/
self_rep_neural_net3.py
executable file
·197 lines (165 loc) · 6.2 KB
/
self_rep_neural_net3.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
#!/usr/bin/python
'''
Created on Nov 16, 2013
@author: jason
'''
from pybrain.structure import FeedForwardNetwork, LinearLayer, SigmoidLayer, FullConnection
from pybrain.tools.shortcuts import buildNetwork
from pybrain.optimization import CMAES, NelderMead, ExactNES, FEM, StochasticHillClimber, GA
import csv, time, cPickle, math
import numpy as np
import random
from sklearn.neighbors import NearestNeighbors
class noveltySearch:
def __init__(self):
self.archive = np.zeros((1,6), dtype=np.float)
self.bestFitness = 1e308
self.bestParams = None
self.kdTree = NearestNeighbors(n_neighbors=10)
self.p_min = 0.5
self.numAdd = 0
self.numNotAdd = 0
self.scaleFactor = 1.0
def computeNovelty(self, vector, params):
if self.archive.shape[0] < 10:
n = self.archive.shape[0]
else:
n = 10
self.kdTree.fit(self.archive)
dist, ind = self.kdTree.kneighbors(vector, n, return_distance=True)
p = np.sum(dist)/n
# add to archive if larger than p_min
if p > self.p_min:
self.archive = np.vstack((self.archive, vector))
self.numAdd += 1
self.numNotAdd = 0
else:
self.numAdd = 0
self.numNotAdd +=1
# adjust p_min
if self.numAdd > 1:
self.p_min *= 1.1
if self.numNotAdd > 32 and self.p_min > 0.05:
self.p_min *= 0.9
fitness = vector[0]
if self.bestFitness > fitness:
self.bestFitness = fitness
self.bestParams = params
# print self.archive
print self.numAdd, self.numNotAdd
print self.bestFitness, self.archive.shape[0], self.p_min
return p
class neuralWrapper:
def __init__(self, percentage=0.9, structure=[8,12,12,1]):
self.networkStructure = structure
self.myNetwork = self.generateNetwork(self.networkStructure)
self.myNetwork._setParameters(np.random.uniform(low=-1.0, high=1.0, size=len(self.myNetwork.params)))
self.indices = self.freezedWeightsIndices(self.myNetwork, percentage)
self.originalWeights = np.copy(self.myNetwork.params)
self.noveltySearch = noveltySearch()
self.metaInfo = {"percentage":percentage}
assert 2**self.networkStructure[0] > len(self.myNetwork.params)
def prettyPrintNet(self):
net = self.myNetwork
for mod in net.modules:
print "Module:", mod.name
if mod.paramdim > 0:
print "--parameters:", mod.params
for conn in net.connections[mod]:
print "-connection to", conn.outmod.name
if conn.paramdim > 0:
print "- parameters", conn.params
if hasattr(net, "recurrentConns"):
print "Recurrent connections"
for conn in net.recurrentConns:
print "-", conn.inmod.name, " to", conn.outmod.name
if conn.paramdim > 0:
print "- parameters", conn.params
def generateNetwork(self, structure):
n = FeedForwardNetwork()
prevLayer = SigmoidLayer(structure[0])
n.addInputModule(prevLayer)
for index, num in enumerate(structure[1:-1]):
tempLayer = SigmoidLayer(num)
n.addModule(tempLayer)
n.addConnection(FullConnection(prevLayer, tempLayer))
prevLayer = tempLayer
lastLayer = SigmoidLayer(structure[-1])
n.addOutputModule(lastLayer)
n.addConnection(FullConnection(prevLayer, lastLayer))
n.sortModules()
return n
def freezedWeightsIndices(self, weights, percentage=0.9):
numSamples = int(percentage*len(weights))
return random.sample(xrange(len(weights)), numSamples)
def fitnessFunction(self, weights):
sumoferror = self.sumOfErrors(weights)
entrophy = self.getEntrophy(weights)
median = np.median(weights)
mean = np.mean(weights)
stdev = np.std(weights)
range = np.max(weights) - np.min(weights)
vector = np.array([sumoferror, entrophy, median, mean, stdev, range])
novelty = self.noveltySearch.computeNovelty(vector, self.myNetwork.params)
return novelty
def sumOfErrors(self, weights):
self.myNetwork.params[self.indices] = weights
# self.myNetwork._setParameters(self.originalWeights)
error = 0.0
for index, weight in enumerate(self.myNetwork.params):
myInput = self.position2input(index)
output = self.myNetwork.activate(myInput)[0]
error += abs(weight - output)
return error
def getEntrophy(self, weights):
term = 0.0
for weight in weights:
weight = abs(weight) + 0.00001
term += math.log(weight)*weight
return -1*term
def logNet(self, outname=None):
if outname != None:
f = open(outname, 'wb')
else:
f = open('results_' + str(self.networkStructure) + '_' + str(self.metaInfo["percentage"]) + '.pkl', 'wb')
cPickle.dump((self.myNetwork, self.networkStructure, self.indices, self.originalWeights, self.metaInfo), f)
f.close()
def loadNet(self, filename):
f = open(filename, 'rb')
self.myNetwork, self.networkStructure, self.indices, self.originalWeights, self.metaInfo = cPickle.load(f)
print "num connections: ", self.myNetwork.params.shape[0]
f.close()
def position2input(self, pos):
myInput = [int(x) for x in list('{0:0b}'.format(pos))]
while len(myInput) < self.networkStructure[0]:
myInput.insert(0,0)
return myInput
def experiment1(self):
l = GA(self.fitnessFunction, self.myNetwork.params[self.indices])
l.minimize = False
l.verbose = True
l.maxLearningSteps = 500
params, fitness = l.learn()
self.myNetwork.params[self.indices] = params
self.metaInfo["numsteps"] = l.maxLearningSteps
self.metaInfo["fitness"] = fitness
# self.myNetwork._setParameters(self.originalWeights)
self.logNet()
def compareWeights(self):
error = 0.0
for i,weight in enumerate(self.myNetwork.params):
output = self.myNetwork.activate(self.position2input(i))[0]
if i in self.indices:
type = "unclamped"
else:
type = "clamped"
print type + ", " + str(weight) + ", " + str(output)
error += abs(weight - output)
print "total error:", error
if __name__ == '__main__':
for p in [1.0]:
x = neuralWrapper(percentage=p)
print "percentage:", p
print "num connections:", x.myNetwork.params.shape[0]
x.experiment1()
x.compareWeights()