-
Notifications
You must be signed in to change notification settings - Fork 1
/
ground_metric_gm.py
142 lines (128 loc) · 6.51 KB
/
ground_metric_gm.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
import torch
import torch.nn.functional as F
# GM stands for graph matching
class Ground_Metric_GM:
def __init__(
self,
model_1_param:torch.tensor=None,
model_2_param:torch.tensor=None,
conv_param:bool=False,
bias_param:bool=False,
pre_conv_param:bool=False,
pre_conv_image_size_squared:int=None ):
'''
if [bias_param] is True, the input parameters should be 1-d tensor of the size
[cur_neuron_num]. Else:
if [conv_param] is False, the input parameters should be 2-d tensors of the size
[cur_neuron_num] x [pre_neuron_num], where each entity is simply a float
if [conv_param] is True, the input parameters should be 3-d tensors of the size
[cur_neuron_num] x [pre_neuron_num] x [kernel_size_squared]
'''
self._sanity_check(
model_1_param, model_2_param, conv_param, bias_param,
pre_conv_param, pre_conv_image_size_squared )
'''
if [pre_conv_param] is False:
transforms the parameters so that the ultimate params are of the size
1 x [cur_neuron_num * pre_neuron_num] x [edge_weight_length],
where [edge_weight_length] is 1 if conv_param=False and [kernel_size_squared] otherwise
elif [conv_param] is False and [bias_param] is False:
transforms the parameters so that the ultimate params are of the size
1 x [cur_neuron_num * pre_neuron_num / pre_kernel_size] x [pre_kernel_size]
after the transformation, the weights in the middle dimension would be:
0~0, 1~0, ..., N-1~0, ..., 0~M-1, 1~M-1, ..., N-1~M-1
which is the order of edges
( N stands for [pre_neuron_num], and M stands for [cur_neuron_num] )
'''
self.model_1_param = model_1_param
self.model_2_param = model_2_param
self.conv_param = conv_param
self.bias_param = bias_param
# bias, or fully-connected from linear
if bias_param is True or (conv_param is False and pre_conv_param is False):
self.model_1_param = self.model_1_param.reshape( 1, -1, 1 )
self.model_2_param = self.model_2_param.reshape( 1, -1, 1 )
# fully-connected from conv
elif conv_param is False and pre_conv_param is True:
self.model_1_param = self.model_1_param.reshape( 1, -1, pre_conv_image_size_squared )
self.model_2_param = self.model_2_param.reshape( 1, -1, pre_conv_image_size_squared )
# conv
else:
self.model_1_param = self.model_1_param.reshape( 1, -1, model_1_param.shape[-1] )
self.model_2_param = self.model_2_param.reshape( 1, -1, model_2_param.shape[-1] )
def _sanity_check(
self,
model_1_param:torch.tensor,
model_2_param:torch.tensor,
conv_param:bool,
bias_param:bool,
pre_conv_param:bool,
pre_conv_image_size_squared:int ):
assert model_1_param is not None
assert model_2_param is not None
if bias_param is True:
assert len( model_1_param.shape ) == 1
elif conv_param is True:
assert len( model_1_param.shape ) == 3
else:
assert len( model_1_param.shape ) == 2
if pre_conv_param is True:
assert type(pre_conv_image_size_squared) == int
assert model_1_param.shape == model_2_param.shape
def process_distance( self, p:int=2 ):
'''
returns the p-norm pair-wise distance between [self.model_1_param] and
[self.model_2_param].
if [bias_param] is False
the returned value is a tensor of the size
[cur_neuron_num * pre_neuron_num] x [cur_neuron_num * pre_neuron_num], such that
` r[i][j] = the distance between edge i in model 1 and edge j in model 2 `
( edge i stands for the weight between neuron a in the previous layer and neuron
b in the current layer, where ` i = a + b*N ` )
else if [bias_param] is True
the returned value is a tensor of the size
[cur_neuron_num] x [cur_neuron_num]
'''
return torch.cdist(
self.model_1_param.to(torch.float),
self.model_2_param.to(torch.float),
p=p )[0]
def process_soft_affinity( self, p:int=2 ):
'''
returns the p-norm pair-wise soft affinity between [self.model_1_param] and
[self.model_2_param].
if [bias_param] is False
the returned value [r] is a tensor of the size
[cur_neuron_num * pre_neuron_num] x [cur_neuron_num * pre_neuron_num], such that
` r[i][j] = the affinity between edge i in model 1 and edge j in model 2 `
( edge i stands for the weight between neuron a in the previous layer and neuron
b in the current layer, where ` i = a + b*N ` )
else if [bias_param] is True
the returned value is of the size
[cur_neuron_num] x [cur_neuron_num]
'''
return torch.exp( 0 - self.process_distance( p=p ) )
if __name__ == '__main__':
print( '---------- testing on fully-connected layers weights ----------' )
param_1 = torch.tensor(
[ [1, 2, 3, 4],
[5, 6, 7, 8] ] )
param_2 = torch.tensor(
[ [2, 4, 6, 8],
[3, 5, 7, 9] ] )
gm = Ground_Metric_GM( param_1, param_2, conv_param=False, bias_param=False )
print( f'\tdistance is \n\t{gm.process_distance()},\n\tsoft_affinity is \n\t{gm.process_soft_affinity()}' )
print( '---------- testing on convlutional layers weights ----------' )
param_1 = torch.tensor(
[ [[1,2,3,4], [1,2,3,4], [1,2,3,4], [1,2,3,4]],
[[1,2,3,4], [1,2,3,4], [1,2,3,4], [1,2,3,4]] ] )
param_2 = torch.tensor(
[ [[1,2,3,4], [5,6,7,8], [1,2,3,4], [5,6,7,8]],
[[1,2,3,4], [5,6,7,8], [1,2,3,4], [5,6,7,8]] ] )
gm = Ground_Metric_GM( param_1, param_2, conv_param=True, bias_param=False )
print( f'\tdistance is \n\t{gm.process_distance()},\n\tsoft_affinity is \n\t{gm.process_soft_affinity()}' )
print( '---------- testing on layers biases ----------' )
param_1 = torch.tensor([1,2,3,4,5,6,7,8])
param_2 = torch.tensor([8,7,6,5,4,3,2,1])
gm = Ground_Metric_GM( param_1, param_2, conv_param=False, bias_param=True )
print( f'\tdistance is \n\t{gm.process_distance()},\n\tsoft_affinity is \n\t{gm.process_soft_affinity()}' )