From 81ac69fa9e3c10e8b357f8725fe60c503fe31fab Mon Sep 17 00:00:00 2001
From: xinyang <txy15436@gmail.com>
Date: Sat, 20 Apr 2019 14:37:16 +0800
Subject: [PATCH] =?UTF-8?q?=E7=8E=B0=E5=8F=AF=E4=BD=BF=E7=94=A8CNN?=
 =?UTF-8?q?=E5=88=86=E7=B1=BB=E5=99=A8=E5=88=A4=E6=96=AD=E8=A3=85=E7=94=B2?=
 =?UTF-8?q?=E6=9D=BF=E5=8F=8A=E5=85=B6=E6=95=B0=E5=AD=97=E3=80=82?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../armor_finder/classifier/classifier.h      |  51 ++++
 .../armor_finder/classifier/classifier.cpp    | 269 ++++++++++++++++++
 2 files changed, 320 insertions(+)
 create mode 100644 armor/include/armor_finder/classifier/classifier.h
 create mode 100644 armor/src/armor_finder/classifier/classifier.cpp

diff --git a/armor/include/armor_finder/classifier/classifier.h b/armor/include/armor_finder/classifier/classifier.h
new file mode 100644
index 0000000..57b3bc6
--- /dev/null
+++ b/armor/include/armor_finder/classifier/classifier.h
@@ -0,0 +1,51 @@
+//
+// Created by xinyang on 19-4-19.
+//
+
+#ifndef _CLASSIFIER_H_
+#define _CLASSIFIER_H_
+
+#include <Eigen/Dense>
+#include <opencv2/core/eigen.hpp>
+#include <vector>
+#include <string>
+#include <opencv2/core.hpp>
+
+using namespace std;
+using namespace Eigen;
+
+class Classifier {
+private:
+    bool state;
+
+    vector<vector<MatrixXd>> conv1_w, conv2_w;
+    vector<double> conv1_b, conv2_b;
+    MatrixXd fc1_w, fc2_w;
+    VectorXd fc1_b, fc2_b;
+
+    vector<vector<MatrixXd>> load_conv_w(const string &file);
+    vector<double> load_conv_b(const string &file);
+    MatrixXd load_fc_w(const string &file);
+    VectorXd load_fc_b(const string &file);
+
+    MatrixXd softmax(const MatrixXd &input);
+    MatrixXd relu(const MatrixXd &input);
+    vector<vector<MatrixXd>> apply_bias(const vector<vector<MatrixXd>> &input, const vector<double> &bias);
+    vector<vector<MatrixXd>> relu(const vector<vector<MatrixXd>> &input);
+    vector<vector<MatrixXd>> pool(const vector<vector<MatrixXd>> &input, int size);
+    vector<vector<MatrixXd>> pand(const vector<vector<MatrixXd>> &input, int val);
+    MatrixXd conv(const MatrixXd &filter, const MatrixXd &input);
+    vector<vector<MatrixXd>> conv2(const vector<vector<MatrixXd>> &filter, const vector<vector<MatrixXd>> &input);
+    MatrixXd flatten(const vector<vector<MatrixXd>> &input);
+
+public:
+    explicit Classifier(const string &folder);
+    ~Classifier() = default;
+
+    MatrixXd calculate(const vector<vector<MatrixXd>> &input);
+    explicit operator bool() const;
+    int operator()(const cv::Mat &image);
+};
+
+
+#endif //RUNCNN_CLASSIFIER_H
diff --git a/armor/src/armor_finder/classifier/classifier.cpp b/armor/src/armor_finder/classifier/classifier.cpp
new file mode 100644
index 0000000..b303375
--- /dev/null
+++ b/armor/src/armor_finder/classifier/classifier.cpp
@@ -0,0 +1,269 @@
+//
+// Created by xinyang on 19-4-19.
+//
+
+#include <armor_finder/classifier/classifier.h>
+#include <log.h>
+#include <cstdio>
+#include <iostream>
+
+
+vector<vector<MatrixXd>> Classifier::load_conv_w(const string &file){
+    vector<vector<MatrixXd>> result;
+    FILE *fp = fopen(file.data(), "r");
+    if(fp == nullptr){
+        LOGE("%s open fail!", file.data());
+        state = false;
+        return result;
+    }
+    int channel_in, channel_out, row, col;
+    fscanf(fp, "%d %d %d %d", &channel_in, &channel_out, &row, &col);
+    for(int o=0; o<channel_in; o++){
+        vector<MatrixXd> sub;
+        for(int i=0; i<channel_out; i++){
+            MatrixXd f(row, col);
+            for(int r=0; r<row; r++){
+                for(int c=0; c<col; c++){
+                    fscanf(fp, "%lf", &f(r, c));
+                }
+            }
+            sub.emplace_back(f);
+        }
+        result.emplace_back(sub);
+    }
+    return result;
+}
+
+vector<double> Classifier::load_conv_b(const string &file){
+    vector<double> result;
+    FILE *fp = fopen(file.data(), "r");
+    if(fp == nullptr){
+        LOGE("%s open fail!", file.data());
+        state = false;
+        return result;
+    }
+    int len;
+    fscanf(fp, "%d", &len);
+    for(int i=0; i<len; i++) {
+        double v;
+        fscanf(fp, "%lf", &v);
+        result.emplace_back(v);
+    }
+    return result;
+}
+
+MatrixXd Classifier::load_fc_w(const string &file){
+    FILE *fp = fopen(file.data(), "r");
+    if(fp == nullptr){
+        LOGE("%s open fail!", file.data());
+        state = false;
+        return MatrixXd::Zero(1,1);
+    }
+    int row, col;
+    fscanf(fp, "%d %d", &col, &row);
+    MatrixXd mat(row, col);
+    for(int c=0; c<col; c++){
+        for(int r=0; r<row; r++){
+            fscanf(fp, "%lf", &mat(r, c));
+        }
+    }
+    return mat;
+}
+
+VectorXd Classifier::load_fc_b(const string &file){
+    FILE *fp = fopen(file.data(), "r");
+    if(fp == nullptr){
+        LOGE("%s open fail!", file.data());
+        state = false;
+        return VectorXd::Zero(1,1);
+    }
+    int row;
+    fscanf(fp, "%d", &row);
+    VectorXd vec(row, 1);
+    for(int r=0; r<row; r++){
+        fscanf(fp, "%lf", &vec(r));
+    }
+    return vec;
+}
+
+MatrixXd Classifier::softmax(const MatrixXd &input){
+    MatrixXd tmp = input.array() - input.maxCoeff();
+    return tmp.array().exp() / tmp.array().exp().sum();
+}
+
+vector<vector<MatrixXd>> Classifier::pool(const vector<vector<MatrixXd>> &input, int size){
+    vector<vector<MatrixXd>> output;
+    for(int sample=0; sample<input.size(); sample++) {
+        vector<MatrixXd> sub;
+        for (int channel = 0; channel < input[0].size(); channel++) {
+            MatrixXd tmp(input[0][0].rows() / size, input[0][0].cols() / size);
+            for (int row = 0; row < input[0][0].rows() / size; row++) {
+                for (int col = 0; col < input[0][0].cols() / size; col++) {
+                    double val = 0;
+                    for (int x = 0; x < size; x++) {
+                        for (int y = 0; y < size; y++) {
+                            val += input[sample][channel](row * size + x, col * size + y);
+                        }
+                    }
+                    tmp(row, col) = val / size / size;
+                }
+            }
+            sub.emplace_back(tmp);
+        }
+        output.emplace_back(sub);
+    }
+    return output;
+}
+
+vector<vector<MatrixXd>> Classifier::apply_bias(const vector<vector<MatrixXd>> &input, const vector<double> &bias){
+    assert(input[0].size()==bias.size());
+    vector<vector<MatrixXd>> result;
+    for(int samples=0; samples<input.size(); samples++){
+        vector<MatrixXd> sub;
+        for(int channels=0; channels<input[0].size(); channels++){
+            MatrixXd mat = input[samples][channels].array() + bias[channels];
+            sub.emplace_back(mat);
+        }
+        result.emplace_back(sub);
+    }
+    return result;
+}
+
+MatrixXd Classifier::relu(const MatrixXd &input){
+    return input.unaryExpr([](double val){
+       return (val>0)?(val):(0);
+    });
+}
+
+vector<vector<MatrixXd>> Classifier::relu(const vector<vector<MatrixXd>> &input){
+    vector<vector<MatrixXd>> result;
+    for(int samples=0; samples<input.size(); samples++){
+        vector<MatrixXd> sub;
+        for(int channels=0; channels<input[0].size(); channels++){
+            sub.emplace_back(relu(input[samples][channels]));
+        }
+        result.emplace_back(sub);
+    }
+    return result;
+}
+
+vector<vector<MatrixXd>> Classifier::pand(const vector<vector<MatrixXd>> &input, int val){
+    vector<vector<MatrixXd>> result;
+    for(int sample=0; sample<input.size(); sample++){
+        vector<MatrixXd> sub;
+        for(int channels=0; channels<input[0].size(); channels++){
+            MatrixXd mat = MatrixXd::Zero(input[0][0].rows()+2*val, input[0][0].cols()+2*val);
+            mat.block(val, val, input[0][0].rows(), input[0][0].cols()) = input[sample][channels];
+            sub.emplace_back(mat);
+        }
+        result.emplace_back(sub);
+    }
+    return result;
+}
+
+MatrixXd Classifier::conv(const MatrixXd &filter, const MatrixXd &input){
+    int result_rows = input.rows()-filter.rows()+1;
+    int result_cols = input.cols()-filter.cols()+1;
+    MatrixXd result(result_rows, result_cols);
+    for(int row=0; row<result_rows; row++){
+        for(int col=0; col<result_cols; col++){
+            double val=0;
+            for(int x=0; x<filter.cols(); x++){
+                for(int y=0; y<filter.cols(); y++){
+                    val += input(row+x, col+y) * filter(x,y);
+                }
+            }
+            result(row, col) = val;
+//            result(row, col) = (input.block(row, col, size, size).array() * input.array()).sum();
+        }
+    }
+    return result;
+}
+
+vector<vector<MatrixXd>> Classifier::conv2(const vector<vector<MatrixXd>> &filter, const vector<vector<MatrixXd>> &input){
+    if(filter.size() != input[0].size()){
+        LOGE("shape du not match, which is filter.size=%d, input[0].size()=%d", filter.size(), input[0].size());
+        exit(-1);
+    }
+    vector<vector<MatrixXd>> result;
+    int result_rows = input[0][0].rows()-filter[0][0].rows()+1;
+    int result_cols = input[0][0].cols()-filter[0][0].cols()+1;
+    for(int col=0; col<input.size(); col++){
+        vector<MatrixXd> sub;
+        for(int row=0; row<filter[0].size(); row++){
+            MatrixXd val = MatrixXd::Zero(result_rows, result_cols);
+            for(int x=0; x<filter.size(); x++){
+                val += conv(filter[x][row], input[col][x]);
+            }
+            sub.emplace_back(val);
+        }
+        result.emplace_back(sub);
+    }
+    return result;
+}
+
+MatrixXd Classifier::flatten(const vector<vector<MatrixXd>> &input){
+    int ones = input[0][0].rows()*input[0][0].cols();
+    int channels = input[0].size();
+    int samples = input.size();
+    int row = input[0][0].rows();
+    int col = input[0][0].cols();
+    MatrixXd output(channels*ones,samples);
+    for(int s=0; s<samples; s++) {
+        for(int r=0, cnt=0; r<row; r++){
+            for(int c=0; c<col; c++){
+                for (int i = 0; i < channels; i++) {
+                    output(cnt++, s) = input[s][i](r, c);
+                }
+            }
+        }
+
+    }
+    return output;
+}
+
+Classifier::Classifier(const string &folder) : state(true){
+    conv1_w = load_conv_w(folder+"conv1_w");
+    conv1_b = load_conv_b(folder+"conv1_b");
+    conv2_w = load_conv_w(folder+"conv2_w");
+    conv2_b = load_conv_b(folder+"conv2_b");
+    fc1_w = load_fc_w(folder+"fc1_w");
+    fc1_b = load_fc_b(folder+"fc1_b");
+    fc2_w = load_fc_w(folder+"fc2_w");
+    fc2_b = load_fc_b(folder+"fc2_b");
+}
+
+//#define PRINT_MAT(name) (cout << #name":\n" << name << endl)
+//#define PRINT_MULTI_MAT(name) (cout << #name":\n" << name[0][0] << endl)
+//#define PRINT_MAT_SHAPE(name) LOGM(#name":(%d, %d)", name.rows(), name.cols())
+//#define PRINT_MULTI_MAT_SHAPE(name) LOGM(#name":(%d, %d)", name[0][0].rows(), name[0][0].cols())
+
+MatrixXd Classifier::calculate(const vector<vector<MatrixXd>> &input) {
+    vector<vector<MatrixXd>> conv1_result = relu(apply_bias(conv2(conv1_w, input), conv1_b));
+    vector<vector<MatrixXd>> pool1_result = pool(conv1_result, 2);
+    vector<vector<MatrixXd>> conv2_result = relu(apply_bias(conv2(conv2_w, pool1_result), conv2_b));
+    vector<vector<MatrixXd>> pool2_result = pool(conv2_result, 2);
+    MatrixXd flattened = flatten(pool2_result);
+    MatrixXd y1 = fc1_w * flattened;
+    y1.colwise() += fc1_b;
+    MatrixXd fc1 = relu(y1);
+    MatrixXd y2 = fc2_w * fc1;
+    y2.colwise() += fc2_b;
+    MatrixXd fc2 = softmax(y2);
+    return fc2;
+}
+
+Classifier::operator bool() const {
+    return state;
+}
+
+int Classifier::operator()(const cv::Mat &image) {
+    MatrixXd x;
+    cv2eigen(image, x);
+    vector<MatrixXd> sub = {x};
+    vector<vector<MatrixXd>> in = {sub};
+    MatrixXd result = calculate(in);
+    MatrixXd::Index minRow, minCol;
+    result.maxCoeff(&minRow, &minCol);
+    return minRow;
+}
\ No newline at end of file