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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
|
#include "ObjFun.h"
#include "Error.h"
#include <limits>
namespace TNet {
ObjectiveFunction* ObjectiveFunction::Factory(ObjFunType type) {
ObjectiveFunction* ret = NULL;
switch(type) {
case MEAN_SQUARE_ERROR: ret = new MeanSquareError; break;
case CROSS_ENTROPY: ret = new CrossEntropy; break;
default: Error("Unknown ObjectiveFunction type");
}
return ret;
}
/*
* MeanSquareError
*/
void MeanSquareError::Evaluate(const Matrix<BaseFloat>& net_out, const Matrix<BaseFloat>& target, Matrix<BaseFloat>* err) {
//check dimensions
assert(net_out.Rows() == target.Rows());
assert(net_out.Cols() == target.Cols());
if(err->Rows() != net_out.Rows() || err->Cols() != net_out.Cols()) {
err->Init(net_out.Rows(),net_out.Cols());
}
//compute global gradient
err->Copy(net_out);
err->AddScaled(-1,target);
//compute loss function
double sum = 0;
for(size_t r=0; r<err->Rows(); r++) {
for(size_t c=0; c<err->Cols(); c++) {
BaseFloat val = (*err)(r,c);
sum += val*val;
}
}
error_ += sum/2.0;
frames_ += net_out.Rows();
}
std::string MeanSquareError::Report() {
std::stringstream ss;
ss << "Mse:" << error_ << " frames:" << frames_
<< " err/frm:" << error_/frames_
<< "\n";
return ss.str();
}
/*
* CrossEntropy
*/
///Find maximum in float array
inline int FindMaxId(const BaseFloat* ptr, size_t N) {
BaseFloat mval = -1e20f;
int mid = -1;
for(size_t i=0; i<N; i++) {
if(ptr[i] > mval) {
mid = i; mval = ptr[i];
}
}
return mid;
}
void
CrossEntropy::Evaluate(const Matrix<BaseFloat>& net_out, const Matrix<BaseFloat>& target, Matrix<BaseFloat>* err)
{
//check dimensions
assert(net_out.Rows() == target.Rows());
assert(net_out.Cols() == target.Cols());
if(err->Rows() != net_out.Rows() || err->Cols() != net_out.Cols()) {
err->Init(net_out.Rows(),net_out.Cols());
}
//allocate confunsion buffers
if(confusion_mode_ != NO_CONF) {
if(confusion_.Rows() != target.Cols() || confusion_.Cols() != target.Cols()) {
confusion_.Init(target.Cols(),target.Cols());
confusion_count_.Init(target.Cols());
diag_confusion_.Init(target.Cols());
}
}
//compute global gradient (assuming on softmax input)
err->Copy(net_out);
err->AddScaled(-1,target);
//collect max values
std::vector<size_t> max_target_id(target.Rows());
std::vector<size_t> max_netout_id(target.Rows());
//check correct classification
int corr = 0;
for(size_t r=0; r<net_out.Rows(); r++) {
int id_netout = FindMaxId(net_out[r].pData(),net_out.Cols());
int id_target = FindMaxId(target[r].pData(),target.Cols());
if(id_netout == id_target) corr++;
max_target_id[r] = id_target;//store the max value
max_netout_id[r] = id_netout;
}
//compute loss function
double sumerr = 0;
for(size_t r=0; r<net_out.Rows(); r++) {
if(target(r,max_target_id[r]) == 1.0) {
//pick the max value..., rest is zero
BaseFloat val = log(net_out(r,max_target_id[r]));
if(val < -1e10f) val = -1e10f;
sumerr += val;
} else {
//process whole posterior vect.
for(size_t c=0; c<net_out.Cols(); c++) {
if(target(r,c) != 0.0) {
BaseFloat val = target(r,c)*log(net_out(r,c));
if(val < -1e10f) val = -1e10f;
sumerr += val;
}
}
}
}
//accumulate confusuion network
if(confusion_mode_ != NO_CONF) {
for(size_t r=0; r<net_out.Rows(); r++) {
int id_target = max_target_id[r];
int id_netout = max_netout_id[r];
switch(confusion_mode_) {
case MAX_CONF:
confusion_(id_target,id_netout) += 1;
break;
case SOFT_CONF:
confusion_[id_target].Add(net_out[r]);
break;
case DIAG_MAX_CONF:
diag_confusion_[id_target] += ((id_target==id_netout)?1:0);
break;
case DIAG_SOFT_CONF:
diag_confusion_[id_target] += net_out[r][id_target];
break;
default:
KALDI_ERR << "unknown confusion type" << confusion_mode_;
}
confusion_count_[id_target] += 1;
}
}
error_ -= sumerr;
frames_ += net_out.Rows();
corr_ += corr;
}
std::string CrossEntropy::Report() {
std::stringstream ss;
ss << "Xent:" << error_ << " frames:" << frames_
<< " err/frm:" << error_/frames_
<< " correct[" << 100.0*corr_/frames_ << "%]"
<< "\n";
if(confusion_mode_ != NO_CONF) {
//read class tags
std::vector<std::string> tag;
{
std::ifstream ifs(output_label_map_);
assert(ifs.good());
std::string str;
while(!ifs.eof()) {
ifs >> str;
tag.push_back(str);
}
}
assert(confusion_count_.Dim() <= tag.size());
//print confusion matrix
if(confusion_mode_ == MAX_CONF || confusion_mode_ == SOFT_CONF) {
ss << "Row:label Col:hyp\n" << confusion_ << "\n";
}
//***print per-target accuracies
for(int i=0; i<confusion_count_.Dim(); i++) {
//get the numerator
BaseFloat numerator = 0.0;
switch (confusion_mode_) {
case MAX_CONF: case SOFT_CONF:
numerator = confusion_[i][i];
break;
case DIAG_MAX_CONF: case DIAG_SOFT_CONF:
numerator = diag_confusion_[i];
break;
default:
KALDI_ERR << "Usupported confusion mode:" << confusion_mode_;
}
//add line to report
ss << std::setw(30) << tag[i] << " "
<< std::setw(10) << 100.0*numerator/confusion_count_[i] << "%"
<< " [" << numerator << "/" << confusion_count_[i] << "]\n";
} //***print per-target accuracies
}// != NO_CONF
return ss.str();
}
void CrossEntropy::MergeStats(const ObjectiveFunction& inst) {
const CrossEntropy& xent = dynamic_cast<const CrossEntropy&>(inst);
frames_ += xent.frames_; error_ += xent.error_; corr_ += xent.corr_;
//sum the confustion statistics
if(confusion_mode_ != NO_CONF) {
if(confusion_.Rows() != xent.confusion_.Rows()) {
confusion_.Init(xent.confusion_.Rows(),xent.confusion_.Cols());
confusion_count_.Init(xent.confusion_count_.Dim());
diag_confusion_.Init(xent.diag_confusion_.Dim());
}
confusion_.Add(xent.confusion_);
confusion_count_.Add(xent.confusion_count_);
diag_confusion_.Add(xent.diag_confusion_);
}
}
} // namespace TNet
|
