-
Notifications
You must be signed in to change notification settings - Fork 0
/
Train.py
196 lines (139 loc) · 5.71 KB
/
Train.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
import sys
import numpy as np
from scipy.optimize import minimize
import SimpleNN2
def trainSciPy(net, x, y, lmb):
net.setRandomWeights()
combinedTheta = net.combineTheta(net.theta)
optimizedTheta = minimize(
fun = lambda p: net.computeCost(p, x, y, lmb),
x0 = combinedTheta,
method = 'BFGS',
jac = lambda p: net.computeGrad(p, x, y, lmb),
#callback = lambda xk: print("Iteration complete!"),
options={'disp': True}) #'maxiter' : 5, 'eps' : 1e-10, 'gtol' : 1e-10
net.theta = net.splitTheta(optimizedTheta.x)
return net
def trainSciPy2(netConfig, x, y, lmb):
th1, th2 = SimpleNN2.initRandomThetas(netConfig)
combinedTheta = SimpleNN2.combineThetas(th1, th2)
optimizedTheta = minimize(
fun = lambda p: SimpleNN2.computeCostComb(netConfig, p, x, y, lmb) ,
x0 = combinedTheta,
method = 'L-BFGS-B',
jac = lambda p: SimpleNN2.computeGradComb(netConfig, p, x, y, lmb),
#callback = lambda xk: print("Iteration complete!"),
options={'disp': False}) #'maxiter' : 5, 'eps' : 1e-10, 'gtol' : 1e-10
return SimpleNN2.splitThetas(netConfig, optimizedTheta.x)
def trainGradientDescent(net, x, y, lmb):
net.setRandomWeights()
alpha = 2
thetaTmp = net.theta
costs = []
while True:
(costBefore, grad) = net.computeCostGrad(thetaTmp, x, y, lmb)
thetaTmp1 = [thetaTmp[0] - alpha*grad[0], thetaTmp[1] - alpha*grad[1]]
(costAfter, _) = net.computeCostGrad(thetaTmp1, x, y, lmb)
skipUpdate = False
if costAfter > costBefore:
alpha = alpha / 1.01
skipUpdate = True
print("Decrease alpha due to cyclic behaviour")
# thetaTmp1 = [thetaTmp[0] + alpha*grad[0], thetaTmp[1] + alpha*grad[1]]
# (costAfter, _) = net.computeCostGrad(thetaTmp1, x, y, lmb)
# if costAfter > costBefore:
# alpha = alpha / 1.5
# skipUpdate = True
# print("Decrease alpha due to cyclic behaviour")
#print("costAfter: {0}".format(costAfter))
if not skipUpdate:
costs.append(costAfter)
thetaTmp = thetaTmp1
if len(costs) > 0 and len(costs) % 10 == 0:
print('Epoch', len(costs), 'with cost', costs[-1], 'and alpha', alpha)
if len(costs) > 2 and abs(costs[-2] - costs[-1]) < 0.00001:
if alpha < 0.02:
break
else:
print("Decrease alpha due to close costs")
alpha = alpha / 1.5
net.theta = thetaTmp
return net
def trainGradientDescent2(netConfig, x, y, lmb):
th1, th2 = SimpleNN2.initRandomThetas(netConfig)
alpha = 2.0
costs = []
while True:
costBefore = SimpleNN2.computeCost(netConfig, th1, th2, x, y, lmb)
grad1, grad2 = SimpleNN2.computeGrad(netConfig, th1, th2, x, y, lmb)
th1p = th1 - alpha*grad1
th2p = th2 - alpha*grad2
costAfter = SimpleNN2.computeCost(netConfig, th1p, th2p, x, y, lmb)
skipUpdate = False
if costAfter > costBefore:
alpha = alpha / 1.01
skipUpdate = True
print("Decrease alpha due to cyclic behaviour")
if not skipUpdate:
costs.append(costAfter)
th1 = th1p
th2 = th2p
if len(costs) > 0 and len(costs) % 10 == 0:
print('Epoch', len(costs), 'with cost', costs[-1], 'and alpha', alpha)
if len(costs) > 2 and abs(costs[-2] - costs[-1]) < 0.00001:
if alpha < 0.02:
break
else:
print("Decrease alpha due to close costs")
alpha = alpha / 1.5
return th1, th2
def findOptimalAlpha(netConfig, theta1, theta2, x, y, lmb, grad1, grad2, alphaFrom, alphaTo):
alphas = np.linspace(alphaFrom, alphaTo, 15)
bestAlpha = 0
bestCost = sys.float_info.max
for a in alphas:
theta1p = theta1 - a*grad1
theta2p = theta2 - a*grad2
cost = SimpleNN2.computeCost(netConfig, theta1p, theta2p, x, y, lmb)
if cost < bestCost:
bestCost = cost
bestAlpha = a
return bestAlpha
def trainSGD(netConfig, x, y, lmb):
th1, th2 = SimpleNN2.initRandomThetas(netConfig)
alpha = 0.1
costs = []
numSamples = x.shape[0]
miniBatchSize = 200
for i in range((numSamples-2)//miniBatchSize):
fr = i*miniBatchSize
to = (i+1)*miniBatchSize
xi = x[fr:to,:]
yi = y[fr:to]
costBefore = 0.0
if len(costs) > 0:
costBefore = costs[-1]
else:
costBefore = SimpleNN2.computeCost(netConfig, th1, th2, xi, yi, lmb)
grad1, grad2 = SimpleNN2.computeGrad(netConfig, th1, th2, xi, yi, lmb)
alpha = findOptimalAlpha(netConfig, th1, th2, xi, yi, lmb, grad1, grad2, alpha/2, alpha*2)
th1p = th1 - alpha*grad1
th2p = th2 - alpha*grad2
costAfter = SimpleNN2.computeCost(netConfig, th1p, th2p, xi, yi, lmb)
if costAfter <= costBefore:
costs.append(costAfter)
th1 = th1p
th2 = th2p
#else:
# # Find optimal alpha in a wide range
# alpha = findOptimalAlpha(netConfig, th1, th2, xi, yi, lmb, grad1, grad2, alpha/50, alpha)
# th1p = th1 - alpha*grad1
# th2p = th2 - alpha*grad2
# costAfter = SimpleNN2.computeCost(netConfig, th1p, th2p, xi, yi, lmb)
# if costAfter <= costBefore:
# costs.append(costAfter)
# th1 = th1p
# th2 = th2p
if len(costs) > 0 and len(costs) % 10 == 0:
print('Epoch', len(costs), 'with cost', costs[-1], 'and alpha', alpha)
return th1, th2