-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathcapsnet_theano.py
164 lines (143 loc) · 5.84 KB
/
capsnet_theano.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
import theano
import cProfile
from theano import tensor as T
from theano.tensor import nnet as NN
from theano import function
import numpy as np
def relu(t_in):
in_size = T.shape(t_in)
zeros = T.zeros(in_size)
results = T.switch(T.lt(in_size,zeros),zeros,in_size)
return results
def softmax(v_in):
ex = T.exp(v_in)
denom = T.sum(ex)
return ex/denom
def squash(v_in):
sq_norm = T.dot(v_in,v_in)
norm = T.sqrt(sq_norm)
factor = sq_norm/(1+sq_norm)/norm
return v_in*factor
"""
handle the transition between convolutional layers and capsule layers
since convolutional layers are scalar valued and capsule layers are vector valued
we solve this by paritioning the convolutional layers into groups of size ratio
and interpreting each of the groups as a capsule output, which is then squashed
"""
#batch x cnum x height x width
def primary_capsule_layer(im_conv, w,ratio, capsj):
#batch x height x width x cnum
im_reshaped = im_conv.dimshuffle((0,2,3,1))
#batch x capsi x capslen x 1
im_dim = list(T.shape(im_reshaped))
im_dim[1] = im_dim[1]*im_dim[2]*im_dim[3]
im_dim[2] = ratio
im_dim[3] = 1
im_reshaped = im_reshaped.reshape(im_dim)
#batch x capsi x capslen
im_reshaped = im_reshaped.dimshuffle((0,1,2))
#squash
im_shape = T.shape(im_reshaped)
im_flattened = im_reshaped.flatten(1)
im_squashed, updates = theano.scan(lambda v: squash(v), sequences=im_flattened)
im_squashed = im_squashed.reshape(im_shape)
return capsule_layer(im_squashed,w,capsj)
#batch x capsi x ilen, capsi x capsj x jlen x ilen -> batch x capsj x capsi x jlen
def capsule_layer(u,weights)
#[ilen] <dot> [jlen x ilen]
udim = T.shape(u)
wdim = T.shape(weights)
def batil(u_b, w):
def dotil(u_i, w_i):
results, updates = theano.scan(lambda w_ij, u_i: T.dot(u_i,w_ij), sequences=w_i, nonsequences=u_i)
return results
results, updates = theano.scan(lambda i, u, w: dotil(u[i],w[i]), n_steps=udim[1], nonsequences=(u_b,w))
return results
results, updates = theano.scan(batil,sequences=u,nonsequences=weights)
return caps_route(results)
#batch x capsj x capsi x ulen
def caps_route(u, r=3):
b = T.dtensor3('b')
v = None
for i in range(r):
#softmax
b_dim = T.shape(b)
b_flat = b.flatten(2)
c, updates = theano.scan(lambda v: softmax(b),sequences=b_flat)
c = c.reshape(b_dim)
#sum
#batch x capsj x vlen
u_sum, updates = theano.scan(lambda u_b: route_sum(u_b,c), sequences = u)
#squash
u_sum_dim = T.shape(u_sum)
u_flat = u_sum.flatten(2)
u_squashed, updates = theano.scan(lambda v: squash(v), sequences = u_flat)
v = u_squashed.reshape(u_sum_dim)
#update b
bat = theano.shared(0)
u_shape = T.shape(u)
bat_nums = u_shape[0]
capsj = u_shape[1]
capsi = u_shape[2]
def caps_route_inner_loop(bat):
bat = bat + 1
#capsj x ulen
v_b = v[bat]
#capsj x capsi x ulen
u_b = u[bat]
j = theano.shared(0)
#loop over j for for u_b
def caps_inner_j(j):
j = j + 1
#capsi x ulen
u_j = u_b[j]
#ulen x capsi
u_jt = u.transpose()
#implicit ij loop with matmul
return T.dot(v_b,u_jt)
results, updates = theano.scan(caps_inner_j,non_sequences=j,n_steps=capsj)
return results
b_deltas, updates = theano.scan(caps_route_inner_loop,non_sequences=bat, n_steps=bat_nums)
b = b + b_deltas
return v
#capsj x capsi x ulen, capsj x capsi
def route_sum(u_b, c):
results, updates = theano.scan(lambda u: T.dot(u,c), sequences=u_b)
return results
class capsnet:
def __init__(self, conv_size, conv_arch, caps_arch, primary_caps_ratio = 10):
image = T.dtensor4('image') #n x 3 (rgb) x height x width
convlayers = len(conv_arch)
capslayers = len(caps_arch)
convolutions = []
for i in range(convlayers):
convolutions.append(T.dtensor4('convolutions')) #n x num_convs x rows x cols
image_convs = [relu(NN.conv2d(image,convolutions[0]))]
for i in range(1,convlayers):
image_convs.append(relu(NN.conv2d(image_convs[-1],convolutions[i])))
weights = []
for i in range(capslayers):
weights.append(T.dmatrix())
caps_layers = [primary_capsule_layer(image_convs[-1],weights[0], primary_caps_ratio,caps_arch[0])]
for i in range(1,capslayers):
caps_layers.append(capsule_layer(caps_layers[-1],weights[i],caps_arch[i]))
self.ff = function([image,*convolutions,*weights],caps_layers[-1], mode=theano.Mode(optimizer="name"), profile=True) #compile
self.convolutions = [theano.shared(np.zeros((cnum,conv_size[0],conv_size[1]))) for cnum in conv_arch]
convis = [conv_arch[-1]/primary_caps_ratio]+caps_arch
self.weights = [theano.shared(np.zeros((convi,primary_capsule_ratio,primary_capsule_ratio))) for convi in convis]
def feed_foreward(images):
batch_size = T.shape(images)[0]
#hack, not doing this is possible but would be annoying
conv_stacks = [T.stack([convs]*batch_size) for convs in self.convolutions]
weight_stacks = [T.stack([weights]*batch_size) for weights in self.weights]
return self.ff(images,*conv_stacks,*weight_stacks)
def main():
conv_size = [10,10]
conv_arch = [30]
caps_arch = [30,20,10]
images = [np.zeros((1000,1000))]
c = capsnet(conv_size,conv_arch,caps_arch)
c.feed_foreward(images)
c.ff.profile.summary()
if __name__=='__main__':
main()