depth=DRAM_DEPTH port=SHARED_DRAM offset=slave bundle=memorybus
#pragma HLS INTERFACE s_axilite register port=weight_offset bundle=axilite
#pragma HLS INTERFACE s_axilite register port=input_offset bundle=axilite
#pragma HLS INTERFACE s_axilite register port=output offset bundle=axilite

四、weight的权重大小

  网络尺寸表

    Pnet

Feature size

name

Kernel size

Stride

Padding

12*12*3

conv1

ReLU1

3*3*10

1

Valid

10*10*10

pool1_conv1

pool1_ReLU1

3*3*16

2

Same

5*5*16

conv2

ReLU2

3*3*32

1

Valid

3*3*32

conv3

ReLU3

3*3*32

1

Valid

1*1*32

 

 

 

 

    Rnet

Feature size

name

Kernel size

Stride

Padding

24*24*3

conv1

relu1

3*3*28

1

Same

24*24*28

pool1_conv1

pool1_relu1

3*3*282Same
12*12*28

conv2

relu2

3*3*481Same
12*12*48

pool2_conv3

poo2_relu3

3*3*482Same
6*6*48

conv3

relu3

3*3*642Same
3*3*64 

 

 

 

    Onet

Feature size

name

Kernel size

Stride

Padding

48*48*3

conv1

relu1

3*3*32

1

Same

48*48*32

conv2

relu2

3*3*32

2

Same

24*24*32

conv3

relu3

3*3*64

1

Same

24*24*64

conv4_

relu4_

3*3*64

2

Same

12*12*64

conv5_

relu5_

3*3*128

2

Same

6*6*128

conv6_

relu6_

3*3*128

2

Same

3*3*128    

4.1 Pnet weight的大小

----------Pnet conv1---------------------
-----------Init conv and Fc for weight---------------------
Current layer weight 1080 byte and bias 40 byte, total= 1120 byte
----------Pnet poo1_conv1---------------------
-----------Init conv and Fc for weight---------------------
Current layer weight 5760 byte and bias 64 byte, total= 5824 byte
----------Pnet conv2---------------------
-----------Init conv and Fc for weight---------------------
Current layer weight 18432 byte and bias 128 byte, total= 18560 byte
----------Pnet conv3---------------------
-----------Init conv and Fc for weight---------------------
Current layer weight 36864 byte and bias 128 byte, total= 36992 byte
----------Pnet conv4c1---------------------
-----------Init conv and Fc for weight---------------------
Current layer weight 256 byte and bias 8 byte, total= 264 byte
----------Pnet conv4c2---------------------
-----------Init conv and Fc for weight---------------------
Current layer weight 512 byte and bias 16 byte, total= 528 byte

  Pnet

Pnet

weight

layer nameweight data byteweight bias bytetotal byte
conv11080401120
pool1_conv15760645824
conv21843212818560
conv33686412836992
conv4c12568264
conv4c251216528

  Rnet

----------Rnet conv1---------------------
Current layer weight 3024 byte and bias 112 byte, total= 3136 byte
----------Rnet pool_conv1---------------------
Current layer weight 28224 byte and bias 112 byte, total= 28336 byte
----------Rnet conv2---------------------
Current layer weight 48384 byte and bias 192 byte, total= 48576 byte
----------Rnet pool2_conv3---------------------
Current layer weight 82944 byte and bias 192 byte, total= 83136 byte
----------Rnet conv3---------------------
Current layer weight 110592 byte and bias 256 byte, total= 110848 byte
----------Rnet fc4---------------------
Current layer weight 294912 byte and bias 512 byte, total= 295424 byte
----------Rnet score---------------------
Current layer weight 1024 byte and bias 8 byte, total= 1032 byte
----------Rnet location---------------------
Current layer weight 2048 byte and bias 16 byte, total= 2064 byte
Create Rnet,Read Rnet.bin ,16 pointers

Rnet

weight

layer nameweight data byteweight bias bytetotal byte
conv130241123136
pool_conv12822411228336
conv24838419248576
pool2_conv38294419283136
conv3110592256110848
fc4294912512295424
score102481032
location2048162064

  Onet

----------Onet conv1---------------------
Current layer weight 3456 byte and bias 128 byte, total= 3584 byte
----------Onet conv2---------------------
Current layer weight 36864 byte and bias 128 byte, total= 36992 byte
----------Onet conv3---------------------
Current layer weight 73728 byte and bias 256 byte, total= 73984 byte
----------Onet conv4_---------------------
Current layer weight 147456 byte and bias 256 byte, total= 147712 byte
----------Onet conv5_---------------------
Current layer weight 294912 byte and bias 512 byte, total= 295424 byte
----------Onet conv6_---------------------
Current layer weight 589824 byte and bias 512 byte, total= 590336 byte
----------Onet fc5---------------------
Current layer weight 1179648 byte and bias 1024 byte, total= 1180672 byte
----------Onet score---------------------
Current layer weight 2048 byte and bias 8 byte, total= 2056 byte
----------Onet location---------------------
Current layer weight 4096 byte and bias 16 byte, total= 4112 byte
Create Onet,Read Onet.bin ,18 pointers
Onet buffer init

Onet

weight

layer nameweight data byteweight bias bytetotal byte
conv134561283584
conv23686412836992
conv37372825673984
conv4_147456256147712
conv5_294912512295424
conv6_589824512590336
fc5117964810241180672
score204882056
location4096164112

五、feature的大小

  5.1 Rnet

    //input image to pBox format
	RnetImage2MatrixInit(rgb);
	featurePadInit(this->rgb, this->rgb_pad, 1, 1);
    //conv1
	convolutionInit(this->conv1_wb, this->rgb_pad, this->conv1_out);
	featurePadInit(this->conv1_out, this->conv1_out_pad, 0, 1);
    //pool1
	convolutionInit(this->pool_conv1_wb, this->conv1_out_pad, this->pool_conv1_out);
	featurePadInit(this->pool_conv1_out, this->pool_conv1_out_pad, 1, 1);
    //conv2
	convolutionInit(this->conv2_wb, this->pool_conv1_out_pad, this->conv2_out);
	featurePadInit(this->conv2_out, this->conv2_out_pad, 0, 1);
    //pool2
	convolutionInit(this->pool2_conv3_wb, this->conv2_out_pad, this->pool2_conv3_out);
	featurePadInit(this->pool2_conv3_out, this->pool2_conv3_out_pad, 0, 1);
    //conv3
	convolutionInit(this->conv3_wb, this->pool2_conv3_out_pad, this->conv3_out);
    //post conv precess
	fullconnectInit(this->fc4_wb, this->fc4_out);
	fullconnectInit(this->score_wb, this->score_);
	fullconnectInit(this->location_wb, this->location_);

据此生成的大小为：

Create Rnet,Read Rnet.bin ,16 pointers
Rnet weight buffer init SUCCESS!Start init feature buffer
-------Rnet image2Matrix buffer byte size 6912 -------
--------Feature pad buffer byte size 8112 -------
--------convolution buffer byte size 64512 -------
--------Feature pad buffer byte size 70000 -------
--------convolution buffer byte size 16128 -------
--------Feature pad buffer byte size 21952 -------
--------convolution buffer byte size 27648 -------
--------Feature pad buffer byte size 32448 -------
--------convolution buffer byte size 6912 -------
--------Feature pad buffer byte size 9408 -------
--------convolution buffer byte size 2304 -------
--------fullconnect buffer byte size 512 -------
--------fullconnect buffer byte size 8 -------
--------fullconnect buffer byte size 16 -------

Rnet

feature

feature namebytesize
rgb6912
rgb_pad8112
conv1_out64512
conv1_out_pad70000
pool_conv1_out16128
pool_conv1_out_pad21952
conv2_out27648
conv2_out_pad32448
pool2_conv3_out6912
pool2_conv3_out_pad9408
conv3_out2304
fc4_out512
score_8
location_16

5.2 Onet

    //Init the network
	printf("Onet weight buffer init SUCCESS! Start init feature buffer \n");
	  //change image to pBox format
    OnetImage2MatrixInit(rgb);
    featurePadInit(this->rgb, this->rgb_pad, 1, 1);//s=1
	  //conv1
    convolutionInit(this->conv1_wb, this->rgb_pad, this->conv1_out);
    featurePadInit(this->conv1_out, this->conv1_out_pad, 0, 1);//s=2
	  //conv2
    convolutionInit(this->conv2_wb, this->conv1_out_pad, this->conv2_out);
    featurePadInit(this->conv2_out, this->conv2_out_pad, 1, 1);//s=1
    //conv3
    convolutionInit(this->conv3_wb, this->conv2_out_pad, this->conv3_out);
    featurePadInit(this->conv3_out, this->conv3_out_pad, 0, 1);//s=2
    //conv4_
    convolutionInit(this->conv4_wb, this->conv3_out_pad, this->conv4_out);
    featurePadInit(this->conv4_out, this->conv4_out_pad, 0, 1);//s=2
    //conv5_
    convolutionInit(this->conv5_wb, this->conv4_out_pad, this->conv5_out);
    featurePadInit(this->conv5_out, this->conv5_out_pad, 0, 1);//s=2
    //conv6_
    convolutionInit(this->conv6_wb, this->conv5_out_pad, this->conv6_out);
    //post conv precess
	  fullconnectInit(this->fc5_wb, this->fc5_out);
    fullconnectInit(this->score_wb, this->score_);
    fullconnectInit(this->location_wb, this->location_);

Onet

feature

feature namebytesize
rgb27648
rgb_pad30000
conv1_out294912
conv1_out_pad307328
conv2_out73728
conv2_out_pad86528
conv3_out147456
conv3_out_pad160000
conv4_out36864
conv4_out_pad43264
conv5_out18432
conv5_out_pad25088
conv6_out4608
fc5_out1024
score_8
location_16

5.3 Pnet

Pnet的feature较多，有很多个pyramid，每个pyramid的尺寸都不一样。

//scales 0.400000  h:116  w:142
//scales 0.340000  h:98  w:120
//scales 0.289000  h:84  w:102
//scales 0.245650  h:72  w:88
//scales 0.208803  h:62  w:74
//scales 0.177482  h:52  w:64
//scales 0.150860  h:44  w:54
//scales 0.128231  h:38  w:46
//scales 0.108996  h:32  w:40
//scales 0.092647  h:28  w:34
//scales 0.078750  h:24  w:28
//scales 0.066937  h:20  w:24
//scales 0.056897  h:18  w:22
//scales 0.048362  h:14  w:18
Create Pnet,Read Pnet.bin ,12 pointers
------Start find Face function
scales 0.400000  h:116  w:142
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 197664 ----------------
--------convolution buffer byte size 638400 -------
--------Feature pad buffer byte size 648600 -------
--------convolution buffer byte size 255360 -------
--------convolution buffer byte size 478720 -------
--------convolution buffer byte size 447744 -------
--------convolution buffer byte size 27984 -------
--------convolution buffer byte size 55968 -------
scales 0.340000  h:98  w:120
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 141120 ----------------
--------convolution buffer byte size 453120 -------
--------Feature pad buffer byte size 461720 -------
--------convolution buffer byte size 181248 -------
--------convolution buffer byte size 335616 -------
--------convolution buffer byte size 309760 -------
--------convolution buffer byte size 19360 -------
--------convolution buffer byte size 38720 -------
scales 0.289000  h:84  w:102
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 102816 ----------------
--------convolution buffer byte size 328000 -------
--------Feature pad buffer byte size 335320 -------
--------convolution buffer byte size 131200 -------
--------convolution buffer byte size 239616 -------
--------convolution buffer byte size 217856 -------
--------convolution buffer byte size 13616 -------
--------convolution buffer byte size 27232 -------
scales 0.245650  h:72  w:88
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 76032 ----------------
--------convolution buffer byte size 240800 -------
--------Feature pad buffer byte size 247080 -------
--------convolution buffer byte size 96320 -------
--------convolution buffer byte size 173184 -------
--------convolution buffer byte size 154752 -------
--------convolution buffer byte size 9672 -------
--------convolution buffer byte size 19344 -------
scales 0.208803  h:62  w:74
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 55056 ----------------
--------convolution buffer byte size 172800 -------
--------Feature pad buffer byte size 178120 -------
--------convolution buffer byte size 69120 -------
--------convolution buffer byte size 121856 -------
--------convolution buffer byte size 106496 -------
--------convolution buffer byte size 6656 -------
--------convolution buffer byte size 13312 -------
scales 0.177482  h:52  w:64
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 39936 ----------------
--------convolution buffer byte size 124000 -------
--------Feature pad buffer byte size 128520 -------
--------convolution buffer byte size 49600 -------
--------convolution buffer byte size 85376 -------
--------convolution buffer byte size 72576 -------
--------convolution buffer byte size 4536 -------
--------convolution buffer byte size 9072 -------
scales 0.150860  h:44  w:54
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 28512 ----------------
--------convolution buffer byte size 87360 -------
--------Feature pad buffer byte size 91160 -------
--------convolution buffer byte size 34944 -------
--------convolution buffer byte size 58368 -------
--------convolution buffer byte size 47872 -------
--------convolution buffer byte size 2992 -------
--------convolution buffer byte size 5984 -------
scales 0.128231  h:38  w:46
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 20976 ----------------
--------convolution buffer byte size 63360 -------
--------Feature pad buffer byte size 66600 -------
--------convolution buffer byte size 25344 -------
--------convolution buffer byte size 40960 -------
--------convolution buffer byte size 32256 -------
--------convolution buffer byte size 2016 -------
--------convolution buffer byte size 4032 -------
scales 0.108996  h:32  w:40
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 15360 ----------------
--------convolution buffer byte size 45600 -------
--------Feature pad buffer byte size 48360 -------
--------convolution buffer byte size 18240 -------
--------convolution buffer byte size 28288 -------
--------convolution buffer byte size 21120 -------
--------convolution buffer byte size 1320 -------
--------convolution buffer byte size 2640 -------
scales 0.092647  h:28  w:34
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 11424 ----------------
--------convolution buffer byte size 33280 -------
--------Feature pad buffer byte size 35640 -------
--------convolution buffer byte size 13312 -------
--------convolution buffer byte size 19712 -------
--------convolution buffer byte size 13824 -------
--------convolution buffer byte size 864 -------
--------convolution buffer byte size 1728 -------
scales 0.078750  h:24  w:28
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 8064 ----------------
--------convolution buffer byte size 22880 -------
--------Feature pad buffer byte size 24840 -------
--------convolution buffer byte size 9152 -------
--------convolution buffer byte size 12672 -------
--------convolution buffer byte size 8064 -------
--------convolution buffer byte size 504 -------
--------convolution buffer byte size 1008 -------
scales 0.066937  h:20  w:24
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 5760 ----------------
--------convolution buffer byte size 15840 -------
--------Feature pad buffer byte size 17480 -------
--------convolution buffer byte size 6336 -------
--------convolution buffer byte size 8064 -------
--------convolution buffer byte size 4480 -------
--------convolution buffer byte size 280 -------
--------convolution buffer byte size 560 -------
scales 0.056897  h:18  w:22
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 4752 ----------------
--------convolution buffer byte size 12800 -------
--------Feature pad buffer byte size 14280 -------
--------convolution buffer byte size 5120 -------
--------convolution buffer byte size 6144 -------
--------convolution buffer byte size 3072 -------
--------convolution buffer byte size 192 -------
--------convolution buffer byte size 384 -------
scales 0.048362  h:14  w:18
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 3024 ----------------
--------convolution buffer byte size 7680 -------
--------Feature pad buffer byte size 8840 -------
--------convolution buffer byte size 3072 -------
--------convolution buffer byte size 3072 -------
--------convolution buffer byte size 1024 -------
--------convolution buffer byte size 64 -------
--------convolution buffer byte size 128 -------

六、网络的总尺寸

6.1 权重尺寸

Pnet

weight

layer nameweight data byteweight bias bytetotal byte
conv11080401120
pool1_conv15760645824
conv21843212818560
conv33686412836992
conv4c12568264
conv4c251216528

Rnet

weight

layer nameweight data byteweight bias bytetotal byte
conv130241123136
pool_conv12822411228336
conv24838419248576
pool2_conv38294419283136
conv3110592256110848
fc4294912512295424
score102481032
location2048162064

Onet

weight

layer nameweight data byteweight bias bytetotal byte
conv134561283584
conv23686412836992
conv37372825673984
conv4_147456256147712
conv5_294912512295424
conv6_589824512590336
fc5117964810241180672
score204882056
location4096164112

6.2 feature大小

Rnet

feature

feature namebytesize
rgb6912
rgb_pad8112
conv1_out64512
conv1_out_pad70000
pool_conv1_out16128
pool_conv1_out_pad21952
conv2_out27648
conv2_out_pad32448
pool2_conv3_out6912
pool2_conv3_out_pad9408
conv3_out2304
fc4_out512
score_8
location_16

 

Onet

feature

feature namebytesize
rgb27648
rgb_pad30000
conv1_out294912
conv1_out_pad307328
conv2_out73728
conv2_out_pad86528
conv3_out147456
conv3_out_pad160000
conv4_out36864
conv4_out_pad43264
conv5_out18432
conv5_out_pad25088
conv6_out4608
fc5_out1024
score_8
location_16
//scales 0.400000  h:116  w:142
//scales 0.340000  h:98  w:120
//scales 0.289000  h:84  w:102
//scales 0.245650  h:72  w:88
//scales 0.208803  h:62  w:74
//scales 0.177482  h:52  w:64
//scales 0.150860  h:44  w:54
//scales 0.128231  h:38  w:46
//scales 0.108996  h:32  w:40
//scales 0.092647  h:28  w:34
//scales 0.078750  h:24  w:28
//scales 0.066937  h:20  w:24
//scales 0.056897  h:18  w:22
//scales 0.048362  h:14  w:18
Create Pnet,Read Pnet.bin ,12 pointers
------Start find Face function
scales 0.400000  h:116  w:142
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 197664 ----------------
--------convolution buffer byte size 638400 -------
--------Feature pad buffer byte size 648600 -------
--------convolution buffer byte size 255360 -------
--------convolution buffer byte size 478720 -------
--------convolution buffer byte size 447744 -------
--------convolution buffer byte size 27984 -------
--------convolution buffer byte size 55968 -------
scales 0.340000  h:98  w:120
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 141120 ----------------
--------convolution buffer byte size 453120 -------
--------Feature pad buffer byte size 461720 -------
--------convolution buffer byte size 181248 -------
--------convolution buffer byte size 335616 -------
--------convolution buffer byte size 309760 -------
--------convolution buffer byte size 19360 -------
--------convolution buffer byte size 38720 -------
scales 0.289000  h:84  w:102
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 102816 ----------------
--------convolution buffer byte size 328000 -------
--------Feature pad buffer byte size 335320 -------
--------convolution buffer byte size 131200 -------
--------convolution buffer byte size 239616 -------
--------convolution buffer byte size 217856 -------
--------convolution buffer byte size 13616 -------
--------convolution buffer byte size 27232 -------
scales 0.245650  h:72  w:88
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 76032 ----------------
--------convolution buffer byte size 240800 -------
--------Feature pad buffer byte size 247080 -------
--------convolution buffer byte size 96320 -------
--------convolution buffer byte size 173184 -------
--------convolution buffer byte size 154752 -------
--------convolution buffer byte size 9672 -------
--------convolution buffer byte size 19344 -------
scales 0.208803  h:62  w:74
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 55056 ----------------
--------convolution buffer byte size 172800 -------
--------Feature pad buffer byte size 178120 -------
--------convolution buffer byte size 69120 -------
--------convolution buffer byte size 121856 -------
--------convolution buffer byte size 106496 -------
--------convolution buffer byte size 6656 -------
--------convolution buffer byte size 13312 -------
scales 0.177482  h:52  w:64
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 39936 ----------------
--------convolution buffer byte size 124000 -------
--------Feature pad buffer byte size 128520 -------
--------convolution buffer byte size 49600 -------
--------convolution buffer byte size 85376 -------
--------convolution buffer byte size 72576 -------
--------convolution buffer byte size 4536 -------
--------convolution buffer byte size 9072 -------
scales 0.150860  h:44  w:54
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 28512 ----------------
--------convolution buffer byte size 87360 -------
--------Feature pad buffer byte size 91160 -------
--------convolution buffer byte size 34944 -------
--------convolution buffer byte size 58368 -------
--------convolution buffer byte size 47872 -------
--------convolution buffer byte size 2992 -------
--------convolution buffer byte size 5984 -------
scales 0.128231  h:38  w:46
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 20976 ----------------
--------convolution buffer byte size 63360 -------
--------Feature pad buffer byte size 66600 -------
--------convolution buffer byte size 25344 -------
--------convolution buffer byte size 40960 -------
--------convolution buffer byte size 32256 -------
--------convolution buffer byte size 2016 -------
--------convolution buffer byte size 4032 -------
scales 0.108996  h:32  w:40
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 15360 ----------------
--------convolution buffer byte size 45600 -------
--------Feature pad buffer byte size 48360 -------
--------convolution buffer byte size 18240 -------
--------convolution buffer byte size 28288 -------
--------convolution buffer byte size 21120 -------
--------convolution buffer byte size 1320 -------
--------convolution buffer byte size 2640 -------
scales 0.092647  h:28  w:34
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 11424 ----------------
--------convolution buffer byte size 33280 -------
--------Feature pad buffer byte size 35640 -------
--------convolution buffer byte size 13312 -------
--------convolution buffer byte size 19712 -------
--------convolution buffer byte size 13824 -------
--------convolution buffer byte size 864 -------
--------convolution buffer byte size 1728 -------
scales 0.078750  h:24  w:28
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 8064 ----------------
--------convolution buffer byte size 22880 -------
--------Feature pad buffer byte size 24840 -------
--------convolution buffer byte size 9152 -------
--------convolution buffer byte size 12672 -------
--------convolution buffer byte size 8064 -------
--------convolution buffer byte size 504 -------
--------convolution buffer byte size 1008 -------
scales 0.066937  h:20  w:24
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 5760 ----------------
--------convolution buffer byte size 15840 -------
--------Feature pad buffer byte size 17480 -------
--------convolution buffer byte size 6336 -------
--------convolution buffer byte size 8064 -------
--------convolution buffer byte size 4480 -------
--------convolution buffer byte size 280 -------
--------convolution buffer byte size 560 -------
scales 0.056897  h:18  w:22
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 4752 ----------------
--------convolution buffer byte size 12800 -------
--------Feature pad buffer byte size 14280 -------
--------convolution buffer byte size 5120 -------
--------convolution buffer byte size 6144 -------
--------convolution buffer byte size 3072 -------
--------convolution buffer byte size 192 -------
--------convolution buffer byte size 384 -------
scales 0.048362  h:14  w:18
Start run Pnet
Pnet feature buffer init
----------image to matrix buffer byte size 3024 ----------------
--------convolution buffer byte size 7680 -------
--------Feature pad buffer byte size 8840 -------
--------convolution buffer byte size 3072 -------
--------convolution buffer byte size 3072 -------
--------convolution buffer byte size 1024 -------
--------convolution buffer byte size 64 -------
--------convolution buffer byte size 128 -------

 

 

 

 

 

 

 

 

c++ MTCNN
jacke121的专栏
09-16 2780
https://github.com/szad670401/Fast-MTCNN 70ms一张，可能用的cpu //Created by Jack Yu #include "stdafx.h" #include <fstream> #include <iostream> #include <opencv2/opencv.hpp> #include &...
mtcnn c++
jacke121的专栏
09-17 2971
ncnn版的：这个好像没有关键点 https://github.com/cpuimage/MTCNN 原版是这个，有关键点 这个把最小人脸改为80，i7处理器10ms处理一张， https://github.com/moli232777144/mtcnn_ncnn ncnn 编译： https://github.com/Tencent/ncnn 这个与上面类似： htt...
mtcnn-light C++ 无框架实现
03-07
Mtcnn-light的C++实现，无需任何框架，将caffe模型转换为txt格式，cpu下进行。
MTCNN人脸检测 附完整C++代码
weixin_34404393的博客
05-05 477
人脸检测识别一直是图像算法领域一个主流话题。 前年SeetaFace开源了人脸识别引擎，一度成为热门话题。 虽然后来SeetaFace又放出来 2.0版本，但是，我说但是。。。 没有训练代码，想要自己训练一下模型那可就犯难了。 虽然可以阅读源码，从前向传播的角度，反过来实现训练代码， 但是谁有那个闲功夫和时间，去折腾这个呢？ 有的时候还是要站在巨人的肩膀上，你才能看得更远。 ...
Windows下cmake编译caffe，实现纯C++版本MTCNN人脸检测和关键点定位
热门推荐
XZZPPP的博客
06-14 1万+
2017.5.28 发现 caffe官方windows版本已经不提供vs工程文件了，需要用cmake编译生成sln文件。详细编译调试过程如下：   一、开发环境要求：         Windows 7/10，64位系统，Visual Studio 2013以上版本，推荐使用VS2015，安装anaconda2.4或以上版本。   二、下载caffe windows版本源码，生成vs
mtcnn_opencv实现mtcnn_人脸检测MTCNN_equaljck_mtcnn_
10-02
【描述】"MTCNN实现人脸实时检测，c++实现编译动态库方式调用" 描述了项目的核心功能，即实现实时的人脸检测，并且是通过C++语言编写的，编译成动态库以便于在不同应用中调用。这意味着代码库不仅包含了模型的实现，...
基于 MTCNN 的人脸检测，使用 kotlin + java + NDK 编写实现.zip
03-23
【标题】"基于 MTCNN 的人脸检测，使用 kotlin + java + NDK 编写实现.zip" 描述了一个项目，该项目旨在实现一种基于 Multi-Task Cascaded Convolutional Neural Networks（MTCNN）的人脸检测技术。MTCNN 是一个深度...
MTCNN-NCNN: C++基于ncnn框架的人脸检测与界标实现
MTCNN（Multi-task Cascaded Convolutional Networks）是一种高效的人脸检测算法，它通过级联的卷积神经网络来同时完成人脸检测和关键点定位的任务。该算法分为三个阶段：P-Net（Proposal Network）、R-Net（Refine ...
将matlab版MTCNN迁移到C++的VS工程实现
2. 多任务级联卷积神经网络（MTCNN）的C++实现： 将MTCNN从Matlab转换为C++并应用于Visual Studio工程中，涉及到了深度学习框架的C++接口的使用。例如，使用Caffe、TensorFlow C++ API或其他支持C++的深度学习框架...
Caffe模型转txt文件的C++实现
03-07
`Caffe` 是一个流行的深度学习框架，广泛用于训练和部署神经网络模型。本话题主要关注如何使用 `C++` 将 `Caffe` 模型转换为 `txt` 文件，以便在不依赖 `Caffe` 框架的情况下调用模型。这种转换对那些希望在嵌入式...
mtcnn-opencv：仅使用OpenCVC ++ MTCNN推理
02-18
mtcnn-opencv 使用此实现，您的机器上不需要Caffe即可进行MTCNN推理。 OpenCV使用提供推理支持的DNN模块。 该模块能够从各种流行的框架（例如Caffe，tensorflow，darknet等）中获取模型和权重。 更多信息在这里 神经网络 [ZHANG2016] Zhang，K.，Zhang，Z.，Li，Z.，and Qiao，Y.（2016）。 使用多任务级联卷积网络进行联合人脸检测和对齐。 IEEE信号处理快报，23（10）：1499–1503。 快速开始 要求 OpenCV 4.1以上 CMake 3.2+ 建造 git clone https://github.com/egcode/mtcnn-opencv.git cd mtcnn-opencv mkdir build cd build cmake .. make 跑 cd mtcnn-opencv
卷积神经网 CNN C++ 实现
05-15
卷积神经网 CNN C++的 实现 基于 mnist 数据 里面已经包含数据集
ncnn-mtcnn
11-28
1.ncnn源代码及模型 2.mtcnn源代码以及个人对mtcnn理解注释
MTCNN人脸检测 附完整C++代码
Fenglin6165的专栏
08-14 506
参照： https://www.cnblogs.com/cpuimage/p/8995600.html https://github.com/Tencent/ncnn https://github.com/cpuimage/MTCNN
Mxnet MTCNN C++
douzi949389的博客
12-19 1656
Mxnet C++ 接口的找了好久,才找到的,虽然有很多坑,慢慢优化吧  #include &lt;iostream&gt; #include &lt;fstream&gt; #include &lt;map&gt; #include &lt;string&gt; #include &lt;vector&gt; #include &lt;time.h&gt; // opencv #includ...
人脸检测多尺度检测函数c语言,MTCNN人脸检测 附完整C++代码
weixin_29038345的博客
05-21 575
人脸检测识别一直是图像算法领域一个主流话题。前年SeetaFace开源了人脸识别引擎，一度成为热门话题。虽然后来SeetaFace又放出来 2.0版本，但是，我说但是。。。没有训练代码，想要自己训练一下模型那可就犯难了。虽然可以阅读源码，从前向传播的角度，反过来实现训练代码，但是谁有那个闲功夫和时间，去折腾这个呢？有的时候还是要站在巨人的肩膀上，你才能看得更远。而SeetaFace不算巨...
c++编写神经网络（二）MTCNN的程序主程序
祥瑞的技术博客
01-17 648
背景：我们得出了MTCNN的内存调用，内存调用是开辟了一系列的存储空间。 目的：固定相应的存储的大小，编写主程序。 目录 一、MTCNN内存占用表 1.1 权重尺寸 6.2 feature大小 二、程序思路 一、MTCNN内存占用表 1.1 权重尺寸 Pnet weight layer name weight data byte weig...
采用mtcnn输出的特征点做人脸对齐C++代码
xiexiecn的专栏
07-16 2175
之前开源过一个ios项目做简单的人脸对比（https://github.com/xuduo35/ncnn-mtcnn-facenet），发现没有做人脸对齐，今天研究了下补上。 1. deep insight提供的一个头文件FaceProprocess.h，见最后 2. 从https://github.com/deepinsight/insightface/issues/481找来的一段code...
【亲测免费】 **探索人脸检测新境界：MTCNN的C++实现**
最新发布
gitblog_00529的博客
08-29 827
探索人脸检测新境界：MTCNN的C++实现 在人脸识别领域，准确而高效的人脸检测算法是构建强大应用的基础。今天，我们为您介绍一个重量级的开源项目——MTCNN的C++实现，它将多任务级联卷积神经网络（MTCNN）的力量带入了C++的世界，为不同深度学习框架提供了完美支持。 项目介绍 MTCNN是一种前沿的脸部检测算法，以其高精度和快速响应著称。本项目通过精心设计的C++代码，实现了这一算法，并兼容...
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

祥瑞Coding

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付：¥1
获取中
扫码支付

您的余额不足，请更换扫码支付或充值

打赏作者

实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值