📄 dhmm_model_mfc.cpp
字号:
}
fclose(fp_DHMM_Model_File);
fclose(fp_DHMM_Model_TXT_File);
fclose(fp_DHMM_Model_BIN_File);
d2dda_Free(d2dda_LogB, pu_DHMM_Model->n_State_Num, pu_DHMM_Model->n_Code_Book_Size);
// 释放模型占用内存
for (n_Model_Index = 0; n_Model_Index < n_DHMM_Model_Num; n_Model_Index++)
{
delete[] pu_DHMM_Model[n_Model_Index].pdPi;
d2dda_Free(pu_DHMM_Model[n_Model_Index].d2dda_A, u_Pro_Config.n_DHMM_Model_State_Num, u_Pro_Config.n_DHMM_Model_State_Num);
d2dda_Free(pu_DHMM_Model[n_Model_Index].d2dda_B, u_Pro_Config.n_DHMM_Model_State_Num, u_Pro_Config.n_VQ_Code_Book_Size);
}
delete[] pu_DHMM_Model;
if ((u_Pro_Config.l_DHMM_Model_Config & MODEL_CONFIG_GENERATE_DHMM_MODEL_MASK) == (MODEL_CONFIG_TRAIN_WITH_SILENCE_MODEL))
u_Pro_Config.n_DHMM_Model_State_Num -= 2;
else if((u_Pro_Config.l_DHMM_Model_Config & MODEL_CONFIG_GENERATE_DHMM_MODEL_MASK) == MODEL_CONFIG_LOAD_WITH_SILENCE_MODEL)
u_Pro_Config.n_DHMM_Model_State_Num -= 2;
return 0;
}
//////////////////////////////////////////////////////////////////////
// 函数名称：DHMM_Model_Train_All_DHMM_Model
// 函数功能：训练所有词的DHMM模型
// 函数性质：API
// 输入参数：
// n_DHMM_Model_Num，要训练的模型个数
// 输出参数：
// pu_DHMM_Model，存放训好的DHMM模型
// 返回值：
// 0 表示成功
// 备注：关于训练的细节参数，均在u_Pro_Config中指定，
// n_Feature_Person_Num，总的语料库人数
// n_DHMM_Model_Train_Set_Person_Num，训练集的人数
// n_DHMM_Model_Person_Start_Index，训练集起始人的标号，
// 训练集将从此人顺序持续指定人数，标号越界自动折返到0
int DHMM_Model_Train_All_DHMM_Model(DHMM_MODEL * pu_DHMM_Model, int n_DHMM_Model_Num)
{
int nRetCode;
char sz_Feature_File_Name[256];
DYNA_2DIM_DOUBLE_ARRAY d2dda_Code_Book;
WORD_SAMPLE * pu_Word_Sample;
int n_Word_Sample_Index, n_Person_Index, n_Train_Set_Person_Index;
// 读入码书
d2dda_Code_Book = d2dda_New(u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim);
ASSERT(d2dda_Code_Book != NULL);
nRetCode = DHMM_VQ_Load_Code_Book_File(u_Pro_Config.sz_Toload_Code_Book_File_Name,
d2dda_Code_Book, u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim);
ASSERT(nRetCode == 0);
ASSERT((0 <= u_Pro_Config.n_DHMM_Model_Person_Start_Index) && (u_Pro_Config.n_DHMM_Model_Person_Start_Index < u_Pro_Config.n_Feature_Person_Num)
&& (0 <= u_Pro_Config.n_DHMM_Model_Person_End_Index) && (u_Pro_Config.n_DHMM_Model_Person_End_Index < u_Pro_Config.n_Feature_Person_Num));
// 为训练集的词条准备内存
pu_Word_Sample = new WORD_SAMPLE[u_Pro_Config.n_DHMM_Model_Train_Set_Person_Num];
ASSERT(pu_Word_Sample != NULL);
// 该重循环，训练每一个词条的模型
for (n_Word_Sample_Index = 0; n_Word_Sample_Index < u_Pro_Config.n_DHMM_Model_Num; n_Word_Sample_Index++)
{
int Sen_Index = n_Word_Sample_Index + u_Pro_Config.n_Sentence_Start_Index;
// 该重循环，对一个词条，用所有训练集进行训练的数据准备
n_Train_Set_Person_Index = 0;
for (n_Person_Index = 0; n_Person_Index < u_Pro_Config.n_Feature_Person_Num; n_Person_Index++)
{
// 跳过测试集的人
if (u_Pro_Config.n_DHMM_Model_Person_Start_Index <= u_Pro_Config.n_DHMM_Model_Person_End_Index)
{
if ((n_Person_Index < u_Pro_Config.n_DHMM_Model_Person_Start_Index) || (n_Person_Index > u_Pro_Config.n_DHMM_Model_Person_End_Index)) continue;
}
else
{
if ((n_Person_Index > u_Pro_Config.n_DHMM_Model_Person_End_Index) && (n_Person_Index < u_Pro_Config.n_DHMM_Model_Person_Start_Index)) continue;
}
// 准备该词条的内存空间
sprintf(sz_Feature_File_Name, u_Pro_Config.sz_Feature_Origin_File_Name_Format, n_Person_Index);
pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len = dfa_Feature_Get_Sentence_Frame_Num(sz_Feature_File_Name, Sen_Index);
pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Dim = u_Pro_Config.n_Feature_Dim;
pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence =
d2dda_New(pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len, pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Dim);
pu_Word_Sample[n_Train_Set_Person_Index].pn_VQed_Feature_Sequence =
new int[pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len];
ASSERT((pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence != NULL)
&& (pu_Word_Sample[n_Train_Set_Person_Index].pn_VQed_Feature_Sequence != NULL));
// 读入特征参数
nRetCode = dfa_Feature_Read_A_Sentence(sz_Feature_File_Name, Sen_Index, u_Pro_Config.n_Feature_Dim,
pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence);
ASSERT(nRetCode == pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len);
// VQ特征
DHMM_VQ_Encode_A_Word_Sample(d2dda_Code_Book, u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim,
&pu_Word_Sample[n_Train_Set_Person_Index]);
// 释放原始的特征参数占用空间
d2dda_Free(pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence,
pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len, pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Dim);
n_Train_Set_Person_Index++;
}
ASSERT(n_Train_Set_Person_Index == u_Pro_Config.n_DHMM_Model_Train_Set_Person_Num);
PRO_LOG("Model:\tTraining Model %4d of %4d.\n", n_Word_Sample_Index, u_Pro_Config.n_DHMM_Model_Num);
// 训练一个词条的模型
DHMM_Model_Train_DHMM_Model(pu_Word_Sample, u_Pro_Config.n_DHMM_Model_Train_Set_Person_Num, &pu_DHMM_Model[n_Word_Sample_Index]);
// 释放VQ后的特征占用的内存
for (n_Train_Set_Person_Index = 0; n_Train_Set_Person_Index < u_Pro_Config.n_DHMM_Model_Train_Set_Person_Num; n_Train_Set_Person_Index++)
{
delete[] pu_Word_Sample[n_Train_Set_Person_Index].pn_VQed_Feature_Sequence;
}
}
// 释放所有训练集的词条所占内存，释放码书内存
delete[] pu_Word_Sample;
d2dda_Free(d2dda_Code_Book, u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim);
return 0;
}
//////////////////////////////////////////////////////////////////////
// 函数名称：DHMM_Model_Train_Silence_DHMM_Model
// 函数功能：训练静音的单状态DHMM模型
// 函数性质：API
// 输入参数：
// n_DHMM_Model_Num，要训练的模型个数
// 输出参数：
// pu_DHMM_Model，存放训好的DHMM模型
// 返回值：
// 0 表示成功
// 备注：关于训练的细节参数，均在u_Pro_Config中指定，
// n_Feature_Person_Num，总的语料库人数
// n_DHMM_Model_Train_Set_Person_Num，训练集的人数
// n_DHMM_Model_Person_Start_Index，训练集起始人的标号，
// 训练集将从此人顺序持续指定人数，标号越界自动折返到0
int DHMM_Model_Train_Silence_DHMM_Model(DHMM_MODEL * pu_DHMM_Model)
{
int nRetCode;
char sz_Feature_File_Name[256];
char sz_Silence_File_Name[256];
DYNA_2DIM_DOUBLE_ARRAY d2dda_Code_Book;
WORD_SAMPLE * pu_Word_Sample;
int n_Word_Sample_Index, n_Person_Index, n_Train_Set_Person_Index;
int n_DHMM_Model_Train_Set_Person_Num;
int i, j/*, k, n, m*/;
int Status_num = 1;
int Output_num = u_Pro_Config.n_VQ_Code_Book_Size; // VQ码本大小（码字个数）
double *Pi; // 初始概率
double **A; // A矩阵：Status_num*Status_num
double **B; // B矩阵：Status_num*Output_num
double w;
// Generate (A, B, Pi)
Pi = pu_DHMM_Model->pdPi; // The above block transform to this line. #DongMing#
// allocate memory for A matrix
A = pu_DHMM_Model->d2dda_A; // The above block transform to this line. #DongMing#
// allocate memory for B matrix
B = pu_DHMM_Model->d2dda_B; // The above block transform to this line. #DongMing#
// 初始化初始概率Pi:状态0概率为1，其余状态概率为0
Pi[0] = 1;
for(i=1; i<Status_num; i++) Pi[i] = 0;
for(i=0; i<Status_num; i++)
{
// 初始化A矩阵
for(j=0; j<Status_num; j++) A[i][j] = 0;
if(i < Status_num-1)
{
A[i][i] = 0.5;
A[i][i+1] = 0.5;
}
else
{
A[i][i] = 1.0;
}
// 归一化A矩阵
w = 0;
for(j=0; j<Status_num; j++) w = w + A[i][j];
for(j=0; j<Status_num; j++) A[i][j] = A[i][j] / w;
w = 0;
// 初始化B矩阵
for(j=0; j<Output_num; j++)
{
B[i][j] = 1.0 / Output_num * (0.9 + ((rand()%1000) / 5000.0));
w = w + B[i][j];
}
// 归一化B矩阵
for(j=0; j<Output_num; j++) B[i][j] = 0; //B[i][j] / w;
}
// 读入码书
d2dda_Code_Book = d2dda_New(u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim);
ASSERT(d2dda_Code_Book != NULL);
nRetCode = DHMM_VQ_Load_Code_Book_File(u_Pro_Config.sz_Toload_Code_Book_File_Name,
d2dda_Code_Book, u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim);
ASSERT(nRetCode == 0);
ASSERT((0 <= u_Pro_Config.n_DHMM_Model_Person_Start_Index) && (u_Pro_Config.n_DHMM_Model_Person_Start_Index < u_Pro_Config.n_Feature_Person_Num)
&& (0 <= u_Pro_Config.n_DHMM_Model_Person_End_Index) && (u_Pro_Config.n_DHMM_Model_Person_End_Index < u_Pro_Config.n_Feature_Person_Num));
// 为训练集的词条准备内存
pu_Word_Sample = new WORD_SAMPLE[u_Pro_Config.n_DHMM_Model_Train_Set_Person_Num * u_Pro_Config.n_DHMM_Model_Num];
ASSERT(pu_Word_Sample != NULL);
n_Train_Set_Person_Index = 0;
// 该重循环，加载每一个词条的静音数据
for (n_Word_Sample_Index = 0; n_Word_Sample_Index < u_Pro_Config.n_DHMM_Model_Num; n_Word_Sample_Index++)
{
int Sen_Index = n_Word_Sample_Index + u_Pro_Config.n_Sentence_Start_Index;
// 该重循环，对一个词条，用所有训练集进行训练的数据准备
for (n_Person_Index = 0; n_Person_Index < u_Pro_Config.n_Feature_Person_Num; n_Person_Index++)
{
// 跳过测试集的人
if (u_Pro_Config.n_DHMM_Model_Person_Start_Index <= u_Pro_Config.n_DHMM_Model_Person_End_Index)
{
if ((n_Person_Index < u_Pro_Config.n_DHMM_Model_Person_Start_Index) || (n_Person_Index > u_Pro_Config.n_DHMM_Model_Person_End_Index)) continue;
}
else
{
if ((n_Person_Index > u_Pro_Config.n_DHMM_Model_Person_End_Index) && (n_Person_Index < u_Pro_Config.n_DHMM_Model_Person_Start_Index)) continue;
}
// 修改文件名
sprintf(sz_Feature_File_Name, u_Pro_Config.sz_Feature_Origin_File_Name_Format, n_Person_Index);
strcpy(sz_Silence_File_Name, sz_Feature_File_Name);
int j = strlen(sz_Silence_File_Name) - 4;
sz_Silence_File_Name[j] = '\0';
strcat(sz_Silence_File_Name,"_sl.dat");
// 准备该词条的内存空间
pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len = dfa_Feature_Get_Silence_Frame_Num(sz_Silence_File_Name, Sen_Index);
//pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len = 2 * RELAX_FRAME;
pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Dim = u_Pro_Config.n_Feature_Dim;
pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence =
d2dda_New(pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len, pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Dim);
pu_Word_Sample[n_Train_Set_Person_Index].pn_VQed_Feature_Sequence =
new int[pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len];
ASSERT((pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence != NULL)
&& (pu_Word_Sample[n_Train_Set_Person_Index].pn_VQed_Feature_Sequence != NULL));
// 读入特征参数
nRetCode = dfa_Feature_Read_A_Silence(sz_Silence_File_Name, Sen_Index, u_Pro_Config.n_Feature_Dim,
pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence);
ASSERT(nRetCode == pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len);
// VQ特征
DHMM_VQ_Encode_A_Word_Sample(d2dda_Code_Book, u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim,
&pu_Word_Sample[n_Train_Set_Person_Index]);
// 释放原始的特征参数占用空间
d2dda_Free(pu_Word_Sample[n_Train_Set_Person_Index].d2dda_Feature_Sequence,
pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Sequence_Len, pu_Word_Sample[n_Train_Set_Person_Index].n_Feature_Dim);
n_Train_Set_Person_Index++;
}
}
PRO_LOG("Model:\tTraining Silence Model...\n");
n_DHMM_Model_Train_Set_Person_Num = n_Train_Set_Person_Index;
int fr_num;
for(i = 0; i < n_DHMM_Model_Train_Set_Person_Num; i++)
{
fr_num = pu_Word_Sample[i].n_Feature_Sequence_Len;
for(j = 0; j < fr_num; j++)
B[0][pu_Word_Sample[i].pn_VQed_Feature_Sequence[j]]++;
}
// 释放VQ后的特征占用的内存
for (n_Train_Set_Person_Index = 0; n_Train_Set_Person_Index < n_DHMM_Model_Train_Set_Person_Num; n_Train_Set_Person_Index++)
{
delete[] pu_Word_Sample[n_Train_Set_Person_Index].pn_VQed_Feature_Sequence;
}
for(i = 0; i < Output_num; i++)
{
B[0][i] /= (fr_num * n_DHMM_Model_Train_Set_Person_Num);
if (B[0][i] < Epsilon_B)
B[0][i] = Epsilon_B;
}
for(j=0; j<Output_num; j++)
w = w + B[0][j];
// 归一化B矩阵
for(j=0; j<Output_num; j++) B[0][j] = B[0][j] / w;
// 释放所有训练集的词条所占内存，释放码书内存
delete[] pu_Word_Sample;
d2dda_Free(d2dda_Code_Book, u_Pro_Config.n_VQ_Code_Book_Size, u_Pro_Config.n_Feature_Dim);
return 0;
}
//////////////////////////////////////////////////////////////////////
// 函数名称：DHMM_Model_Train_DHMM_Model
// 函数功能：训练一个词的DHMM模型
// 函数性质：API
// 输入参数：
// pu_Word_Sample，存放该词训练集的数据
// n_Word_Sample_Num，训练集的词数
// 输出参数：
// pu_DHMM_Model，存放训好的DHMM模型
// 返回值：
// 0 表示成功
// 备注：WORD_SAMPLE的定义，参见公共的.H文件
// 该函数是一个分发函数，按照u_Pro_Config.l_DHMM_Model_Config配置的信息
// 将具体调用不同的函数体
int DHMM_Model_Train_DHMM_Model(WORD_SAMPLE * pu_Word_Sample, int n_Word_Sample_Num,
DHMM_MODEL * pu_DHMM_Model)
{
switch (u_Pro_Config.l_DHMM_Model_Config & MODEL_CONFIG_TRAIN_PROCEDURE_MASK)
{
case MODEL_CONFIG_TRAIN_PROCEDURE_STD:
ASSERT(0);
break;
case MODEL_CONFIG_TRAIN_PROCEDURE_GL:
return DHMM_Model_Train_DHMM_Model_GL(pu_Word_Sample, n_Word_Sample_Num, pu_DHMM_Model);
break;
case MODEL_CONFIG_TRAIN_PROCEDURE_LHS:
ASSERT(0); //LHS的训练过程，与GL的完全相同，故略去
break;
case MODEL_CONFIG_TRAIN_PROCEDURE_HQ:
return DHMM_Model_Train_DHMM_Model_HQ(pu_Word_Sample, n_Word_Sample_Num, pu_DHMM_Model);
break;
case MODEL_CONFIG_TRAIN_PROCEDURE_WP:
ASSERT(0);
break;
default:
ASSERT(0);
break;
}
return 0;
}
⌨️ 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 + -