/
AutoEncoder.cpp
172 lines (147 loc) · 5.67 KB
/
AutoEncoder.cpp
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
/*******************************************************************
* Copyright(c) 2015
* All rights reserved.
*
* Name: Sparse/Basic Auto Encoder
* Description: A sparse model to find the interesting feature of data
*
* Date: 2015-9-10
* Author: Yang
* Instruction: Use 10000 random 8 * 8 block of ten 512 * 512 image to
* train the model
******************************************************************/
#include "AutoEncoder.h"
AutoEncoder::AutoEncoder(std::string add,
const int rows, const int cols,UINT hiddenUnits,
double sparsity):
IO_dim(cols), hidden_units(hiddenUnits), samples(rows), st_sparsity(sparsity)
{
/*
* Description:
* A constructed function to initialize the Sparse AutoEncoder
*
* @param rows: number of samples
* @param cols: dimension of samples
* @param hiddenUnits: number of hidden units
* @param sparsity: standard sparsity we want to approximate
*
*/
LoadMatrix(*OriginalData, add, cols, rows);
HiddenMatrix = MatrixXf::Zero(samples, hidden_units);
Sparsity_average = VectorXf::Zero(hidden_units);
OutputMatrix = MatrixXf::Zero(samples, IO_dim);
/* Initialize the en/decode weight matrix */
double w_init_interval = sqrt(6./(IO_dim + hidden_units + 1));
Weight_encode = w_init_interval * MatrixXf::Random(hidden_units, IO_dim);
Weight_decode = w_init_interval * MatrixXf::Random(IO_dim, hidden_units);
Bias_encode = VectorXf::Ones(hidden_units);
Bias_decode = VectorXf::Ones(IO_dim);
}
AutoEncoder::AutoEncoder(MatrixXf& matrix,
UINT hiddenUnits, double sparsity):
IO_dim(matrix.cols()), hidden_units(hiddenUnits), samples(matrix.rows()), st_sparsity(sparsity)
{
OriginalData = &matrix;
HiddenMatrix = MatrixXf::Zero(samples, hidden_units);
Sparsity_average = VectorXf::Zero(hidden_units);
OutputMatrix = MatrixXf::Zero(samples, IO_dim);
/* Initialize the en/decode weight matrix */
double w_init_interval = sqrt(6./(IO_dim + hidden_units + 1));
Weight_encode = w_init_interval * MatrixXf::Random(hidden_units, IO_dim);
Weight_decode = w_init_interval * MatrixXf::Random(IO_dim, hidden_units);
Bias_encode = VectorXf::Ones(hidden_units);
Bias_decode = VectorXf::Ones(IO_dim);
}
VectorXf AutoEncoder::Calculate(int index)
{
/*
* Description:
* Calculate the i-th samples by feedforward and store hiddenvector
* in the row i of hidden matrix, output in row i of output matrix
*
* @return outputVector: The output of FF
*/
VectorXf HiddenVector = Weight_encode * OriginalData->row(index).transpose() + Bias_encode;
for (int i = 0; i < HiddenVector.size(); i++)
{
HiddenVector(i) = sigmoid(HiddenVector(i));
}
HiddenMatrix.row(index) = HiddenVector.transpose();
VectorXf output_vector = VectorXf(IO_dim);
output_vector = Weight_decode * HiddenVector + Bias_decode;
for (int i = 0; i < output_vector.size(); i++)
{
output_vector(i) = sigmoid(output_vector(i));
}
OutputMatrix.row(index) = output_vector.transpose();
return output_vector;
}
void AutoEncoder::BackPropagate(int index, double etaLearningRate)
{
/*
* Description:
* BP in the SAE and considering the cross-entropy cost
*
* @param etaLearningRate: Step to convergence
*
*/
VectorXf delta_3th = OriginalData->row(index) - OutputMatrix.row(index) ;
for (int i = 0; i < delta_3th.size(); i++)
{
delta_3th(i) = DSIGMOID(OutputMatrix(index, i)) * delta_3th(i);
}
VectorXf diff_2th = Weight_decode.transpose() * delta_3th;
VectorXf delta_2th = VectorXf(diff_2th.size());
/* If need sparse restriction */
if (hidden_units >= IO_dim)
{
for (int i = 0; i < delta_2th.size(); i++)
{
double d_sp = -st_sparsity / Sparsity_average(i) +
(1 - st_sparsity) / (1 - Sparsity_average(i));
delta_2th(i) = DSIGMOID(HiddenMatrix(index, i)) * (diff_2th(i) + d_sp);
}
}
else
{
for (int i = 0; i < delta_2th.size(); i++)
{
delta_2th(i) = DSIGMOID(HiddenMatrix(index, i)) * diff_2th(i);
}
}
Weight_decode += etaLearningRate * delta_3th * HiddenMatrix.row(index);
Weight_encode += etaLearningRate * delta_2th * OriginalData->row(index);
Bias_decode += etaLearningRate * delta_3th;
Bias_encode += etaLearningRate * delta_2th;
}
void AutoEncoder::Visualizing(int ImageCols, int ImageRows,
int FeatureCols, int FeatureRows)
{
/*
* Description:
* Visualize the feature that hidden units read
*
* @Param ImageCols: The number of features contained in the column
* @Param ImageRows: The number of features contained in the row
* @Param FeatureCols/Rows: The cols/rows of feature image.
*/
cv::Mat image = cv::Mat(ImageRows * FeatureRows, ImageCols * FeatureCols, CV_8UC1);
for (int i = 0; i < hidden_units; i++)
{
VectorXf out = Weight_encode.row(i) / sqrt((Weight_encode.row(i) * (Weight_encode.row(i).transpose())));
for (int j = 0; j < FeatureRows; j++)
{
for (int k = 0; k < FeatureCols; k++)
{
image.at<uchar>(i / ImageCols * FeatureRows + j, i % ImageCols * FeatureCols + k) = uchar(int(out(j * FeatureCols + k) * 256));
}
}
}
cv::imshow("Feature", image);
cv::waitKey();
}
void AutoEncoder::sparsity_averaging()
{
Sparsity_average = VectorXf::Ones(samples).transpose() * HiddenMatrix;
Sparsity_average = Sparsity_average / samples;
}