1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
const char *help = "\
speech_hmm_train (c) Samy Bengio & Co 2001\n\
\n\
This program is used to do speech recognition \n";
#include <torch/LexiconInfo.h>
#include <torch/EMTrainer.h>
#include <torch/HMM.h>
#include <torch/DiagonalGMM.h>
#include <torch/MatDataSet.h>
#include <torch/CmdLine.h>
#include <torch/NLLMeasurer.h>
#include <torch/Random.h>
#include <torch/DiskHTKDataSet.h>
#include <torch/HTKDataSet.h>
#include <torch/SpeechHMM.h>
#include <torch/FileListCmdOption.h>
#include <torch/string_utils.h>
using namespace Torch;
#include "speech_include.cc"
int main(int argc, char **argv)
{
char* phoneme_name;
char* lex_name;
char* silence_name;
char* silence_phone;
char* sp_name;
char* sp_phone;
real accuracy;
real threshold;
int max_iter;
real prior;
int max_load;
int the_seed;
char *dir_name;
char *model_file;
char *init_model;
bool add_sil_to_targets;
bool htk_model;
bool viterbi;
bool train_separate;
bool disk;
Allocator *allocator = new Allocator;
FileListCmdOption input_file_list("file name", "the list of inputs files or one data file");
input_file_list.isArgument(true);
FileListCmdOption target_file_list("file name", "the list of target files or one target file");
target_file_list.isArgument(true);
//===============================================================
//=================== The command-line ==========================
//===============================================================
// Construct the command line
CmdLine cmd;
// Put the help line at the beginning
cmd.info(help);
// Train mode
cmd.addText("\nArguments:");
cmd.addSCmdArg("phoneme_name", &phoneme_name, "the list of phonemes file");
cmd.addSCmdArg("lex_name", &lex_name, "the lexicon file");
cmd.addCmdOption(&input_file_list);
cmd.addCmdOption(&target_file_list);
cmd.addSCmdArg("init model file", &init_model, "the initialized model file");
cmd.addText("\nModel Options:");
cmd.addSCmdOption("-silence_word", &silence_name,"sil", "name of silence word");
cmd.addSCmdOption("-silence_phone", &silence_phone,"h#", "name of silence phone");
cmd.addSCmdOption("-sp_word", &sp_name,"", "name of short pause word");
cmd.addSCmdOption("-sp_phone", &sp_phone,"", "name of short pause phoneme");
cmd.addBCmdOption("-add_sil_to_targets", &add_sil_to_targets, false, "add silence at begining of targets");
cmd.addText("\nLearning Options:");
cmd.addBCmdOption("-train_separate", &train_separate, false, "first train separate models");
cmd.addBCmdOption("-viterbi", &viterbi, false, "viterbi training (instead of EM)");
cmd.addRCmdOption("-threshold", &threshold, 0.001, "variance threshold");
cmd.addRCmdOption("-prior", &prior, 0.001, "prior on the weights");
cmd.addICmdOption("-iter", &max_iter, 25, "max number of iterations of HMM");
cmd.addRCmdOption("-e", &accuracy, 0.00001, "end accuracy");
cmd.addText("\nMisc Options:");
cmd.addICmdOption("-seed", &the_seed, -1, "the random seed");
cmd.addICmdOption("-load", &max_load, -1, "max number of examples to load for train");
cmd.addSCmdOption("-dir", &dir_name, ".", "directory to save measures");
cmd.addSCmdOption("-save", &model_file, "", "the model file");
cmd.addBCmdOption("-htk_model", &htk_model, false, "load and save in HTK model format");
cmd.addBCmdOption("-disk", &disk, false, "keep data on disk");
// Read the command line
cmd.read(argc, argv);
cmd.setWorkingDirectory(dir_name);
DiskXFile::setBigEndianMode();
//====================================================================
//=================== Training Mode =================================
//====================================================================
if(the_seed == -1)
Random::seed();
else
Random::manualSeed((long)the_seed);
// read lexicon
if (strlen(sp_phone) == 0)
sp_phone = NULL;
if (strlen(silence_phone) == 0)
silence_phone = NULL;
LexiconInfo lex(phoneme_name,silence_phone,sp_phone,lex_name,add_sil_to_targets ? silence_name : NULL,NULL);
//====================================================================
//=================== Create the DataSet ... =========================
//====================================================================
// some basic tests on the files
if (input_file_list.n_files != target_file_list.n_files) {
error("the input and target files should have the same number of files (%d != %d)\n",input_file_list.n_files,target_file_list.n_files);
}
for (int i=0;i<input_file_list.n_files;i++) {
char* si = strRemoveSuffix(input_file_list.file_names[i]);
char* st = strRemoveSuffix(target_file_list.file_names[i]);
if (strcmp(strBaseName(si),strBaseName(st)))
warning("input file (%s) do not correspond to target file (%s)",input_file_list.file_names[i],target_file_list.file_names[i]);
free(si);
free(st);
}
DataSet* data;
if (disk)
data = (DataSet*)new(allocator) DiskHTKDataSet(input_file_list.file_names,target_file_list.file_names, input_file_list.n_files, true, max_load, &lex);
else
data = (DataSet*)new(allocator) HTKDataSet(input_file_list.file_names,target_file_list.file_names, input_file_list.n_files, true, max_load, &lex);
//=================== Create the HMM... =========================
real* thresh = (real*)allocator->alloc(data->n_inputs*sizeof(real));
initializeThreshold(data,thresh,threshold);
SpeechHMM* shmm = NULL;
if (htk_model)
shmm = newSpeechHMMFromHTK(cmd.getPath(init_model),&lex,thresh,prior,allocator);
else
shmm = newSpeechHMMFromTorch(cmd.getPath(init_model),&lex,thresh,prior,allocator,&cmd);
//=================== Measurers and Trainer ===============================
// Measurers on the training dataset
MeasurerList measurers;
char* n_nll = strConcat(2,"hmm_train_val_",viterbi ? "viterbi" : "em");
allocator->retain(n_nll);
NLLMeasurer nll_meas(shmm->log_probabilities, data, cmd.getXFile(n_nll));
measurers.addNode(&nll_meas);
// The Gradient Machine Trainer
EMTrainer trainer(shmm);
trainer.setIOption("max iter", max_iter);
trainer.setROption("end accuracy", accuracy);
if (viterbi)
trainer.setBOption("viterbi", true);
//=================== Let's go... ===============================
trainer.train(data, &measurers);
if(strcmp(model_file, "")) {
if (htk_model) {
save_htk_model(cmd.getPath(model_file),shmm,lex.phone_info->phone_names);
} else {
DiskXFile model_(cmd.getPath(model_file), "w");
cmd.saveXFile(&model_);
int n_models = shmm->n_models;
int n_gaussians = ((DiagonalGMM*)((HMM*)shmm->models[n_models-2])->states[1])->n_gaussians;
int* states = (int*)allocator->alloc(sizeof(int)*n_models);
for (int i=0;i<n_models;i++)
states[i] = shmm->models[i]->n_states;
model_.taggedWrite(&n_gaussians, sizeof(int), 1, "n_gaussians");
model_.taggedWrite(&n_models, sizeof(int), 1, "n_phonemes");
model_.taggedWrite(states,sizeof(int), n_models, "n_states");
model_.taggedWrite(&data->n_inputs, sizeof(int), 1, "n_inputs");
shmm->saveXFile(&model_);
}
}
delete allocator;
return(0);
}
|
