-
Notifications
You must be signed in to change notification settings - Fork 333
/
sparse_autoencoder.py
181 lines (138 loc) · 7.4 KB
/
sparse_autoencoder.py
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
import numpy as np
def sigmoid(x):
return 1 / (1 + np.exp(-x))
def sigmoid_prime(x):
return sigmoid(x) * (1 - sigmoid(x))
def KL_divergence(x, y):
return x * np.log(x / y) + (1 - x) * np.log((1 - x) / (1 - y))
def initialize(hidden_size, visible_size):
# we'll choose weights uniformly from the interval [-r, r]
r = np.sqrt(6) / np.sqrt(hidden_size + visible_size + 1)
W1 = np.random.random((hidden_size, visible_size)) * 2 * r - r
W2 = np.random.random((visible_size, hidden_size)) * 2 * r - r
b1 = np.zeros(hidden_size, dtype=np.float64)
b2 = np.zeros(visible_size, dtype=np.float64)
theta = np.concatenate((W1.reshape(hidden_size * visible_size),
W2.reshape(hidden_size * visible_size),
b1.reshape(hidden_size),
b2.reshape(visible_size)))
return theta
# visible_size: the number of input units (probably 64)
# hidden_size: the number of hidden units (probably 25)
# lambda_: weight decay parameter
# sparsity_param: The desired average activation for the hidden units (denoted in the lecture
# notes by the greek alphabet rho, which looks like a lower-case "p").
# beta: weight of sparsity penalty term
# data: Our 64x10000 matrix containing the training data. So, data(:,i) is the i-th training example.
#
# The input theta is a vector (because minFunc expects the parameters to be a vector).
# We first convert theta to the (W1, W2, b1, b2) matrix/vector format, so that this
# follows the notation convention of the lecture notes.
# Returns: (cost,gradient) tuple
def sparse_autoencoder_cost(theta, visible_size, hidden_size,
lambda_, sparsity_param, beta, data):
# The input theta is a vector (because minFunc expects the parameters to be a vector).
# We first convert theta to the (W1, W2, b1, b2) matrix/vector format, so that this
# follows the notation convention of the lecture notes.
W1 = theta[0:hidden_size * visible_size].reshape(hidden_size, visible_size)
W2 = theta[hidden_size * visible_size:2 * hidden_size * visible_size].reshape(visible_size, hidden_size)
b1 = theta[2 * hidden_size * visible_size:2 * hidden_size * visible_size + hidden_size]
b2 = theta[2 * hidden_size * visible_size + hidden_size:]
# Number of training examples
m = data.shape[1]
# Forward propagation
z2 = W1.dot(data) + np.tile(b1, (m, 1)).transpose()
a2 = sigmoid(z2)
z3 = W2.dot(a2) + np.tile(b2, (m, 1)).transpose()
h = sigmoid(z3)
# Sparsity
rho_hat = np.sum(a2, axis=1) / m
rho = np.tile(sparsity_param, hidden_size)
# Cost function
cost = np.sum((h - data) ** 2) / (2 * m) + \
(lambda_ / 2) * (np.sum(W1 ** 2) + np.sum(W2 ** 2)) + \
beta * np.sum(KL_divergence(rho, rho_hat))
# Backprop
sparsity_delta = np.tile(- rho / rho_hat + (1 - rho) / (1 - rho_hat), (m, 1)).transpose()
delta3 = -(data - h) * sigmoid_prime(z3)
delta2 = (W2.transpose().dot(delta3) + beta * sparsity_delta) * sigmoid_prime(z2)
W1grad = delta2.dot(data.transpose()) / m + lambda_ * W1
W2grad = delta3.dot(a2.transpose()) / m + lambda_ * W2
b1grad = np.sum(delta2, axis=1) / m
b2grad = np.sum(delta3, axis=1) / m
# After computing the cost and gradient, we will convert the gradients back
# to a vector format (suitable for minFunc). Specifically, we will unroll
# your gradient matrices into a vector.
grad = np.concatenate((W1grad.reshape(hidden_size * visible_size),
W2grad.reshape(hidden_size * visible_size),
b1grad.reshape(hidden_size),
b2grad.reshape(visible_size)))
return cost, grad
def sparse_autoencoder(theta, hidden_size, visible_size, data):
"""
:param theta: trained weights from the autoencoder
:param hidden_size: the number of hidden units (probably 25)
:param visible_size: the number of input units (probably 64)
:param data: Our matrix containing the training data as columns. So, data(:,i) is the i-th training example.
"""
# We first convert theta to the (W1, W2, b1, b2) matrix/vector format, so that this
# follows the notation convention of the lecture notes.
W1 = theta[0:hidden_size * visible_size].reshape(hidden_size, visible_size)
b1 = theta[2 * hidden_size * visible_size:2 * hidden_size * visible_size + hidden_size]
# Number of training examples
m = data.shape[1]
# Forward propagation
z2 = W1.dot(data) + np.tile(b1, (m, 1)).transpose()
a2 = sigmoid(z2)
return a2
# visible_size: the number of input units (probably 64)
# hidden_size: the number of hidden units (probably 25)
# lambda_: weight decay parameter
# sparsity_param: The desired average activation for the hidden units (denoted in the lecture
# notes by the greek alphabet rho, which looks like a lower-case "p").
# beta: weight of sparsity penalty term
# data: Our 64x10000 matrix containing the training data. So, data(:,i) is the i-th training example.
#
# The input theta is a vector (because minFunc expects the parameters to be a vector).
# We first convert theta to the (W1, W2, b1, b2) matrix/vector format, so that this
# follows the notation convention of the lecture notes.
# Returns: (cost,gradient) tuple
def sparse_autoencoder_linear_cost(theta, visible_size, hidden_size,
lambda_, sparsity_param, beta, data):
# The input theta is a vector (because minFunc expects the parameters to be a vector).
# We first convert theta to the (W1, W2, b1, b2) matrix/vector format, so that this
# follows the notation convention of the lecture notes.
W1 = theta[0:hidden_size * visible_size].reshape(hidden_size, visible_size)
W2 = theta[hidden_size * visible_size:2 * hidden_size * visible_size].reshape(visible_size, hidden_size)
b1 = theta[2 * hidden_size * visible_size:2 * hidden_size * visible_size + hidden_size]
b2 = theta[2 * hidden_size * visible_size + hidden_size:]
# Number of training examples
m = data.shape[1]
# Forward propagation
z2 = W1.dot(data) + np.tile(b1, (m, 1)).transpose()
a2 = sigmoid(z2)
z3 = W2.dot(a2) + np.tile(b2, (m, 1)).transpose()
h = z3
# Sparsity
rho_hat = np.sum(a2, axis=1) / m
rho = np.tile(sparsity_param, hidden_size)
# Cost function
cost = np.sum((h - data) ** 2) / (2 * m) + \
(lambda_ / 2) * (np.sum(W1 ** 2) + np.sum(W2 ** 2)) + \
beta * np.sum(KL_divergence(rho, rho_hat))
# Backprop
sparsity_delta = np.tile(- rho / rho_hat + (1 - rho) / (1 - rho_hat), (m, 1)).transpose()
delta3 = -(data - h)
delta2 = (W2.transpose().dot(delta3) + beta * sparsity_delta) * sigmoid_prime(z2)
W1grad = delta2.dot(data.transpose()) / m + lambda_ * W1
W2grad = delta3.dot(a2.transpose()) / m + lambda_ * W2
b1grad = np.sum(delta2, axis=1) / m
b2grad = np.sum(delta3, axis=1) / m
# After computing the cost and gradient, we will convert the gradients back
# to a vector format (suitable for minFunc). Specifically, we will unroll
# your gradient matrices into a vector.
grad = np.concatenate((W1grad.reshape(hidden_size * visible_size),
W2grad.reshape(hidden_size * visible_size),
b1grad.reshape(hidden_size),
b2grad.reshape(visible_size)))
return cost, grad