about lnet i want to transfer to ncnn.

[zys1994]

about lnet i want to transfer to ncnn. about my code is that.

#include <iostream>
#include <stdio.h>
#include <algorithm>
#include <vector>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include "platform.h"
#include "net.h"
#include "mat.h"

class Noop : public ncnn::Layer {};
DEFINE_LAYER_CREATOR(Noop)

int main() {
    ncnn::Net net;

    net.register_custom_layer("LinearRegressionOutput", Noop_layer_creator);
    net.register_custom_layer("Custom", Noop_layer_creator);
    net.load_param("../model/L106.param");
    net.load_model("../model/L106.bin");


    cv::Mat img = cv::imread("../test_img/AE_.jpg");

    cv::resize(img, img, cv::Size(96, 96));
    cv::Mat img_cp =img.clone();
    cv::cvtColor(img, img, CV_BGR2RGB);  //mxnet支持的也是rgb
    unsigned char *rgbdata = img.data;
    ncnn::Mat in = ncnn::Mat::from_pixels(rgbdata, ncnn::Mat::PIXEL_RGB, 96, 96);
    //ncnn::Mat in = ncnn::Mat::from_pixels(rgbdata, ncnn::Mat::PIXEL_BGR, 96, 96);
    const float mean_vals[3] = {127.5f, 127.5f, 127.5f};
    const float norm_vals[3] = {0.0078125f, 0.0078125f, 0.0078125f};
    in.substract_mean_normalize(mean_vals, norm_vals);


    ncnn::Mat out;

    ncnn::Extractor ex = net.create_extractor();
    //ex.set_light_mode(true);
    ex.input("data", in);


    double t_start = cv::getTickCount();
    ex.extract("conv6_3", out);
    double t_end = cv::getTickCount();
    float costTime = (t_end - t_start) / cv::getTickFrequency();
    std::cout << "cost Time :" << costTime << std::endl;


    std::cout << "Hello, World!" << std::endl;
    ncnn::Mat out_flatterned = out.reshape(out.w * out.h * out.c);
    for (int i = 0; i < 106; i++)
    {
        cv::Point pt = cv::Point(96 * out_flatterned[i * 2], 96 * out_flatterned[i * 2 + 1]);
        //cv::Point pt = cv::Point(96 * out_flatterned[i ], 96 * out_flatterned[i+106]);
        cv::circle(img_cp, pt, 2, cv::Scalar(0, 0, 250), 2);
    }
    cv::imshow("show",img_cp);
    cv::waitKey(0);
  
    return 0;
}

The result is not correct, Can you share your experience with me ?

[aidlearning]

you don't need this :
net.register_custom_layer("LinearRegressionOutput", Noop_layer_creator);
net.register_custom_layer("Custom", Noop_layer_creator);
because forward don't need this layer!

[zys1994]

if i comment it, it will get wrong.

 layer LinearRegressionOutput not exists or registered

it doesn't matter. i turns mxnet model to ncnn. i wonder why i get wrong result. the index landmark is wrong? or the means and scales is wrong?

[aidlearning]

you must delete the layers after conv6_3 because of the layers LinearRegressionOutput is for training and don't need in forward.

[aidlearning]

I mean you open the L106.param ,and delete the layers after conv6_3 in the file.

[aidlearning]

if you like the project ,pls give me a star,thank u!

[aidlearning]

7767517
66 67
Input data 0 1 data
Convolution conv1 1 1 data conv1 0=32 1=3 11=3 5=1 6=864
BatchNorm bn1 1 1 conv1 bn1 0=32
PReLU prelu1 1 1 bn1 prelu1 0=32
ConvolutionDepthWise conv2_dw 1 1 prelu1 conv2_dw 0=32 1=2 11=2 5=1 6=128 7=32
BatchNorm bn2_dw 1 1 conv2_dw bn2_dw 0=32
PReLU prelu2_dw 1 1 bn2_dw prelu2_dw 0=32
Convolution conv2_sep 1 1 prelu2_dw conv2_sep 0=32 1=1 11=1 5=1 6=1024
BatchNorm bn2_sep 1 1 conv2_sep bn2_sep 0=32
PReLU prelu2_sep 1 1 bn2_sep prelu2_sep 0=32
ConvolutionDepthWise conv3_dw 1 1 prelu2_sep conv3_dw 0=32 1=3 11=3 3=2 13=2 5=1 6=288 7=32
BatchNorm bn3_dw 1 1 conv3_dw bn3_dw 0=32
PReLU prelu3_dw 1 1 bn3_dw prelu3_dw 0=32
Convolution conv3_sep 1 1 prelu3_dw conv3_sep 0=64 1=1 11=1 5=1 6=2048
BatchNorm bn3_sep 1 1 conv3_sep bn3_sep 0=64
PReLU prelu3_sep 1 1 bn3_sep prelu3_sep 0=64
ConvolutionDepthWise conv4_dw 1 1 prelu3_sep conv4_dw 0=64 1=2 11=2 5=1 6=256 7=64
BatchNorm bn4_dw 1 1 conv4_dw bn4_dw 0=64
PReLU prelu4_dw 1 1 bn4_dw prelu4_dw 0=64
Convolution conv4_sep 1 1 prelu4_dw conv4_sep 0=64 1=1 11=1 5=1 6=4096
BatchNorm bn4_sep 1 1 conv4_sep bn4_sep 0=64
PReLU prelu4_sep 1 1 bn4_sep prelu4_sep 0=64
ConvolutionDepthWise conv5_dw 1 1 prelu4_sep conv5_dw 0=64 1=3 11=3 3=2 13=2 5=1 6=576 7=64
BatchNorm bn5_dw 1 1 conv5_dw bn5_dw 0=64
PReLU prelu5_dw 1 1 bn5_dw prelu5_dw 0=64
Convolution conv5_sep 1 1 prelu5_dw conv5_sep 0=64 1=1 11=1 5=1 6=4096
BatchNorm bn5_sep 1 1 conv5_sep bn5_sep 0=64
PReLU prelu5_sep 1 1 bn5_sep prelu5_sep 0=64
ConvolutionDepthWise conv6_dw 1 1 prelu5_sep conv6_dw 0=64 1=2 11=2 5=1 6=256 7=64
BatchNorm bn6_dw 1 1 conv6_dw bn6_dw 0=64
PReLU prelu6_dw 1 1 bn6_dw prelu6_dw 0=64
Convolution conv6_sep 1 1 prelu6_dw conv6_sep 0=64 1=1 11=1 5=1 6=4096
BatchNorm bn6_sep 1 1 conv6_sep bn6_sep 0=64
PReLU prelu6_sep 1 1 bn6_sep prelu6_sep 0=64
ConvolutionDepthWise conv7_dw 1 1 prelu6_sep conv7_dw 0=64 1=3 11=3 3=2 13=2 5=1 6=576 7=64
BatchNorm bn7_dw 1 1 conv7_dw bn7_dw 0=64
PReLU prelu7_dw 1 1 bn7_dw prelu7_dw 0=64
Convolution conv7_sep 1 1 prelu7_dw conv7_sep 0=128 1=1 11=1 5=1 6=8192
BatchNorm bn7_sep 1 1 conv7_sep bn7_sep 0=128
PReLU prelu7_sep 1 1 bn7_sep prelu7_sep 0=128
ConvolutionDepthWise conv8_dw 1 1 prelu7_sep conv8_dw 0=128 1=2 11=2 5=1 6=512 7=128
BatchNorm bn8_dw 1 1 conv8_dw bn8_dw 0=128
PReLU prelu8_dw 1 1 bn8_dw prelu8_dw 0=128
Convolution conv8_sep 1 1 prelu8_dw conv8_sep 0=128 1=1 11=1 5=1 6=16384
BatchNorm bn8_sep 1 1 conv8_sep bn8_sep 0=128
PReLU prelu8_sep 1 1 bn8_sep prelu8_sep 0=128
ConvolutionDepthWise conv9_dw 1 1 prelu8_sep conv9_dw 0=128 1=3 11=3 3=2 13=2 5=1 6=1152 7=128
BatchNorm bn9_dw 1 1 conv9_dw bn9_dw 0=128
PReLU prelu9_dw 1 1 bn9_dw prelu9_dw 0=128
Convolution conv9_sep 1 1 prelu9_dw conv9_sep 0=256 1=1 11=1 5=1 6=32768
BatchNorm bn9_sep 1 1 conv9_sep bn9_sep 0=256
PReLU prelu9_sep 1 1 bn9_sep prelu9_sep 0=256
ConvolutionDepthWise conv10_dw 1 1 prelu9_sep conv10_dw 0=256 1=2 11=2 5=1 6=1024 7=256
BatchNorm bn10_dw 1 1 conv10_dw bn10_dw 0=256
PReLU prelu10_dw 1 1 bn10_dw prelu10_dw 0=256
Convolution conv10_sep 1 1 prelu10_dw conv10_sep 0=256 1=1 11=1 5=1 6=65536
BatchNorm bn10_sep 1 1 conv10_sep bn10_sep 0=256
PReLU prelu10_sep 1 1 bn10_sep prelu10_sep 0=256
ConvolutionDepthWise conv11_dw 1 1 prelu10_sep conv11_dw 0=256 1=3 11=3 5=1 6=2304 7=256
BatchNorm bn11_dw 1 1 conv11_dw bn11_dw 0=256
PReLU prelu11_dw 1 1 bn11_dw prelu11_dw 0=256
Convolution conv11_sep 1 1 prelu11_dw conv11_sep 0=256 1=1 11=1 5=1 6=65536
BatchNorm bn11_sep 1 1 conv11_sep bn11_sep 0=256
PReLU prelu11_sep 1 1 bn11_sep prelu11_sep 0=256
InnerProduct conv6_3 1 1 prelu11_sep conv6_3 0=212 1=1 2=54272
BatchNorm bn6_3

like this!

[aidlearning]

L106.paramhttps://github.com/aidlearning/AidLearning-FrameWork/blob/master/examples/landmark106/L106.param

[zys1994]

the result is the same incorrect. your means and scales is [127.5, 127.5, 127.5], [0.0078,0.0078, 0.0078]?

[zys1994]

Do you mind sharing your L106.bin with me to make sure whether i transfer mxnet to ncnn correctly or not. My email is zyslongjuanfeng@gmail.com

[aidlearning]

yes!pls give me a star

[zys1994]

i had given your star when i see your excellent work!

[zys1994]

i have not received the email.
What's the problem?

[aidlearning]

I have send it by postmaster@aidlearning.net.may be blocked by gmail,i sent u again at once!

[zys1994]

try 1158007644@qq.com, thanks

[aidlearning]

ok,i email u again, pls check it

[zys1994]

thanks for your model. i have used your model. but it get incorrect result.

[zys1994]

this is some problem for gmail, 1158007644@qq.com is ok.

[aidlearning]

my models is fine, we use it in projects ,it is working good...

[zys1994]

ok, i use "bn6_3" and get a good result. The SampleLnet106 is conv6-3, leading me a wrong direction. Do you use "bn6_3"?

[aidlearning]

yes! I use the bn6_3

[Tomhouxin]

有个问题请教：

1. 在使用106关键点检测模型之前是否用了mtcnn检测人脸
2. mtcnn ---》ncnn模型转换是怎么转的

[aidlearning]

用了mtcnn人脸模型，用ncnn直接就可以转啊

[Tomhouxin]

这个mtcnn是zuoqing用maxnet训练的吗，然后maxnet2ncnn？

[qidiso]

是的

…

---Original--- From: "Tomhouxin"<notifications@github.com> Date: Fri, May 31, 2019 14:03 PM To: "aidlearning/AidLearning-FrameWork"<AidLearning-FrameWork@noreply.github.com>; Cc: "Subscribed"<subscribed@noreply.github.com>; Subject: Re: [aidlearning/AidLearning-FrameWork] about lnet i want to transfer to ncnn. (#17) 这个mtcnn是zuoqing用maxnet训练的吗，然后maxnet2ncnn？ — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.