/
multilayer.py
251 lines (221 loc) · 9.56 KB
/
multilayer.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
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
#!/usr/bin/env python
from __future__ import division
import sys
import math
import random
from scipy.stats import bernoulli
# read data
def readData(fileName):
fr = open(fileName, 'r')
lines = fr.read().strip().splitlines()
instances = []
for line in lines:
instance = line.split(', ')
instances.append([float(x) for x in instance])
fr.close()
return instances
#forward to predict Y and store phi(nets)
def forward(parameters, instance, nodes):
features = instance[1:]
nets = []
nets.append(features) #initial input layer
#initial hidden layer phi(nets) to zero
for i in range(len(nodes)):
numOfCurrentLayerNodes = nodes[i]
nets.append([0 for k in xrange(numOfCurrentLayerNodes)]) #initial hidden layer outputs
#nets.append([0]) #initial the phi(nets) of output Y
Y = [0 for x in xrange(nodes[len(nodes)-1])]
for i in range(len(parameters)):
numOfNodesOfPre = len(parameters[i])
numOfNodesOfCurrent = len(parameters[i][0])
for j in range(numOfNodesOfCurrent):
netJ = 0
for k in range(numOfNodesOfPre):
netJ = netJ + parameters[i][k][j] * nets[i][k]
nets[i + 1][j] = 1.0 / (1 + math.exp(-1 * netJ))
#if i == len(parameters) - 1:
for j in xrange(len(Y)):
if nets[len(nets) - 1][j] < 0.5:
Y[j] = 0
else:
Y[j] = 1
return [Y, nets]
#predict Y
def predict(parameters, instance, nodes):
features = instance[1:]
nets = []
nets.append(features) #initial input layer
#initial hidden layer phi(nets) to zero
for i in range(len(nodes)):
numOfCurrentLayerNodes = nodes[i]
nets.append([0 for k in xrange(numOfCurrentLayerNodes)]) #initial hidden layer outputs
#nets.append([0]) #initial the phi(nets) of output Y
Y = [0 for x in xrange(nodes[len(nodes)-1])]
for i in range(len(parameters)):
numOfNodesOfPre = len(parameters[i])
numOfNodesOfCurrent = len(parameters[i][0])
for j in range(numOfNodesOfCurrent):
netJ = 0
for k in range(numOfNodesOfPre):
netJ = netJ + parameters[i][k][j] * nets[i][k]
nets[i + 1][j] = 1.0 / (1 + math.exp(-1 * netJ))
#if i == len(parameters) - 1:
maxNet = 0
for j in xrange(len(Y)):
if nets[len(nets) - 1][j] > nets[len(nets)-1][maxNet]:
maxNet = j
Y[maxNet] = 1
return Y
#get accuracy
def evaluate(parameters, instances, nodes):
correct = 0
numOfYs = nodes[len(nodes)-1]
for instance in instances:
label = [0 for x in xrange(numOfYs)]
label[int(instance[0])-1] = 1
Y = predict(parameters, instance, nodes)
c = True;
for i in range(len(Y)-1):
if label[i] != Y[i]:
c = False
if c == True:
correct = correct + 1
return (1.0 * correct / len(instances))
def preTrainingParameters(parameters, instances, nodes):
a = 0.1
###convert feature to binary
for i in range(len(instances)):
instance = instances[i]
feature = instance[1:]
instances[i] = [bernoulli.rvs(1.0/(1 + math.exp(-1 * x)), size=1)[0] for x in feature]
currentLayerNodes = instances
for parameter in range(len(parameters)):
currentParameter = parameters[parameter]
for instance in range(len(currentLayerNodes)):
currentLayerNode = currentLayerNodes[instance]
### remove y for input layer
#if (parameter == 0):
# currentLayerNode = currentLayerNode[1:]
###use visible to get hidden
nextLayerNodes = [0 for x in range(nodes[parameter])]
for nextLayerNode in range(len(nextLayerNodes)):
tmp = 0
for w in range(len(currentParameter)):
tmp = tmp + currentLayerNode[w]*currentParameter[w][nextLayerNode]
nextLayerNodes[nextLayerNode] = bernoulli.rvs(1.0/(1 + math.exp(-1 * tmp)), size=1)[0]
###use hidden to get visible, negative phase
visibleNodesNegative = [0 for x in range(len(currentParameter))]
for visibleNodeNegative in range(len(visibleNodesNegative)):
tmp = 0
for w in range(len(nextLayerNodes)):
tmp = tmp + nextLayerNodes[w]*currentParameter[visibleNodeNegative][w]
visibleNodesNegative[visibleNodeNegative] = bernoulli.rvs(1.0/(1 + math.exp(-1 * tmp)), size=1)[0]
###use negative visible to get negative hidden
nextLayerNodesNegative = [0 for x in range(nodes[parameter])]
for nextLayerNodeNegative in range(len(nextLayerNodesNegative)):
tmp = 0
for w in range(len(currentParameter)):
tmp = tmp + visibleNodesNegative[w]*currentParameter[w][nextLayerNode]
nextLayerNodesNegative[nextLayerNodeNegative] = bernoulli.rvs(1.0/(1 + math.exp(-1 * tmp)), size=1)[0]
###compute delta parameters
#deltaParameter = [[0 for x in range(len(currentParameter[0]))] for y in range(len(currentParameter))]
for i in range(len(currentParameter)):
for j in range(len(currentParameter[0])):
currentParameter[i][j] = currentParameter[i][j] + a * (currentLayerNode[i]*nextLayerNodes[j]-visibleNodesNegative[i]*nextLayerNodesNegative[j])
parameters[parameter] = currentParameter
###update currentLayerNodes for pre-training next layer parameters
for instance in range(len(currentLayerNodes)):
currentLayerNode = currentLayerNodes[instance]
for nextLayerNode in range(len(nextLayerNodes)):
tmp = 0
for w in range(len(currentParameter)):
tmp = tmp + currentLayerNode[w]*currentParameter[w][nextLayerNode]
nextLayerNodes[nextLayerNode] = bernoulli.rvs(1.0/(1 + math.exp(-1 * tmp)), size=1)[0]
currentLayerNodes[instance] = nextLayerNodes
#main function
cmdParameters = sys.argv
if (cmdParameters[1] == 'train'):
alpha = 1
instances = readData(cmdParameters[2])
preTrainInstances = readData(cmdParameters[3])
# initial arrays of parameters matrix, set all the parameters to 0
numOfInputs = len(instances[0]) - 1
#numOfNodesOfOutputLayer = int(cmdParameters[3])
nodes = [int(x) for x in cmdParameters[4:]]
parameters = []
parameters.append(
[[random.random() for k in xrange(nodes[0])] for j
in xrange(numOfInputs)])
for i in range(len(nodes) - 1):
numOfCurrentLayerNodes = nodes[i]
numOfNextLayerNodes = nodes[i + 1]
parameters.append([[random.random() for k in
xrange(numOfNextLayerNodes)] for j in xrange(numOfCurrentLayerNodes)])
#parameters.append(
# [[random.random() for k in xrange(1)] for j in
# xrange(nodes[len(nodes) - 1])])
#preTrainingParameters(parameters, preTrainInstances, nodes)
for ite in range(0, 20):
for instance in instances:
numOfYs = nodes[len(nodes)-1]
label = [0 for x in xrange(numOfYs)]
label[int(instance[0])-1] = 1
preSegma = []
#forward to get predicted Y
[Y, nets] = forward(parameters, instance, nodes)
#backward propagation
for i in reversed(range(len(nets))):
if i > 0:
numOfNodesOfCurrent = len(nets[i])
numOfNodesOfLower = len(nets[i - 1])
currentSegma = []
if i == len(nets) - 1:
for j in range(numOfYs):
currentSegma.append((Y[j] - label[j]) * nets[i][j] * (1 - nets[i][j]))
else:
numOfNodesOfUpper = len(nets[i + 1])
for j in range(numOfNodesOfCurrent):
tmp = 0
for k in range(numOfNodesOfUpper):
tmp = tmp + preSegma[k] * parameters[i][j][k]
segma = tmp * nets[i][j] * (1 - nets[i][j])
currentSegma.append(segma)
for j in range(numOfNodesOfLower):
for k in range(numOfNodesOfCurrent):
parameters[i - 1][j][k] = parameters[i - 1][j][k] - alpha * currentSegma[k] * nets[i - 1][j]
preSegma = currentSegma
print 'accuracy = ' + str(evaluate(parameters, instances, nodes))
print '\n'
parametersMatrix = '\n\n'.join('\n'.join(' '.join(str(x) for x in e) for e in matrix) for matrix in parameters)
f = open('parameters', 'w')
f.write(parametersMatrix)
f.close()
if cmdParameters[1] == 'test':
parametersFile = cmdParameters[2]
fr = open(parametersFile, 'r')
lines = fr.read().strip().splitlines()
parameters = []
node = []
words = []
for line in lines:
#print line
#print repr(line)
if line in ['\n', '', ' ']:
parameters.append(node)
node = []
else:
words = line.split(' ')
node.append([float(w) for w in words])
parameters.append(node)
#print parameters
fr.close()
nodes = []
instances = readData(cmdParameters[3])
for i in xrange(len(parameters)):
nodes.append(len(parameters[i]))
print 'accuracy = ' + str(evaluate(parameters, instances, nodes))
#pprint.pprint(parameters)
print '\n'
#pprint.pprint(nets[1][1]*parameters[1][1][0])
#print '\n'
#pprint.pprint(parameters )