-
Notifications
You must be signed in to change notification settings - Fork 3
/
convnet.py
261 lines (204 loc) · 10.2 KB
/
convnet.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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
import numpy as np
from datetime import datetime
from skimage.util.shape import view_as_windows
from numpy.random import random_sample
from scipy.ndimage.filters import gaussian_filter
import sys #for stdio flush
import lcdnorm4
import lpool4
class ConvNet():
print 'Divnorm apenas na primeira camada'
DEBUG = 0
n_filters = None
shape_norm = None
shape_conv = None
shape_pool = None
stride_pool = None
filters_norm = None
filters_conv = None
filters_pooling = None
#fbstochastic = None
n_layers = None
stoc_pool = False
div_norm = None
training = True
retif = True
def extract(self, sample):
#check if the filter bank has the right number of filters
if self.filters_conv== None:
self.set_filters()
for i in range(len(self.n_filters)):
if self.filters_conv[i].shape[-1] != self.n_filters[i]:
self.set_filters() #if it does not have the right number of filters, set the new filters
break
sample = sample.astype(np.float32)
sample.shape = sample.shape + (1,)
for n_layer in np.arange(self.n_layers):
if self.DEBUG>=2:
print 'layer = ', n_layer
antes1 = datetime.now()
sample = self.convolution(sample,n_layer)
if self.DEBUG>=2:
print 'convolution=', (datetime.now() - antes1)
antes1 = datetime.now()
if self.div_norm and n_layer==0:
#sample = self.subtractive_normalization(sample,self.filters_norm[n_layer])
#sample = self.divisive_normalization(sample,self.filters_norm[n_layer])
sample.shape = (1,) + sample.shape
sample = lcdnorm4.lcdnorm4(sample, self.shape_norm[n_layer], contrast=False, divisive=True)
sample.shape = sample.shape[1:]
if self.DEBUG>=2:
print 'divisive_normalization=', (datetime.now() - antes1)
antes1 = datetime.now()
if self.stoc_pool:
sample = self.stochastic_pooling(sample,n_layer)
else:
#sample.shape = (1,)+ sample.shape#reshape to four dimensions
#sample = lpool4.lpool4(sample, neighborhood = self.shape_pool[n_layer], stride=self.stride_pool[n_layer])
#sample.shape = sample.shape[1:]#reshape to three dimensions
sample = self.max_pooling(sample,n_layer)
if self.DEBUG>=2:
print 'max_pooling=', (datetime.now() - antes1)
if self.DEBUG >=2:
print 'final sample.shape=', sample.shape
sys.stdout.flush()#force print when running child/sub processes
#return sample.reshape(-1)
return sample
def similarity3D(self, X, fb):
assert X.ndim == 3
assert fb.ndim == 4
assert X.shape[-1] == fb.shape[2]
Xw = view_as_windows(X, fb.shape[:3])
Xwa = np.abs(Xw-fb)
return Xwa.sum(axis=(3,4,5))
def conv3D(self, X, fb):
assert X.ndim == 3
assert fb.ndim == 4
assert X.shape[-1] == fb.shape[2]
n_filters = fb.shape[-1]
fb2 = fb.copy()
fb2.shape = -1, n_filters
X_rfi = view_as_windows(X, fb.shape[:3])
outh, outw = X_rfi.shape[:2]
X_rfim = X_rfi.reshape(outh * outw, -1)
ret = np.dot(X_rfim, fb2) # -- convolution as matrix multiplication
return ret.reshape(outh, outw, n_filters)
def convolution(self, X, n_layer):
sample = self.conv3D(X, self.filters_conv[n_layer])
#sample = self.similarity3D(X, self.filters_conv[n_layer])
# -- post nonlinearities/retified linear
if self.retif:
np.maximum(sample,0.,sample) #in place operations are faster because it avoids a copy. That is: a +=b is faster than a = a + b
return sample
else:
return sample
def max_pooling(self, X, n_layer):
X_rfi = view_as_windows(X, self.filters_pooling[n_layer].shape + (1,))
X_rfi = X_rfi[::self.stride_pool[n_layer][0], ::self.stride_pool[n_layer][1],:,:,:,:]
return np.amax(X_rfi, axis=(3,4,5))
def subtractive_normalization(self, X, filter_norm ):
# -- pre nonlinearities
rf_shape_side = (np.asarray(filter_norm.shape) - 1)/2
if len(X.shape)==2:
X.shape = X.shape + (1,)
inh, inw, ind = X.shape
ret = self.conv3D(X, filter_norm)
return X[rf_shape_side[0]:inh - rf_shape_side[0], rf_shape_side[1]:inw - rf_shape_side[1], :] - ret
def divisive_normalization(self, X, filter_norm):
# -- pre nonlinearities
rf_shape_side = (np.asarray(filter_norm.shape) - 1)/2
if len(X.shape)==2:
X.shape = X.shape + (1,)
inh, inw, _ = X.shape
ret = X ** 2
#ret = self.conv3D(ret, filter_norm)
X_rfi = view_as_windows(X, filter_norm.shape[:3])
ret = X_rfi.sum(axis=(2,3,4))
# -- post nonlinearities
ret = np.sqrt(ret)
np.maximum(ret,1.,ret) #avoids that very small numbers will cause the nominator have an greater value
return X[rf_shape_side[0]:inh - rf_shape_side[0], rf_shape_side[1]:inw - rf_shape_side[1], :] / ret
def stochastic_pooling(self, X, n_layer):
inh = X.shape[0] - self.shape_pool[0]+1
inw = X.shape[1] - self.shape_pool[1]+1
n_filter = self.n_filters[n_layer]
filtersize = self.shape_pool[0]*self.shape_pool[1]
randomsamples = random_sample((inh)*(inw)*n_filter).reshape((inh),(inw),n_filter) #generate random values
randomsamples = np.repeat(randomsamples,repeats = filtersize, axis=2).reshape((inh),(inw),n_filter,filtersize)
X_rfi = view_as_windows(X, self.shape_pool + (1,))
sumpool = np.repeat(np.sum(X_rfi,axis=(3,4,5)),repeats=filtersize).reshape((inh, inw, n_filter,self.shape_pool[0],self.shape_pool[1],1))
probabilities = X_rfi/sumpool
probabilities[np.isnan(probabilities)] = 1/float(filtersize)#get where the sum is zero and replace by one, so the division by zero error do not occur
probabilities = probabilities.reshape((inh, inw, n_filter,filtersize))
if self.training:
bins = np.add.accumulate(probabilities,axis=3)
binsbefore = np.concatenate((np.zeros((inh, inw, n_filter,1)),bins[:,:,:,:-1]),axis=3)
ret = X_rfi[np.where((((binsbefore<= randomsamples) * (bins> randomsamples)))==True)]
ret = ret.reshape(((inh),(inw),n_filter))[::self.stride_pool,::self.stride_pool]
else: #for testing
ret = probabilities*X_rfi
sumpool[sumpool==0.] = 1.
ret = np.sum(ret,axis=(3))/sumpool[:,:,:,0,0,0]
ret = ret[::self.stride_pool, ::self.stride_pool]
def apply_ZCA(self, Xin):
X = Xin.copy()
#ensure data has zero mean
m = np.mean(X,axis=1).reshape(-1,1)
X = X-m
XXt = np.dot(X,np.transpose(X))
P, D, Pt = np.linalg.svd(XXt)
D = D**(-1/2)
D = np.identity(D.shape[0])*D
W = np.dot(np.dot(P,D),Pt)
return np.dot(W,X)
"""
def get_stochastic_filters(self, n_filters):
filtersize = self.shape_pool[0]*self.shape_pool[1]
filters =[]
for n_filter in n_filters:
#generate random samples for stochastic pooling
randomsamples = random_sample((inh-12)*(inw-12)*n_filter).reshape((inh-12),(inw-12),n_filter) #generate the random values
randomsamples = np.repeat(randomsamples,repeats = filtersize, axis=2).reshape((inh-12),(inw-12),n_filter,filtersize)
filters.append(randomsamples)
return filters
"""
def get_random_filtersconv(self, n_filters):
n_filter_before =1
filters =[]
for n_layer in range(len(n_filters)):
#create filters for convolution
f_shape = self.shape_conv[n_layer] + (n_filter_before,)
f_size = f_shape[0] * f_shape[1] * f_shape[2]
# -- set random filter weights
fbconv = np.random.RandomState(42).randn(f_size, self.n_filters[n_layer]).astype(np.float32)
# -- subtract mean
f_mean = fbconv.mean(axis=0)
fbconv -= f_mean
# -- set to unit-norm
f_norm = np.sqrt((fbconv ** 2).sum(axis=0))
fbconv /= f_norm
fbconv.shape = f_shape + (self.n_filters[n_layer],)
filters.append(fbconv)
n_filter_before = self.n_filters[n_layer]
return filters
def set_filters(self):
#create filters
self.filters_conv = self.get_random_filtersconv(self.n_filters)
#self.fbstochastic = self.get_stochastic_filters(self.n_filters)
self.filters_norm = []
self.filters_pooling = []
aux = np.zeros(self.shape_norm[0])
aux[(aux.shape[0]-1)/2,(aux.shape[1]-1)/2] = 1
gf = gaussian_filter(aux, sigma = 1).astype(np.float32)
#gf.shape = gf.shape+(1,1)
#self.filters_norm.append(gf)
for n_layer in range(len(self.n_filters)):
aux = np.zeros(self.shape_norm[n_layer])
aux[(aux.shape[0]-1)/2,(aux.shape[1]-1)/2] = 1
gf = gaussian_filter(aux, sigma = 1).astype(np.float32)
gf.shape = gf.shape+(1,)
gf = np.repeat(gf,self.n_filters[n_layer],axis=2)
#gf = np.repeat(gf,self.n_filters[n_layer],axis=3)
self.filters_norm.append(gf)
self.filters_pooling.append(np.ones(self.shape_pool[n_layer], dtype=np.float32))
self.n_layers = len(self.n_filters)