-
Notifications
You must be signed in to change notification settings - Fork 2
/
AutoEncoder.py
223 lines (188 loc) · 8.71 KB
/
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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
__author__ = 'stgy'
import numpy as np
from utils import sigmoid,corrupt,loadData
import matplotlib.pyplot as plt
import matplotlib as mpl
from random import randint
import time
from layers import *
class SparseAutoEncoder:
def __init__(self,input_size,hidden_size,W1=None,W2=None,b1=None,b2=None):
if W1 is None:
low = -4 * np.sqrt(6.0 / (input_size + hidden_size))
high = 4 * np.sqrt(6.0 / (input_size + hidden_size))
self.W1 = np.random.uniform(low,high,size = (input_size,hidden_size))
else:
self.W1 = W1
if W2 is None:
low = -4 * np.sqrt(6.0 / (input_size + hidden_size))
high = 4 * np.sqrt(6.0 / (input_size + hidden_size))
self.W2 = np.random.uniform(low,high,size = (hidden_size,input_size))
else:
self.W2 = W2
if b1 is None:
self.b1 = np.zeros((1,hidden_size))
else:
self.b1 = b1
if b2 is None:
self.b2 = np.zeros((1,input_size))
else:
self.b2 = b2
def get_hidden_output(self,x):
return sigmoid(x.dot(self.W1) + self.b1)
def train(self,x,epochs = 15,lr = 0.1, batch_size = 20, regularization = 1e-3,beta = 1,activation_level = 0.005):
n_batch = x.shape[0] / batch_size
learning_curve_list = []
for i in xrange(epochs):
Loss = []
if i == 0:
start_time = time.clock()
for j in xrange(n_batch):
batch_x = x[j * batch_size: (j + 1) * batch_size]
hidden_in,cache1 = affine_forward(batch_x,self.W1,self.b1)
hidden_out,cache2 = sigmoid_forward(hidden_in)
average_activation = np.sum(hidden_out, axis = 0) / batch_size
reconstruct_in,cache3 = affine_forward(hidden_out,self.W2,self.b2)
batch_loss,dscore = square_loss(reconstruct_in,batch_x)
reg_loss = regularization * 0.5 * (np.sum(self.W1 * self.W1) + np.sum(self.W2 * self.W2))
kl_divergence_activation = beta * np.sum((activation_level * np.log(activation_level / average_activation) +
(1 - activation_level) * np.log((1 - activation_level) / (1 - average_activation))))
loss = batch_loss + reg_loss + kl_divergence_activation
Loss.append(loss)
"""back propagation"""
grad_W2,grad_b2,grad_hidden_out = affine_backward(dscore,cache3)
grad_hidden_out += beta * (
(1 - activation_level) / (1 - average_activation) - activation_level / average_activation) / batch_size
grad_hidden_in = sigmoid_backward(grad_hidden_out,cache2)
grad_W1,grad_b1,_ = affine_backward(grad_hidden_in,cache1)
"""update parameters"""
self.W2 -= lr * (grad_W2 + regularization * self.W2)
self.W1 -= lr * (grad_W1 + regularization * self.W1)
self.b1 -= lr * (grad_b1)
self.b2 -= lr * (grad_b2)
mean_loss = np.mean(Loss)
learning_curve_list.append(mean_loss)
print "average loss is: %f, at epoch %d" % (mean_loss,i)
if i % 10 == 0:
cmap = mpl.cm.gray_r
norm = mpl.colors.Normalize(vmin=0)
rand_index = randint(0,x.shape[0])
plt.subplot(1,2,1)
plt.imshow(x[rand_index].reshape(28,28),cmap = cmap)
plt.subplot(1,2,2)
hidden_random = sigmoid(x[rand_index].dot(self.W1) + self.b1)
recons_random = sigmoid(hidden_random.dot(self.W2) + self.b2)
plt.imshow(recons_random.reshape(28,28),cmap = cmap)
plt.show()
for j in xrange(100):
plt.subplot(10,10,j)
plt.axis('off')
plt.imshow(self.W1.T[j,:].reshape(28,28),cmap = cmap)
plt.show()
if i == 0:
stop_time = time.clock()
print "one single epoch runs %f minutes!" % ((stop_time - start_time) / 60.0)
plt.plot(learning_curve_list)
plt.show()
class DenoisingAutoEncoder:
'''initlize weights and bias'''
def __init__(self,input_size,hidden_size,W1=None,b1=None,b2=None):
if W1 is None:
low = -4 * np.sqrt(6.0 / (input_size + hidden_size))
high = 4 * np.sqrt(6.0 / (input_size + hidden_size))
self.W1 = np.random.uniform(low,high,size = (input_size,hidden_size))
else:
self.W1 = W1
if b1 is None:
self.b1 = np.zeros((1,hidden_size))
else:
self.b1 = b1
if b2 is None:
self.b2 = np.zeros((1,input_size))
else:
self.b2 = b2
self.W2 = self.W1.T # tie weight
def get_hidden_output(self,x):
return sigmoid(x.dot(self.W1) + self.b1)
def train(self,x,epochs = 15,lr = 0.01, batch_size = 20, corruption_level = 0.3, regularization = 0):
n_batch = x.shape[0] / batch_size
corrupt_x = corrupt(x,corruption_level) # add noise to the original data
learning_curve_list = []
for i in xrange(epochs):
Loss = []
if i == 0:
start_time = time.clock()
for j in xrange(n_batch):
batch_x = x[j * batch_size : (j + 1) * batch_size] # get minibatch of original data
corrupt_batch_x = corrupt_x[j * batch_size : (j + 1) * batch_size] # get minibatch of corrupted data
hidden_in,cache1 = affine_forward(corrupt_batch_x,self.W1,self.b1)
hidden_out,cache2 = sigmoid_forward(hidden_in)
reconstruct_in,cache3 = affine_forward(hidden_out,self.W2,self.b2)
batch_loss,dscore = cross_entropy_loss(reconstruct_in,batch_x)
reg_loss = regularization * 0.5 * self.W1 * self.W1
loss = batch_loss + reg_loss
Loss.append(loss)
"""back propagation"""
grad_W2,grad_b2,grad_hidden_out = affine_backward(dscore,cache3)
grad_hidden_in = sigmoid_backward(grad_hidden_out,cache2)
grad_W1,grad_b1,_ = affine_backward(grad_hidden_in,cache1)
"""update parameters"""
self.W1 -= lr * (grad_W1 + grad_W2.T + regularization * self.W1)
self.b1 -= lr * (grad_b1)
self.b2 -= lr * (grad_b2)
mean_loss = np.mean(Loss)
learning_curve_list.append(mean_loss)
print "average loss is: %f, at epoch: %d" % (mean_loss,i)
'''visualize weight'''
if i % 10 == 0:
cmap = mpl.cm.gray_r
norm = mpl.colors.Normalize(vmin=0)
rand_index = randint(0,x.shape[0])
plt.subplot(1,3,1)
plt.imshow(x[rand_index].reshape(28,28),cmap = cmap)
plt.subplot(1,3,2)
plt.imshow(corrupt_x[rand_index].reshape(28,28),cmap = cmap)
hidden_random = sigmoid(corrupt_x[rand_index].dot(self.W1) + self.b1)
recons_random = sigmoid(hidden_random.dot(self.W2) + self.b2)
plt.subplot(1,3,3)
plt.imshow(recons_random.reshape(28,28),cmap = cmap)
plt.show()
for i in xrange(100):
plt.subplot(10,10,i)
plt.axis('off')
plt.imshow(self.W1.T[i,:].reshape(28,28),cmap = cmap)
plt.show()
if i == 0:
stop_time = time.clock()
print "one single epoch runs %i minutes!" % ((stop_time - start_time) / 60.0)
plt.plot(learning_curve_list)
plt.show()
if __name__ == "__main__":
dataset = "mnist.pkl.gz"
train_set, valid_set, test_set = loadData(dataset)
train_x,train_y = train_set
valid_x,valid_y = valid_set
test_x,test_y = test_set
print "the size of training set is:(%d,%d)" % train_x.shape
n_sample,feature_size = train_x.shape
n_hidden = 500
epochs = 100
'''lr = 0.1
batch_size = 20
corruption_level = 0.3
regularization = 0
print "initializing AutoEncoder......"
dA = DenoisingAutoEncoder(feature_size,n_hidden)
print "start training......"
dA.train(train_x,epochs,lr,batch_size,corruption_level,regularization)
print "finish training!"'''
lr = 0.1
batch_size = 20
regularization = 1e-4
activation_level = 0.01
beta = 3
print "initializing AutoEncoder......"
dA = SparseAutoEncoder(feature_size,n_hidden)
print "start training......"
dA.train(train_x,epochs,lr,batch_size,regularization,beta,activation_level)
print "finish training!"