-
Notifications
You must be signed in to change notification settings - Fork 0
/
NN2.py
226 lines (185 loc) · 6.4 KB
/
NN2.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
from random import random
from numpy import *
from ast import literal_eval
from math import e
from math import sqrt
from scipy import optimize
def sigmoid(x):
return 1/(1+e**-x)
def msigmoid(x):
v = vectorize(lambda z: 1/(1+e**-helper(z)))
return v(x)
def helper(x):
#print str(x)
if x < -700:
return -700
return x
def sigmoidGradient(x):
return msigmoid(x) * (1 - msigmoid(x));
def safe_ln(x, minval=0.0000000001):
return log(x.clip(min=minval))
def generate_weights(layers):
input_layer_size = layers[0]
output_layer_size = layers[len(layers)-1]
hidden_layers = layers[1:len(layers)-1]
thetas = []
prevSize = input_layer_size
for i in range(0,len(hidden_layers)):
eint = sqrt(6)/sqrt(prevSize + hidden_layers[i])
temp = []
for j in range(0,(hidden_layers[i] * (prevSize + 1))):
temp.append(random.random() * 2 * eint - eint)
thetas.append(array(temp))
thetas[i].resize(hidden_layers[i],(prevSize + 1))
prevSize = hidden_layers[i]
eint = sqrt(6)/sqrt(prevSize + output_layer_size)
temp = []
for j in range(0,output_layer_size * (prevSize + 1)):
temp.append(random.random() * 2 * eint - eint)
thetas.append(array(temp))
thetas[len(thetas)-1].resize(output_layer_size,(prevSize + 1))
return thetas
#random epilson Theta numbers working properly
def runNet(nn_params, layers, X):
m = X.shape[0]
X = X.reshape(1,m)
input_layer_size = layers[0]
output_layer_size = layers[len(layers)-1]
hidden_layers = layers[1:len(layers)-2]
#re-roll nn_params into individuals thetas
thetas = []
ii = 0
prevSize = input_layer_size
for i in range(0,len(hidden_layers)):
thetas.append(nn_params[ii:ii + hidden_layers[i] * (prevSize + 1)])
thetas[i].resize(hidden_layers[i], (prevSize + 1))
ii = hidden_layers[i] * (prevSize + 1)
prevSize = hidden_layers[i]
thetas.append(nn_params[ii:ii + output_layer_size * (prevSize + 1)])
thetas[len(thetas)-1].resize(output_layer_size, (prevSize + 1))
#Feed forward
nX = append(ones((1,1)),X,axis=1)
hidden_zs = []
hidden_as = []
prevNodes = nX
for i in range(0,len(thetas)):
hidden_zs.append(dot(prevNodes,thetas[i].transpose()))
n = prevNodes.shape[0]
hidden_as.append(msigmoid(hidden_zs[i]))
if (i != len(thetas)-1):
hidden_as[i] = append(ones((n,1)),hidden_as[i],axis=1)
prevNodes = hidden_as[i]
ol = hidden_as[len(hidden_as)-1]
#print "Running network...\n"+str(ol)
return ol
def nnCostFunction(nn_params, hidden_layers, X, y, reg):
#print "NN_params:\n"+str(nn_params)
m = X.shape[0]
if m != y.shape[0]:
print "Must have the same number of training inputs as outputs"
return 0
input_layer_size = X.shape[1]
output_layer_size = y.shape[1]
#re-roll nn_params into individuals thetas
thetas = []
ii = 0
prevSize = input_layer_size
for i in range(0,len(hidden_layers)):
thetas.append(nn_params[ii:ii + hidden_layers[i] * (prevSize + 1)])
thetas[i].resize(hidden_layers[i], (prevSize + 1))
#thetas[i].resize((prevSize + 1), hidden_layers[i])
#thetas[i] = thetas[i].transpose()
ii = hidden_layers[i] * (prevSize + 1)
prevSize = hidden_layers[i]
thetas.append(nn_params[ii:ii + output_layer_size * (prevSize + 1)])
thetas[len(thetas)-1].resize(output_layer_size,(prevSize + 1))
#thetas[len(thetas)-1].resize((prevSize + 1), output_layer_size)
#thetas[len(thetas)-1] = thetas[len(thetas)-1].transpose()
#Set return values
J = 0
theta_grads = []
for i in range(0,len(thetas)):
theta_grads.append(zeros(thetas[i].shape))
#Feed forward
nX = append(ones((m,1)),X,axis=1)
hidden_zs = []
hidden_as = []
prevNodes = nX
for i in range(0,len(thetas)):
hidden_zs.append(dot(prevNodes,thetas[i].transpose()))
n = prevNodes.shape[0]
hidden_as.append(msigmoid(hidden_zs[i]))
if (i != len(thetas)-1):
hidden_as[i] = append(ones((n,1)),hidden_as[i],axis=1)
prevNodes = hidden_as[i]
ol = hidden_as[len(hidden_as)-1]
J = (1.0/m) * sum(-y * safe_ln(ol) - (1 - y) * safe_ln(1 - ol))
#print "Theta1: \n"+str(thetas[0])
#print "Thetas2: \n"+str(thetas[1])
#print "Output Layer: "+str(ol)
#print "Unregularized J: "+str(J)
#drop below to unregularize
s = 0
for t in thetas:
s = s + sum(t[:,1:] * t[:,1:])
J = J + ((reg*1.0)/(2*m)) * s
#print "Regularized J: "+str(J)
#print str(J)
#cost function J working properly
#Backpropagation
deltas = []
deltas_big = []
deltas.append(ol - y) #deltas for output layer
for i in range(len(thetas)-1,0,-1):
deltas.insert(0,dot(deltas[0], thetas[i][:,1:]))
deltas[0] = deltas[0] * sigmoidGradient(hidden_zs[i-1])
deltas_big.append(dot(deltas[0].transpose(),nX))
for i in range(1,len(thetas)):
deltas_big.append(dot(deltas[i].transpose(), hidden_as[i-1]))
for i in range(0,len(theta_grads)):
#theta_grads[i] = (1.0/m) * deltas_big[i] == unregularized
theta_grads[i] = (1.0/m) * deltas_big[i] + ((reg*1.0)/m) * thetas[i]
theta_grads[i][:,0] = theta_grads[i][:,0] - ((reg*1.0)/m) * thetas[i][:,0]
#print "ThetaGrads:\n"+str(unroll(theta_grads))
return (J, theta_grads)
def unroll(xs):
acc = xs[0].ravel()
for i in range(1,len(xs)):
acc = append(acc, xs[i].ravel())
return acc
def getJ(nnparams, *args):
layers, X, y, reg = args
answer = nnCostFunction(nnparams, layers, X, y, reg)[0]
print "error: "+str(answer)
return answer
def getGrad(nnparams, *args):
layers, X, y, reg = args
answer = nnCostFunction(nnparams, layers, X, y, reg)[1]
return unroll(answer)
def save_theta(filename, theta):
st = "["+str(theta[0])
for i in theta[1:len(theta)]:
st = st + ","+str(i)
st = st+"]"
f = open(filename, 'w')
s = "theta\n"+st+"\n"
f.write(s)
f.close
def load_theta(filename):
f = open(filename, 'r+')
lines = f.readlines()
theta = literal_eval(lines[1])
return theta
"""
#Xor gate testing, 4 reals
l = [2,2,1]
t = unroll(generate_weights(l))
print "Making a [2,2,1] XOR gate"
X = array([[0,0],[1,0],[0,1],[1,1]])
y = array([[0], [1], [1], [0]])
# Xor lambda should be ~ < 0.00001
args = ([2], X, y, 0.00001)
res1 = optimize.fmin_cg(getJ, t, fprime=getGrad, args=args, maxiter=500)#, gtol=0.0000000005)
#print 'res1 = ', res1
print "Output: \n"+str(runNet(res1, [2], X, y))
"""