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
|
#include "cuSparseLinearity.h"
#include <cmath>
#include <cstdlib>
namespace TNet
{
void
CuSparseLinearity::
PropagateFnc(const CuMatrix<BaseFloat>& X, CuMatrix<BaseFloat>& Y)
{
Y.AddScaledRow(1.0,mBias,0.0);
Y.Gemm('N','N', 1.0, X, mLinearity, 1.0);
}
void
CuSparseLinearity::
BackpropagateFnc(const CuMatrix<BaseFloat>& X, CuMatrix<BaseFloat>& Y)
{
Y.Gemm('N', 'T', 1.0, X, mLinearity, 0.0);
}
void
CuSparseLinearity::
Update()
{
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;
mLinearityCorrection.Gemm('T','N',1.0,GetInput(),GetErrorInput(),mMomentum);
mBiasCorrection.AddColSum(1.0,GetErrorInput(),mMomentum);
mLinearity.AddScaled(-mLearningRate/N,mLinearityCorrection,1.0);
mBias.AddScaled(-mLearningRate/N,mBiasCorrection,1.0);
mLinearityCorrectionAccu.AddScaled(1.0,mLinearityCorrection,1.0);
mLinearity.ApplyMask(mSparsityMask);
//L1 regularization lasso...
//each update? everty 1000th update?
if(mL1Const > 0) {
BaseFloat L1_const = mLearningRate*mL1Const*(mGradDivFrm?1.0:GetInput().Rows());
mLinearity.ApplyL1(L1_const);
}
//L2 regularization weight decay (from actual weights only)
if(mWeightcost > 0) {
BaseFloat L2_decay = -mLearningRate*mWeightcost*(mGradDivFrm?1.0:GetInput().Rows());
mLinearity.AddScaled(L2_decay, mLinearity,1.0);
}
mNFrames += GetInput().Rows();
}
void
CuSparseLinearity::
UpdateMask()
{
//move data to host
Matrix<BaseFloat> linearity, linearity_correction_accu;
Matrix<BaseFloat> sparsity_mask;
mLinearity.CopyTo(linearity);
mLinearityCorrectionAccu.CopyTo(linearity_correction_accu);
mSparsityMask.CopyTo(sparsity_mask);
//decide on new sparsity mask
for(size_t r=0; r<sparsity_mask.Rows(); r++) {
for(size_t c=0; c<sparsity_mask.Cols(); c++) {
if(sparsity_mask(r,c) == 1.0f) { //weight active
if(fabs(linearity(r,c)) < mSparsifyWeightThreshold) {
sparsity_mask(r,c) = 0;//deactivate
linearity(r,c) = 0;
}
} else { //weight inactive
if(abs(linearity_correction_accu(r,c))/(BaseFloat)mNFrames > mUnsparsifyAccu) {
sparsity_mask(r,c) = 1;//activate
}
}
}
}
//move data to the device
mLinearity.CopyFrom(linearity);
mSparsityMask.CopyFrom(sparsity_mask);
}
void
CuSparseLinearity::
ReadFromStream(std::istream& rIn)
{
//matrix is stored transposed as SNet does
BfMatrix transpose;
rIn >> transpose;
mLinearity.CopyFrom(BfMatrix(transpose, TRANS));
//biases stored normally
BfVector bias;
rIn >> bias;
mBias.CopyFrom(bias);
//sparsity mask
rIn >> std::ws;
Matrix<BaseFloat> mask_transp;
if(rIn.peek() == 'm') {//load from file
rIn >> mask_transp;
} else {//or set all elements active
mask_transp.Init(transpose.Rows(),transpose.Cols());
int items=transpose.Rows()*transpose.Stride();
BaseFloat* p = mask_transp.pData();
for(int i=0; i<items; i++) {//set all elements to one
*p++ = 1;
}
}
mSparsityMask.CopyFrom(BfMatrix(mask_transp,TRANS));
//dummy matrix with acumulated gradients
rIn >> std::ws;
if(rIn.peek() == 'm') {//load from file
BfMatrix dummy;
rIn >> dummy;
}
if(transpose.Cols()*transpose.Rows() == 0) {
Error("Missing linearity matrix in network file");
}
if(bias.Dim() == 0) {
Error("Missing bias vector in network file");
}
if(mLinearity.Cols() != GetNOutputs() ||
mLinearity.Rows() != GetNInputs() ||
mBias.Dim() != GetNOutputs()
){
std::ostringstream os;
os << "Wrong dimensionalities of matrix/vector in network file\n"
<< "Inputs:" << GetNInputs()
<< "Outputs:" << GetNOutputs()
<< "\n"
<< "linearityCols:" << mLinearity.Cols()
<< "linearityRows:" << mLinearity.Rows()
<< "biasDims:" << mBias.Dim()
<< "\n";
Error(os.str());
}
assert(mLinearity.Rows() == mSparsityMask.Rows());
assert(mLinearity.Cols() == mSparsityMask.Cols());
}
void
CuSparseLinearity::
WriteToStream(std::ostream& rOut)
{
UpdateMask();
//matrix is stored transposed as SNet does
BfMatrix tmp;
mLinearity.CopyTo(tmp);
BfMatrix transpose(tmp, TRANS);
rOut << transpose;
//biases stored normally
BfVector vec;
mBias.CopyTo(vec);
rOut << vec;
rOut << std::endl;
//store mask
mSparsityMask.CopyTo(tmp);
rOut << BfMatrix(tmp,TRANS);
//store accu
mLinearityCorrectionAccu.CopyTo(tmp);
rOut << BfMatrix(tmp,TRANS);
}
} //namespace
|
