-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathCNM.py
207 lines (148 loc) · 7.25 KB
/
CNM.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
from scipy.optimize import fmin_ncg, fmin_tnc
import scipy
import __main__
import numpy
import pylab
import sys
sys.path.append('/home/antolikjan/topographica/Theano/')
import theano
from theano import tensor as T
from topo.misc.filepath import normalize_path, application_path
from contrib.JanA.ofestimation import *
from contrib.modelfit import *
import contrib.dd
import contrib.JanA.dataimport
from contrib.JanA.regression import laplaceBias
from contrib.JanA.visualization import compareModelPerformanceWithRPI, showRFS, visualize2DOF
class ContrastNormalizationModel(object):
def __init__(self,XX,YY,ZZ,sizex,sizey,of_aff='Exp',of_surr='Linear'):
(self.num_pres,self.kernel_size) = numpy.shape(XX)
self.Y = theano.shared(YY)
self.X = theano.shared(XX)
self.Z = theano.shared(ZZ)
self.xx = theano.shared(numpy.repeat([numpy.arange(0,sizex,1)],sizey,axis=0).T.flatten())
self.yy = theano.shared(numpy.repeat([numpy.arange(0,sizey,1)],sizex,axis=0).flatten())
self.K = T.dvector('K')
self.x = self.K[0]
self.y = self.K[1]
self.surr_size = self.K[2]
self.surr_gain = self.K[3]
self.surr_c50 = self.K[4]
self.n1 = self.K[5]
self.n2 = self.K[6]
self.k = self.K[7:sizex*sizey+7]
self.of_aff = of_aff
self.of_surr = of_surr
#self.a = T.reshape(self.K[5:sizex*sizey],(sizex,sizey))
def model_output(self):
#surr = self.surr_gain *T.var(self.X,axis=1)**2
#surr = self.surr_gain *T.mean(self.X,axis=1)**2
surr = self.surr_gain * T.dot(self.X,T.exp(-T.div_proxy(((self.xx - self.x)**2 + (self.yy - self.y)**2),self.surr_size)).T/T.sqrt(self.surr_size*numpy.pi))
aff = T.dot(self.X,self.k.T)
#lgn_output = theano.printing.Print(message='lgn output:')(lgn_output)
lin = self.construct_of(aff / (self.surr_c50 + self.construct_of(surr,self.of_surr)) - self.n1,self.of_aff)
return lin
def log_likelyhood(self):
mo = self.model_output()
ll = T.sum(mo) - T.sum(T.dot(self.Y.T, T.log(mo)))
ll = ll + T.sum(T.dot(self.k ,T.dot(__main__.__dict__.get('LaplaceBias',0.0004)*self.Z,self.k.T)))
return ll
def func(self):
return theano.function(inputs=[self.K], outputs=self.log_likelyhood(),mode='FAST_RUN')
def der(self):
g_K = T.grad(self.log_likelyhood(), self.K)
return theano.function(inputs=[self.K], outputs=g_K,mode='FAST_RUN')
def hess(self):
g_K = T.grad(self.log_likelyhood(), self.K,consider_constant=[self.Y,self.X])
H, updates = theano.scan(lambda i,v: T.grad(g_K[i],v), sequences= T.arange(g_K.shape[0]), non_sequences=self.K)
f = theano.function(inputs=[self.K], outputs=H,mode='FAST_RUN')
return f
def construct_of(self,inn,of):
if of == 'Exp':
return T.exp(inn)
elif of == 'Sigmoid':
return 1 / (1 + T.exp(-inn))
elif of == 'Square':
return T.sqr(inn)
elif of == 'ExpExp':
return T.exp(T.exp(inn))
elif of == 'Linear':
return inn
elif of == 'ExpSquare':
return T.exp(T.sqr(inn))
elif of == 'LogisticLoss':
return 1*T.log(1+T.exp(1*inn))
elif of == 'Zero':
return inn*0
elif of == 'Linear':
return inn
def response(self,X,kernels):
self.X.value = X
resp = theano.function(inputs=[self.K], outputs=self.model_output())
(a,b) = numpy.shape(kernels)
(c,d) = numpy.shape(X)
responses = numpy.zeros((c,a))
for i in xrange(0,a):
responses[:,i] = resp(kernels[i,:]).T
return responses
def runCNM():
res = contrib.dd.loadResults("newest_dataset.dat")
(sizex,sizey,training_inputs,training_set,validation_inputs,validation_set,ff,db_node) = contrib.JanA.dataimport.sortOutLoading(res)
raw_validation_set = db_node.data["raw_validation_set"]
params={}
params["SCM"]=True
db_node = db_node.get_child(params)
params={}
params["LaplacaBias"] = __main__.__dict__.get('LaplaceBias',0.0004)
params["OFAff"] = __main__.__dict__.get('OFAff','Exp')
params["OFSurr"] = __main__.__dict__.get('OFSurr','Linear')
params["num_neurons"] = __main__.__dict__.get('NumNeurons',103)
significant = [0,1,6,7,8,9,12,13,15,16,17,19,20,21,22,23,24,25,26,27,28,30,31,32,33,34,38,39,40,42,43,45,46,47,48,49,50,51,52,53,56,58,59,61,62,63,64,65,66,69,71,72,75,77,78,79,80,83,85,89,91,92,93,94,96,100,101,102]
# creat history
training_set = numpy.mat(training_set)[:,significant]
validation_set = numpy.mat(validation_set)[:,significant]
training_inputs= numpy.mat(training_inputs)
validation_inputs= numpy.mat(validation_inputs)
for i in xrange(0,len(raw_validation_set)):
raw_validation_set[i] = numpy.mat(raw_validation_set[i])[:,significant]
db_node1 = db_node
db_node = db_node.get_child(params)
num_pres,num_neurons = numpy.shape(training_set)
num_pres,kernel_size = numpy.shape(training_inputs)
num_neurons_to_run=params["num_neurons"]
Ks = numpy.zeros((num_neurons,kernel_size+7))
print 'Kernel size',kernel_size
laplace = laplaceBias(sizex,sizey)
rpi = numpy.linalg.pinv(training_inputs.T*training_inputs + __main__.__dict__.get('RPILaplaceBias',0.0004)*laplace) * training_inputs.T * training_set
bounds = []
for i in xrange(0,kernel_size):
bounds.append((-1000000000,10000000000))
bounds = [(6,26),(6,25),(1,1000000),(0.0,500000),(0.001,100000),(-100,100),(-100,100)] + bounds
for i in xrange(0,num_neurons_to_run):
print i
k0 = [15,15,20,100000,1,0,0] + rpi[:,i].getA1().tolist()
print numpy.shape(k0)
print sizex,sizey
scm = ContrastNormalizationModel(training_inputs,numpy.mat(training_set[:,i]),laplace,sizex,sizey,of_aff=params["OFAff"],of_surr=params["OFSurr"])
#K = fmin_ncg(scm.func(),numpy.array(k0),scm.der(),fhess = scm.hess(),avextol=0.0000001,maxiter=20)
(K,success,c)=fmin_tnc(scm.func(),numpy.array(k0)[:],fprime=scm.der(),bounds = bounds,maxfun = 100000,messages=0)
print scm.func()(K)
Ks[i,:] = K
pred_act = scm.response(training_inputs,Ks)
pred_val_act = scm.response(validation_inputs,Ks)
from contrib.JanA.sparsness_analysis import TrevesRollsSparsness
showRFS(numpy.reshape(numpy.array(rpi.T),(-1,sizex,sizey)))
showRFS(numpy.reshape(Ks[:,7:kernel_size+7],(-1,sizex,sizey)))
print Ks[:,:7]
pylab.figure()
pylab.hist(TrevesRollsSparsness(numpy.mat(pred_val_act)).flatten())
pylab.figure()
pylab.hist(TrevesRollsSparsness(numpy.mat(pred_val_act.T)).flatten())
pylab.figure()
pylab.hist(TrevesRollsSparsness(numpy.mat(validation_set)).flatten())
pylab.figure()
pylab.hist(TrevesRollsSparsness(numpy.mat(validation_set.T)).flatten())
compareModelPerformanceWithRPI(training_set[:,:num_neurons_to_run],validation_set[:,:num_neurons_to_run],training_inputs,validation_inputs,numpy.mat(pred_act)[:,:num_neurons_to_run],numpy.mat(pred_val_act)[:,:num_neurons_to_run],numpy.array(raw_validation_set)[:,:,:num_neurons_to_run],sizex,sizey,'SCM')
db_node.add_data("Kernels",Ks,force=True)
db_node.add_data("GLM",scm,force=True)
#contrib.dd.saveResults(res,"newest_dataset.dat")