我搜了下boost算法的官方代码,居然没有详细的注释,很多代码夹杂了C和C++的东西,读起来不是那么顺畅,所以拼命加注释!
#include "ml.h"
#include "cv.h"
#include "highgui.h"
#include <opencv2/opencv.hpp>
#include <stdio.h>
/**
The sample demonstrates how to train Random Trees classifier
(or Boosting classifier, or MLP - see main()) using the provided dataset.
We use the sample database letter-recognition.data
from UCI Repository, here is the link:
Newman, D.J. & Hettich, S. & Blake, C.L. & Merz, C.J. (1998).
UCI Repository of machine learning databases
[http://www.ics.uci.edu/~mlearn/MLRepository.html].
Irvine, CA: University of California, Department of Information and Computer Science.
The dataset consists of 20000 feature vectors along with the
responses - capital latin letters A..Z.
The first 16000 (10000 for boosting)) samples are used for training
and the remaining 4000 (10000 for boosting) - to test the classifier.
*/
// This function reads data and responses from the file <filename>
//var_count代表什么意思
static int
read_num_class_data( const char* filename, int var_count,
CvMat** data, CvMat** responses )
{
const int M = 1024;
FILE* f = fopen( filename, "rt" );
CvMemStorage* storage; //分配的存储空间
CvSeq* seq; //cv的序列,相当于双向链表的数据结构
char buf[M+2];
float* el_ptr;
CvSeqReader reader;
int i, j;
if( !f )
return 0;
el_ptr = new float[var_count+1];
storage = cvCreateMemStorage();
//cvCreateSeq的第二个参数表示头的大小,一般用sizeof(CvSeq)代替,这里更准确
seq = cvCreateSeq( 0, sizeof(*seq), (var_count+1)*sizeof(float), storage );
for(;;)
{
char* ptr;
if( !fgets( buf, M, f ) || !strchr( buf, ',' ) )
break;
el_ptr[0] = buf[0]; //第一个值赋值,比较特殊,是和数据相关的。
ptr = buf+2;
for( i = 1; i <= var_count; i++ )
{
int n = 0;
sscanf( ptr, "%f%n", el_ptr + i, &n );
ptr += n + 1;
}
if( i <= var_count )
break;
cvSeqPush( seq, el_ptr );
}
fclose(f);
//将传递过来的指针赋值,调用这个函数后,指针就指向了具体的mat
*data = cvCreateMat( seq->total, var_count, CV_32F );
*responses = cvCreateMat( seq->total, 1, CV_32F );
cvStartReadSeq( seq, &reader );
for( i = 0; i < seq->total; i++ )
{
const float* sdata = (float*)reader.ptr + 1;
float* ddata = data[0]->data.fl + var_count*i;
float* dr = responses[0]->data.fl + i;
for( j = 0; j < var_count; j++ )
ddata[j] = sdata[j];
*dr = sdata[-1];
CV_NEXT_SEQ_ELEM( seq->elem_size, reader );
}
cvReleaseMemStorage( &storage );
delete el_ptr;
return 1;
}
static
int build_boost_classifier( char* data_filename,
char* filename_to_save, char* filename_to_load )
{
const int class_count = 26; //分为26个类别
CvMat* data = 0;
CvMat* responses = 0;
//--var_type 定义responses的类型
// -- CV_VAR_CATEGORICAL 分类标签
// -- CV_VAR_ORDERED(CV_VAR_NUMERICAL)数值,用于回归问题
CvMat* var_type = 0;
CvMat* temp_sample = 0; //临时的样本
CvMat* weak_responses = 0; //弱分类器的输出
int ok = read_num_class_data( data_filename, 16, &data, &responses ); //16个特征
int nsamples_all = 0, ntrain_samples = 0; //样本总数(/训练样本)的个数
int var_count; //特征的个数
double train_hr = 0, test_hr = 0;
CvBoost boost;
if( !ok )
{
printf( "Could not read the database %s\n", data_filename );
return -1;
}
printf( "The database %s is loaded.\n", data_filename );
nsamples_all = data->rows;
ntrain_samples = (int)(nsamples_all*0.5); //样本总数的一半用来训练?
var_count = data->cols; //特征的个数
// 生成或者加载已生成的boost分类器
if( filename_to_load )
{
// load classifier from the specified file
boost.load( filename_to_load );
ntrain_samples = 0;
if( !boost.get_weak_predictors() )
{
printf( "Could not read the classifier %s\n", filename_to_load );
return -1;
}
printf( "The classifier %s is loaded.\n", data_filename );
}
else
{
// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
//
// As currently boosted tree classifier in MLL can only be trained
// for 2-class problems, we transform the training database by
// "unrolling" each training sample as many times as the number of
// classes (26) that we have.
//
// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
/*它是把一个multiple-class classifcation的问题,巧妙的变成了binary classification的问题。
每一个原来的sample被重复26次,构建出26个sample data; 其中,对于每一个sample data:
a. 在feature vector中新增了一项feature,这个feature就是 0 - 26 +‘A’ 的 ASCII值;
因为字母种类是categorical的value,所以var_type的最后一列是 CV_VAR_CATEGORICAL;
b. response vector 变成了true/false, true or false取决于当前sample data的字母值
是不是对应于raw sample data的response.*/
CvMat* new_data = cvCreateMat( ntrain_samples*class_count, var_count + 1, CV_32F );
CvMat* new_responses = cvCreateMat( ntrain_samples*class_count, 1, CV_32S );
// 1. unroll the database type mask
printf( "Unrolling the database...\n");
for( int i = 0; i < ntrain_samples; i++ )
{
float* data_row = (float*)(data->data.ptr + data->step*i); //每次指向原始数据的一行
for( int j = 0; j < class_count; j++ ) //小于类型的个数,也就是复制的次数
{
float* new_data_row = (float*)(new_data->data.ptr +
new_data->step*(i*class_count+j)); //找到新分配的数据的行
for( int k = 0; k < var_count; k++ ) //将每行的特征复制过来
new_data_row[k] = data_row[k];
new_data_row[var_count] = (float)j; //最后一个放了复制的序号
new_responses->data.i[i*class_count + j] = responses->data.fl[i] == j+'A'; //新的response取决于是否与原来的值相同
}
}
// 2. create type mask
var_type = cvCreateMat( var_count + 2, 1, CV_8U ); //为啥加2,下面有分析
cvSet( var_type, cvScalarAll(CV_VAR_ORDERED) ); //这些都是数值型的
// the last indicator variable, as well
// as the new (binary) response are categorical
cvSetReal1D( var_type, var_count, CV_VAR_CATEGORICAL ); //这是新加的一位,如前面所说,也应该是离散类型标签
cvSetReal1D( var_type, var_count+1, CV_VAR_CATEGORICAL ); //最后一位作为返回的类型
// 3. train classifier
printf( "Training the classifier (may take a few minutes)...\n");
boost.train( new_data, CV_ROW_SAMPLE, new_responses, 0, 0, var_type, 0,
CvBoostParams(CvBoost::REAL, 100, 0.95, 5, false, 0 ));
//CvBoostParams的参数含义分别是:(1)使用REAL adaboost,
//(2)100个分类器,
//(3)样本总权值小于1.0-0.95 = 0.05的点将不参加下一次的迭代
//(4)训练树的最大深度是5
//(5)不使用代理分裂
//(6)没有定义自己的错分类代价
cvReleaseMat( &new_data );
cvReleaseMat( &new_responses );
printf("\n");
}
temp_sample = cvCreateMat( 1, var_count + 1, CV_32F );
weak_responses = cvCreateMat( 1, boost.get_weak_predictors()->total, CV_32F );
// compute prediction error on train and test data
for( i = 0; i < nsamples_all; i++ )
{
int best_class = 0;
double max_sum = -DBL_MAX;
double r;
CvMat sample;
cvGetRow( data, &sample, i );
for( k = 0; k < var_count; k++ )
temp_sample->data.fl[k] = sample.data.fl[k];
for( j = 0; j < class_count; j++ )
{
temp_sample->data.fl[var_count] = (float)j;
boost.predict( temp_sample, 0, weak_responses );
double sum = cvSum( weak_responses ).val[0];
if( max_sum < sum )
{
max_sum = sum;
best_class = j + 'A';
}
}
r = fabs(best_class - responses->data.fl[i]) < FLT_EPSILON ? 1 : 0;
if( i < ntrain_samples )
train_hr += r;
else
test_hr += r;
}
test_hr /= (double)(nsamples_all-ntrain_samples);
train_hr /= (double)ntrain_samples;
printf( "Recognition rate: train = %.1f%%, test = %.1f%%\n",
train_hr*100., test_hr*100. );
printf( "Number of trees: %d\n", boost.get_weak_predictors()->total );
// Save classifier to file if needed
if( filename_to_save )
boost.save( filename_to_save );
cvReleaseMat( &temp_sample );
cvReleaseMat( &weak_responses );
cvReleaseMat( &var_type );
cvReleaseMat( &data );
cvReleaseMat( &responses );
return 0;
}