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
|
#include "cuDiscreteLinearity.h"
#include "cumath.h"
namespace TNet
{
void
CuDiscreteLinearity::
PropagateFnc(const CuMatrix<BaseFloat>& X, CuMatrix<BaseFloat>& Y)
{
//Y.SetConst(0.0);
//precopy bias
Y.AddScaledRow(1.0,mBias,0.0);
//mulitply with the matrices
int offset_in=0, offset_out=0;
for (int i=0; i<mNBlocks; i++) {
CuMath<BaseFloat>::OffsetGemm('N','N', 1.0, X, mLinearity[i], 1.0, Y,
offset_in, 0, offset_out);
offset_in += mLinearity[i].Rows();
offset_out += mLinearity[i].Cols();
}
}
void
CuDiscreteLinearity::
BackpropagateFnc(const CuMatrix<BaseFloat>& X, CuMatrix<BaseFloat>& Y)
{
//Y.SetConst(0.0);
int offset_in=0, offset_out=0;
for(int i=0; i<mNBlocks; i++) {
CuMath<BaseFloat>::OffsetGemm('N', 'T', 1.0, X, mLinearity[i], 0.0, Y,
offset_in, 0, offset_out);
offset_in += mLinearity[i].Cols();
offset_out += mLinearity[i].Rows();
}
}
void
CuDiscreteLinearity::
Update()
{
//new implementation
BaseFloat N = 1;
if(mGradDivFrm) {
N = static_cast<BaseFloat>(GetInput().Rows());
}
BaseFloat mmt_gain = static_cast<BaseFloat>(1.0/(1.0-mMomentum));
N *= mmt_gain; //compensate higher gradient estimates due to momentum
//get gradients of discrete linearities
int offset_in=0, offset_out=0;
for(int i=0; i<mNBlocks; i++) {
CuMath<BaseFloat>::OffsetGemm('T','N',1.0,
GetInput(),GetErrorInput(),
mMomentum, mLinearityCorrection[i],
offset_in,offset_out,0);
offset_in += mLinearity[i].Rows();
offset_out += mLinearity[i].Cols();
}
for(int i=0; i<mNBlocks; i++) {
//perform update
mLinearity[i].AddScaled(-mLearningRate/N,mLinearityCorrection[i],1.0);
//regularization weight decay
mLinearity[i].AddScaled(-mLearningRate*mWeightcost,mLinearity[i],1.0);
}
//get gradient of bias
mBiasCorrection.AddColSum(1.0,GetErrorInput(),mMomentum);
//update biases
mBias.AddScaled(-mLearningRate/N,mBiasCorrection,1.0);
}
void
CuDiscreteLinearity::
ReadFromStream(std::istream& rIn)
{
rIn >> std::ws >> mNBlocks;
if(mNBlocks < 1) {
KALDI_ERR << "Bad number of blocks:" << mNBlocks;
}
mLinearity.resize(mNBlocks);
mLinearityCorrection.resize(mNBlocks);
int in_dim = 0, out_dim = 0;
for(int i=0; i<mNBlocks; i++) {
//matrix is stored transposed as SNet does
BfMatrix transpose;
rIn >> transpose;
mLinearity[i].CopyFrom(BfMatrix(transpose, TRANS));
if(transpose.Cols()*transpose.Rows() == 0) {
Error("Missing linearity matrix in network file");
}
//allocate training buffers
mLinearityCorrection[i].Init(mLinearity[i].Rows(),mLinearity[i].Cols());
mLinearityCorrection[i].SetConst(0.0);
in_dim += transpose.Cols();
out_dim += transpose.Rows();
}
//biases stored normally
BfVector bias;
rIn >> bias;
mBias.CopyFrom(bias);
if(bias.Dim() == 0) {
Error("Missing bias vector in network file");
}
mBiasCorrection.Init(mBias.Dim());
mBiasCorrection.SetConst(0.0);
if(out_dim != GetNOutputs() ||
in_dim != GetNInputs() ||
mBias.Dim() != GetNOutputs()
){
std::ostringstream os;
os << "Wrong dimensionalities of matrix/vector in network file\n"
<< "Inputs:" << GetNInputs()
<< "Outputs:" << GetNOutputs()
<< "\n"
<< "linearityCols:" << in_dim
<< "linearityRows:" << out_dim
<< "biasDims:" << mBias.Dim()
<< "\n";
Error(os.str());
}
}
void
CuDiscreteLinearity::
WriteToStream(std::ostream& rOut)
{
rOut << mNBlocks << "\n";
for(int i=0; i< mNBlocks; i++) {
//matrix is stored transposed as SNet does
BfMatrix tmp;
mLinearity[i].CopyTo(tmp);
BfMatrix transpose(tmp, TRANS);
rOut << transpose;
}
//biases stored normally
BfVector vec;
mBias.CopyTo(vec);
rOut << vec;
rOut << std::endl;
}
} //namespace
|
