-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbackprop_test.py
234 lines (189 loc) · 7.85 KB
/
backprop_test.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
from numpy import *
import matplotlib.pyplot as plt
from random import shuffle
NEURONS = 20
LAMBD = 0.2
def evaluate(inpt, Theta_1, Theta_2):
#Theta_1 = self.get_theta1()
#Theta_2 = self.get_theta2()
# print ("input",inpt)
z_1 = a_1 = append(ones((shape(inpt)[0],1)), inpt, axis=1)
# print ("a_1", shape(a_1))
# print ("T1", shape(Theta_1))
z_2 = dot(a_1,Theta_1)
# print ("z_2", shape(z_2))
a_2 = relu(z_2) #relu
# print ("a_2", shape(a_2))
a_2 = append(ones((shape(a_2)[0],1)), a_2, axis=1)
# print ("a_2", shape(a_2))
# print ("T_2", shape(Theta_2))
z_3 = dot(a_2, Theta_2)
# print ("z_3", shape(z_3))
out = sigmoid(z_3) #softmax
# print ("out", shape(out))
return out, z_1, z_2, z_3, a_1, a_2
def cost(inpt, Y, Theta_1, Theta_2):
h, z_1, z_2,z_3, a_1, a_2 = evaluate(inpt, Theta_1, Theta_2)
m = shape(h)[0]
sum_theta = 0
sum_theta += sum(sum(dot(Theta_1[1:,:].T,Theta_1[1:,:]))) #droppping bias
sum_theta += sum(sum(dot(Theta_2[1:,:].T,Theta_2[1:,:])))
# print (shape(h))
# print (Y)
# a = negative(Y) * log(h) - (ones(shape(Y)) - Y) * log(ones(shape(h)) - h);
# J = 1/m * sum(sum(a)) + lambda/(2*m)*(sum(sum(Theta1(:,2:end) .^2))+sum(sum(Theta2(:,2:end) .^2)));
return sum(sum(power((h-Y),2)))/(2*m) + LAMBD/(2*m)*sum_theta
# return 1./m *sum(sum(a))
def gradinet_check(X_train, Theta_1, Theta_2):
eps = 0.0001
grad_1 = zeros(shape(Theta_1))
for i in range(shape(Theta_1)[0]):
for j in range(shape(Theta_1)[1]):
temp_theta_1 = copy(Theta_1)
temp_theta_1[i][j] += eps
# car.set_theta1(temp_theta_1)
J_plus_eps = cost(X_train, Y_train, temp_theta_1, Theta_2)
temp_theta_1[i][j] -= 2*eps
# car.set_theta1(temp_theta_1)
# evaluation,z_1,z_2 = evaluate(X_train, temp_theta_1, Theta_2)
J_minus_eps = cost(X_train, Y_train, temp_theta_1, Theta_2)
grad_1[i][j] = (J_plus_eps - J_minus_eps)/(2*eps)
print ("jplus", J_plus_eps)
print ("jminus", J_minus_eps)
print ((J_plus_eps - J_minus_eps)/(2*eps))
print ("gradient_1", grad_1)
# car.set_theta1(Theta_1)
grad_2 = zeros(shape(Theta_2))
for i in range(shape(Theta_2)[0]):
for j in range(shape(Theta_2)[1]):
temp_theta_2 = copy(Theta_2)
temp_theta_2[i][j] += eps
# car.set_theta2(temp_theta_2)
# evaluation,z_1,z_2 = car.evaluate(X_train)
J_plus_eps = cost(X_train, Y_train, Theta_1, temp_theta_2)
temp_theta_2[i][j] -= 2*eps
# car.set_theta2(temp_theta_2)
# evaluation,z_1,z_2 = car.evaluate(X_train)
J_minus_eps = cost(X_train, Y_train, Theta_1, temp_theta_2)
grad_2[i][j] = (J_plus_eps - J_minus_eps)/(2*eps)
print ("jplus", J_plus_eps)
print ("jminus", J_minus_eps)
print ((J_plus_eps - J_minus_eps)/(2*eps))
# car.set_theta2(Theta_2)
return grad_1, grad_2
def get_data():
Theta_1 = (random.rand(14*NEURONS) - 0.5).reshape(14, NEURONS) * 1
Theta_2 = (random.rand((NEURONS+1)*2) - 0.5).reshape(NEURONS+1, 2) * 1
data = load("training_data_c" + '.npy', encoding="latin1")
print (shape(data))
# random.shuffle(data)
print (shape(data))
X = data[0][0]
X = list(X)
# random.shuffle(X)
Y = data[1][0]
Y = list(Y)
random.shuffle(Y)
# Y[:][Y<0.5] = 0.5
train_test_split = int(0.7*shape(X)[0])
X_train = X[:train_test_split]
# X_train = ones(shape(X_train))
X_test= X[train_test_split:]
Y_train = Y[:train_test_split]
Y_test= Y[train_test_split:]
evaluation,z_1, z_2, z_3, a_1, a_2 = evaluate(X_train, Theta_1, Theta_2)
return evaluation, Theta_1, Theta_2, X_train, X_test, Y_train, Y_test
def sigmoid(a):
return 2./(1+exp(-a))-1 #[-1;1]
def relu(a):
a[a<0] = 0
return a
def sigmoid_gradient(a):
return 2./(1+exp(-a))*(1-1./(1+exp(-a)))
def relu_gradient(a):
a[a > 0] = 1
a[a <= 0] = 0
return a
def gradiend_descent(X_train, Y_train, Theta_1, Theta_2, alpha, epochs = 100, test = False, X_test = NaN, Y_test = NaN):
m = shape(X_train)[0]
J = []
J_test = []
for i in range(epochs):
D1, D2 = train_brain(X_train, Y_train, Theta_1, Theta_2)
Theta_1 = Theta_1 - alpha*D1 - LAMBD/(2*m)
Theta_2 = Theta_2 - alpha*D2 - LAMBD/(2*m)
J.append(cost(X_train, Y_train, Theta_1, Theta_2))
if test == True:
J_test.append(cost(X_test, Y_test, Theta_1, Theta_2))
print (Theta_1)
if test == True:
return J, J_test
return J
def get_gradient(X_train, Y_train, Theta_1, Theta_2):
evaluation,z_1,z_2,z_3, a_1, a_2 = evaluate(X_train, Theta_1, Theta_2)
m = shape(evaluation)[0]
delta_3 = (evaluation - Y_train)*sigmoid_gradient(z_3)
# print ("delta 3", shape(delta_3))
# print ("Thet 2.T", shape(Theta_2.T))
# print ("D2 shape", shape(dot(a_2.T,delta_3)))
# # D2 = 1./m*(dot(a_2.T,delta_3) + lambd*Theta_2)
# print ("shape h", shape(sigmoid_gradient(z_3)))
# print ("shape a_2", shape(a_2))
D2 = dot(a_2.T, delta_3)/m
# Theta_2 = Theta_2[1:,:]
delta_2 = dot(delta_3, Theta_2.T) * relu_gradient(append(ones((shape(z_2)[0],1)), z_2, axis=1))
# print ("Thet 2.T", shape(Theta_2.T))
# print ("Z_2", shape(z_2))
# print ("dot shape", shape(dot(delta_3, Theta_2.T)))
# print ("delta 2", shape(delta_2))
#
# Theta_2[0:1,:]=0 #not normiizing bias
# delta_2 = delta_2[:,1:] #dropping bias
# print ("delta_2", shape(delta_2))
# print ("theta 1", shape(Theta_1))
# print ("z_1", shape(z_1))
delta_2 = delta_2[:,1:]
D1 = (dot(z_1.T,delta_2))/m
# Theta_1 = Theta_1[1:,:]
# delta_1 = dot(delta_2, Theta_1.T) * sigmoid_gradient(z_1)
# print ("delta_1", shape(delta_1))
# Theta_1[0:1,:]=0
return D1, D2
def train_brain(X_train, Y_train, Theta_1, Theta_2, alpha = 1, epochs = 200, X_test = NaN, Y_test = NaN):
m = shape(evaluation)[0]
J = []
J_test = []
for i in range(epochs):
D1, D2 = get_gradient(X_train, Y_train, Theta_1, Theta_2)
Theta_1 = Theta_1 - alpha*D1 - LAMBD/(2*m)
Theta_2 = Theta_2 - alpha*D2 - LAMBD/(2*m)
J.append(cost(X_train, Y_train, Theta_1, Theta_2))
J_test.append(cost(X_test, Y_test, Theta_1, Theta_2))
print (Theta_1)
return J, J_test
return J
evaluation, Theta_1, Theta_2, X_train, X_test, Y_train, Y_test = get_data()
# G1, G2 = gradinet_check(X_train, Theta_1, Theta_2)
# D1, D2 = train_brain(X_train, Y_train, Theta_1, Theta_2)
alphas = [0.06*power(i,2) for i in range(20)]
# J = array([gradiend_descent(X_train, Y_train, Theta_1, Theta_2, alpha, 200) for alpha in alphas])
# best_alpha = argmin(J[:,190])
print ()
# J = array([gradiend_descent(X_train, Y_train, Theta_1, Theta_2, alpha, 200) for alpha in alphas])
Res = []
NEURONS = 25
Theta_1 = (random.rand(14*NEURONS) - 0.5).reshape(14, NEURONS) * 1
Theta_2 = (random.rand((NEURONS+1)*2) - 0.5).reshape(NEURONS+1, 2) * 1
# J, J_test = gradiend_descent(X_train, Y_train, Theta_1, Theta_2, 0.4, 200, True, X_test, Y_test)
# Res.append(J[len(J_test)-1]/J_test[len(J_test)-1])
# print ("test/train error", J[len(J_test)-1]/J_test[len(J_test)-1])
# print (J_test)
J, J_test = train_brain(X_train, Y_train, Theta_1, Theta_2, 1, 100, X_test, Y_test)
plt.plot(J_test)
plt.plot(J)
# plt.plot(J_test)
plt.ylabel('some numbers')
plt.show()
# print ("test/train error", J[len(J_test)-1]/J_test[len(J_test)-1])
# print ("theta_1 grad", D1, G1)
# print ("theta_2 grad", D2, G2)