📄 dhmm_hq.cpp
字号:
// DHMM_HQ.cpp:
// Implementation of the DHMM_HQ Module.
// That is the C rewritten
// of previous Matlab DHMM program by HeQiang.
//
// Created 2001/08, By DongMing, MDSR.
//
//////////////////////////////////////////////////////////////////////
#include "stdafx.h"
#include "DHMM_HQ.h"
#include "kwspot.h"
#include <math.h>
#include "DHMM_GL.h"
extern PRO_CONFIG u_Pro_Config;
#define VQ_TRAIN_CODE_BOOK_LOOP_TIME 40 // 设置VQ中K-mean的最大迭代次数
// The best loop: 100, better than 80
#define MODEL_TRAIN_CODE_BOOK_LOOP_TIME 32 // 设置模型训练时的最大迭代次数
// 20, 40 are better
//////////////////////////////////////////////////////////////////////
// Private Function Head
int DHMM_Model_Calculate_Gamma(WORD_SAMPLE * pu_Word_Sample, DHMM_MODEL * pu_DHMM_Model);
//////////////////////////////////////////////////////////////////////
// API functions
int DHMM_VQ_Train_Code_Book_HQ(DYNA_2DIM_DOUBLE_ARRAY d2dda_Code_Word, int n_Code_Word_Num, int n_Code_Word_Dim,
DYNA_2DIM_DOUBLE_ARRAY d2dda_Initial_Code_Book, DYNA_2DIM_DOUBLE_ARRAY d2dda_Code_Book, int n_Code_Book_Size)
{
int * pn_Initial_Code_Book_Index;
int * pn_Code_Word_Min_Index; // 存储每个特征的最近类中心
double * pd_Code_Word_Min_Distance; // 存储每个特征到最近类中心的距离
int n_Loop_Index, n_Feature_Index, n_Code_Word_Index, n_Dim_Index;
double dTmp;
double d_Old_Distance, d_Distance; // 上次与此次的类内平均距离
int nTmp;
BOOL boolTmp;
PRO_LOG("\tVQ = HQ, loop = %4d.\n", VQ_TRAIN_CODE_BOOK_LOOP_TIME);
// 若设置初始码本，拷贝之
if (d2dda_Initial_Code_Book != NULL)
{
for (n_Code_Word_Index = 0; n_Code_Word_Index < n_Code_Book_Size; n_Code_Word_Index++)
{
memcpy(d2dda_Code_Book[n_Code_Word_Index], d2dda_Initial_Code_Book[n_Code_Word_Index], sizeof(double) * n_Code_Word_Dim);
}
}
// 若没有设置初始码本，则随机生成
else
{
pn_Initial_Code_Book_Index = new int[n_Code_Book_Size];
ASSERT(pn_Initial_Code_Book_Index);
for (n_Code_Word_Index = 0; n_Code_Word_Index < n_Code_Book_Size; n_Code_Word_Index++)
{
do {
pn_Initial_Code_Book_Index[n_Code_Word_Index] = int((double(rand()) / (double(RAND_MAX) + 1)) * n_Code_Word_Num);
boolTmp = FALSE;
for (nTmp = 0; nTmp < n_Code_Word_Index; nTmp++)
{
if (pn_Initial_Code_Book_Index[n_Code_Word_Index] == pn_Initial_Code_Book_Index[nTmp]) boolTmp = TRUE;
}
} while (boolTmp);
}
for (n_Code_Word_Index = 0; n_Code_Word_Index < n_Code_Book_Size; n_Code_Word_Index++)
{
memcpy(d2dda_Code_Book[n_Code_Word_Index], d2dda_Code_Word[pn_Initial_Code_Book_Index[n_Code_Word_Index]], sizeof(double) * n_Code_Word_Dim);
}
delete[] pn_Initial_Code_Book_Index;
}
pn_Code_Word_Min_Index = new int[n_Code_Word_Num];
pd_Code_Word_Min_Distance = new double[n_Code_Word_Num];
ASSERT((pn_Code_Word_Min_Index != NULL) && (pd_Code_Word_Min_Distance != NULL));
d_Distance = MAX_DOUBLE_VALUE;
for (n_Loop_Index = 0; n_Loop_Index < VQ_TRAIN_CODE_BOOK_LOOP_TIME; n_Loop_Index++)
{
DEBUG_PRINTF("VQ:\tLoop = %4d of %d.\n", n_Loop_Index, VQ_TRAIN_CODE_BOOK_LOOP_TIME);
// 计算每个特征的最近类中心及距离
for (n_Feature_Index = 0; n_Feature_Index < n_Code_Word_Num; n_Feature_Index++)
{
pn_Code_Word_Min_Index[n_Feature_Index] = -1;
pd_Code_Word_Min_Distance[n_Feature_Index] = MAX_DOUBLE_VALUE;
for (n_Code_Word_Index = 0; n_Code_Word_Index < n_Code_Book_Size; n_Code_Word_Index++)
{
dTmp = 0.0F;
for (n_Dim_Index = 0; n_Dim_Index < n_Code_Word_Dim; n_Dim_Index++)
{
dTmp += (d2dda_Code_Word[n_Feature_Index][n_Dim_Index] - d2dda_Code_Book[n_Code_Word_Index][n_Dim_Index])
* (d2dda_Code_Word[n_Feature_Index][n_Dim_Index] - d2dda_Code_Book[n_Code_Word_Index][n_Dim_Index]);
}
if (dTmp < pd_Code_Word_Min_Distance[n_Feature_Index])
{
pn_Code_Word_Min_Index[n_Feature_Index] = n_Code_Word_Index;
pd_Code_Word_Min_Distance[n_Feature_Index] = dTmp;
}
}
}
// 计算类内平均距离
d_Old_Distance = d_Distance;
d_Distance = 0.0F;
for (n_Feature_Index = 0; n_Feature_Index < n_Code_Word_Num; n_Feature_Index++)
{
d_Distance += pd_Code_Word_Min_Distance[n_Feature_Index];
}
d_Distance /= n_Code_Word_Num;
DEBUG_PRINTF("\tAverage Distance = %10.4f, Changed %15.4E.\n", d_Distance, ((d_Old_Distance - d_Distance) / d_Old_Distance));
// 按最近距离原则给每类分派样点，并用该类样点的平均值更新类中心
for (n_Code_Word_Index = 0; n_Code_Word_Index < n_Code_Book_Size; n_Code_Word_Index++)
{
nTmp = 0;
for (n_Dim_Index = 0; n_Dim_Index < n_Code_Word_Dim; n_Dim_Index++)
{
d2dda_Code_Book[n_Code_Word_Index][n_Dim_Index] = 0.0F;
}
for (n_Feature_Index = 0; n_Feature_Index < n_Code_Word_Num; n_Feature_Index++)
{
if (pn_Code_Word_Min_Index[n_Feature_Index] == n_Code_Word_Index)
{
nTmp++;
for (n_Dim_Index = 0; n_Dim_Index < n_Code_Word_Dim; n_Dim_Index++)
{
d2dda_Code_Book[n_Code_Word_Index][n_Dim_Index] += d2dda_Code_Word[n_Feature_Index][n_Dim_Index];
}
}
}
// 该类分到了样点，直接计算
if (nTmp != 0)
{
for (n_Dim_Index = 0; n_Dim_Index < n_Code_Word_Dim; n_Dim_Index++)
{
d2dda_Code_Book[n_Code_Word_Index][n_Dim_Index] /= nTmp;
}
}
// 该类没有分到样点，则丢弃原类中心，随机选中一个样点作为该类中心
else
{
do {
nTmp = int((double(rand()) / (double(RAND_MAX) + 1)) * n_Code_Word_Num);
boolTmp = FALSE;
if (pd_Code_Word_Min_Distance[nTmp] < EPSILON_DOUBLE_VALUE) boolTmp = TRUE;
} while (boolTmp);
memcpy(d2dda_Code_Book[n_Code_Word_Index], d2dda_Code_Word[nTmp], sizeof(double) * n_Code_Word_Dim);
}
}
}
PRO_LOG("VQ:\tLoop = %4d, Average Distance = %10.4f, Changed %15.4E.\n", n_Loop_Index, d_Distance, ((d_Old_Distance - d_Distance) / d_Old_Distance));
delete[] pn_Code_Word_Min_Index;
delete[] pd_Code_Word_Min_Distance;
return 0;
}
int DHMM_Model_Train_DHMM_Model_HQ(WORD_SAMPLE * pu_Word_Sample, int n_Word_Sample_Num,
DHMM_MODEL * pu_DHMM_Model)
{
DYNA_2DIM_DOUBLE_ARRAY d2dda_Nom; // 状态数 * 输出个数
double * pd_Denom; // 状态数
int n_Loop_Index, n_State_Index, n_Code_Word_Index, n_Word_Sample_Index, n_Frame_Index;
int nTmp;
double d_Old_Total_Likelihood, d_Total_Likelihood; // 上次与此次迭代的模型对所有词条匹配分数的加和
double dTmp;
PRO_LOG("\tModel = HQ, loop = %4d.\n", MODEL_TRAIN_CODE_BOOK_LOOP_TIME);
// 初始化Pi，初始状态为状态0
for (n_State_Index = 0; n_State_Index < pu_DHMM_Model->n_State_Num; n_State_Index++)
{
pu_DHMM_Model->pdPi[n_State_Index] = 0.0F;
}
pu_DHMM_Model->pdPi[0] = 1.0F;
// 初始化A，只能停留或跳向下一个状态，并且等概0.5，最后一个状态只能停留
for (n_State_Index = 0; n_State_Index < pu_DHMM_Model->n_State_Num; n_State_Index++)
{
for (nTmp = 0; nTmp < pu_DHMM_Model->n_State_Num; nTmp++)
{
pu_DHMM_Model->d2dda_A[n_State_Index][nTmp] = 0.0F;
}
}
for (n_State_Index = 0; n_State_Index < (pu_DHMM_Model->n_State_Num - 1); n_State_Index++)
{
pu_DHMM_Model->d2dda_A[n_State_Index][n_State_Index] = 0.5F;
pu_DHMM_Model->d2dda_A[n_State_Index][n_State_Index + 1] = 0.5F;
}
pu_DHMM_Model->d2dda_A[pu_DHMM_Model->n_State_Num - 1][pu_DHMM_Model->n_State_Num - 1] = 1.0F;
// 初始化B，均匀初始化，每个状态输出各个输出码字的概率均等，为码字总数的倒数
for (n_State_Index = 0; n_State_Index < pu_DHMM_Model->n_State_Num; n_State_Index++)
{
for (n_Code_Word_Index = 0; n_Code_Word_Index < pu_DHMM_Model->n_Code_Book_Size; n_Code_Word_Index++)
{
pu_DHMM_Model->d2dda_B[n_State_Index][n_Code_Word_Index] = 1.0F / pu_DHMM_Model->n_Code_Book_Size;
}
}
d2dda_Nom = d2dda_New(pu_DHMM_Model->n_State_Num, pu_DHMM_Model->n_Code_Book_Size);
pd_Denom = new double[pu_DHMM_Model->n_State_Num];
ASSERT((d2dda_Nom != NULL) && (pd_Denom != NULL));
for (n_Word_Sample_Index = 0; n_Word_Sample_Index < n_Word_Sample_Num; n_Word_Sample_Index++)
{
pu_Word_Sample[n_Word_Sample_Index].d2dda_Gamma = d2dda_New(pu_Word_Sample[n_Word_Sample_Index].n_Feature_Sequence_Len, pu_DHMM_Model->n_State_Num);
ASSERT(pu_Word_Sample[n_Word_Sample_Index].d2dda_Gamma != NULL);
}
d_Total_Likelihood = -MAX_DOUBLE_VALUE;
for (n_Loop_Index = 0; n_Loop_Index < MODEL_TRAIN_CODE_BOOK_LOOP_TIME; n_Loop_Index++)
{
// 计算每个词条的Gamma系数，二维数组，帧数 * 状态数
for (n_Word_Sample_Index = 0; n_Word_Sample_Index < n_Word_Sample_Num; n_Word_Sample_Index++)
{
DHMM_Model_Calculate_Gamma(&pu_Word_Sample[n_Word_Sample_Index], pu_DHMM_Model);
}
// 准备重估B
for (n_State_Index = 0; n_State_Index < pu_DHMM_Model->n_State_Num; n_State_Index++)
{
pd_Denom[n_State_Index] = 0.0F;
for (n_Code_Word_Index = 0; n_Code_Word_Index < pu_DHMM_Model->n_Code_Book_Size; n_Code_Word_Index++)
{
d2dda_Nom[n_State_Index][n_Code_Word_Index] = 0.0F;
}
}
// 计算重估B时每项的分子，和每行的分母
for (n_State_Index = 0; n_State_Index < pu_DHMM_Model->n_State_Num; n_State_Index++)
{
for (n_Word_Sample_Index = 0; n_Word_Sample_Index < n_Word_Sample_Num; n_Word_Sample_Index++)
{
for (n_Frame_Index = 0; n_Frame_Index < pu_Word_Sample[n_Word_Sample_Index].n_Feature_Sequence_Len; n_Frame_Index++)
{
d2dda_Nom[n_State_Index][pu_Word_Sample[n_Word_Sample_Index].pn_VQed_Feature_Sequence[n_Frame_Index]]
+= pu_Word_Sample[n_Word_Sample_Index].d2dda_Gamma[n_Frame_Index][n_State_Index];
pd_Denom[n_State_Index] += pu_Word_Sample[n_Word_Sample_Index].d2dda_Gamma[n_Frame_Index][n_State_Index];
}
}
}
// 重新计算B
for (n_State_Index = 0; n_State_Index < pu_DHMM_Model->n_State_Num; n_State_Index++)
{
for (n_Code_Word_Index = 0; n_Code_Word_Index < pu_DHMM_Model->n_Code_Book_Size; n_Code_Word_Index++)
{
pu_DHMM_Model->d2dda_B[n_State_Index][n_Code_Word_Index] =
d2dda_Nom[n_State_Index][n_Code_Word_Index] / pd_Denom[n_State_Index];
}
}
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -