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
|
#ifndef _CUSPARSE_LINEARITY_H_
#define _CUSPARSE_LINEARITY_H_
#include "cuComponent.h"
#include "cumatrix.h"
#include "Matrix.h"
#include "Vector.h"
namespace TNet {
/**
* \brief CuSparseLinearity summation function
*
* \ingroup CuNNUpdatable
* Using weight masks to avoid fluctuation in the output
* -Weights are masked when it is lower than certain threshold - mSparsifyWeightThreshold
* -Weights are activated when the accumulated change is larger than certan value - mUnsparsifyAccu
* -L1 lasso function zeroing weights
* .
* \sa CuBiasedLinearity
*/
class CuSparseLinearity : public CuUpdatableComponent
{
public:
CuSparseLinearity(size_t nInputs, size_t nOutputs, CuComponent *pPred);
~CuSparseLinearity();
ComponentType GetType() const;
const char* GetName() const;
void PropagateFnc(const CuMatrix<BaseFloat>& X, CuMatrix<BaseFloat>& Y);
void BackpropagateFnc(const CuMatrix<BaseFloat>& X, CuMatrix<BaseFloat>& Y);
void Update();
void UpdateMask();
void ReadFromStream(std::istream& rIn);
void WriteToStream(std::ostream& rOut);
void L1(BaseFloat l1) {
mL1Const = l1;
}
protected:
CuMatrix<BaseFloat> mLinearity; ///< Matrix with neuron weights
CuVector<BaseFloat> mBias; ///< Vector with biases
CuMatrix<BaseFloat> mSparsityMask; ///< Mask which selects active weights
CuMatrix<BaseFloat> mLinearityCorrection; ///< Matrix for linearity updates
CuVector<BaseFloat> mBiasCorrection; ///< Vector for bias updates
CuMatrix<BaseFloat> mLinearityCorrectionAccu; ///< Accumulator for linearity updates
BaseFloat mL1Const; ///< L1 regularization constant
size_t mNFrames; ///< Number of accumulated frames
BaseFloat mSparsifyWeightThreshold; ///< Cutoff
BaseFloat mUnsparsifyAccu; ///< Threshold to unsparsify the Cutoff
};
////////////////////////////////////////////////////////////////////////////
// INLINE FUNCTIONS
// CuSparseLinearity::
inline
CuSparseLinearity::
CuSparseLinearity(size_t nInputs, size_t nOutputs, CuComponent *pPred)
: CuUpdatableComponent(nInputs, nOutputs, pPred),
mLinearity(nInputs,nOutputs), mBias(nOutputs), mSparsityMask(nInputs,nOutputs),
mLinearityCorrection(nInputs,nOutputs), mBiasCorrection(nOutputs),
mLinearityCorrectionAccu(nInputs,nOutputs),
mNFrames(0), mSparsifyWeightThreshold(1.0e-3),
mUnsparsifyAccu(1e20f)
{
mLinearityCorrection.SetConst(0.0f);
mBiasCorrection.SetConst(0.0f);
mLinearityCorrectionAccu.SetConst(0.0f);
}
inline
CuSparseLinearity::
~CuSparseLinearity()
{ }
inline CuComponent::ComponentType
CuSparseLinearity::
GetType() const
{
return CuComponent::SPARSE_LINEARITY;
}
inline const char*
CuSparseLinearity::
GetName() const
{
return "<sparselinearity>";
}
} //namespace
#endif
|
