-
Notifications
You must be signed in to change notification settings - Fork 0
/
equivariant_layers_expand.py
233 lines (192 loc) · 8.13 KB
/
equivariant_layers_expand.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
import pdb
import torch
import torch.nn as nn
def check_shape(x, shape):
assert len(x.shape) == len(shape)
for xs, s in zip(x.shape, shape):
assert xs == s
def eops_1_to_1(inputs, normalize=False):
dim = inputs.shape[-1]
N, d, m = inputs.shape
dim = inputs.shape[-1]
sums = inputs.sum(dim=2, keepdim=True)
op1 = inputs
op2 = sums.expand(-1, -1, dim)
return torch.stack([op1, op2], dim=2)
def eops_1_to_2(inputs, normalize=False):
'''
inputs: N x D x m tensor
'''
N, D, m = inputs.shape
dim = inputs.shape[-1]
sum_all = inputs.sum(dim=2, keepdim=True) # N x D x 1
op1 = torch.diag_embed(inputs)
op2 = torch.diag_embed(sum_all.expand(-1, -1, dim))
op3 = inputs.unsqueeze(2).expand(-1, -1, dim, -1)
op4 = inputs.unsqueeze(3).expand(-1, -1, -1, dim)
op5 = sum_all.unsqueeze(3).expand(-1, -1, dim, dim)
return torch.stack([op1, op2, op3, op4, op5], dim=2)
def eops_2_to_1(inputs, normalize=False):
N, D, m, m = inputs.shape
dim = inputs.shape[-1]
diag_part = torch.diagonal(inputs, dim1=-2, dim2=-1)
sum_diag_part = diag_part.sum(dim=2, keepdims=True)
sum_rows = inputs.sum(dim=3)
sum_cols = inputs.sum(dim=2)
sum_all = inputs.sum(dim=(2,3))
op1 = diag_part
op2 = sum_diag_part.expand(-1, -1, dim)
op3 = sum_rows
op4 = sum_cols
op5 = sum_all.unsqueeze(2).expand(-1, -1, dim)
ops = [op1, op2, op3, op4, op5]
return torch.stack([op1, op2, op3, op4, op5], dim=2)
def eops_2_to_2(inputs, normalize=False):
N, D, m, m = inputs.shape
dim = inputs.shape[-1]
diag_part = torch.diagonal(inputs, dim1=-2, dim2=-1) # N x D x m
sum_diag_part = diag_part.sum(dim=2, keepdims=True) # N x D x 1
sum_rows = inputs.sum(dim=3) # N x D x m
sum_cols = inputs.sum(dim=2) # N x D x m
sum_all = inputs.sum(dim=(2,3)) # N x D
ops = [None] * (15 + 1)
ops[1] = torch.diag_embed(diag_part) # N x D x m x m
ops[2] = torch.diag_embed(sum_diag_part.expand(-1, -1, dim))
ops[3] = torch.diag_embed(sum_rows)
#ops[4] = torch.diag_embed(sum_rows)
ops[4] = torch.diag_embed(sum_cols)
ops[5] = torch.diag_embed(sum_all.unsqueeze(-1).expand(-1, -1, dim))
ops[6] = sum_cols.unsqueeze(3).expand(-1, -1, -1, dim)
ops[7] = sum_rows.unsqueeze(3).expand(-1, -1, -1, dim)
ops[8] = sum_cols.unsqueeze(2).expand(-1, -1, dim, -1)
ops[9] = sum_rows.unsqueeze(2).expand(-1, -1, dim, -1)
ops[10] = inputs
ops[11] = torch.transpose(inputs, 2, 3)
ops[12] = diag_part.unsqueeze(3).expand(-1, -1, -1, dim)
ops[13] = diag_part.unsqueeze(2).expand(-1, -1, dim, -1)
ops[14] = sum_diag_part.unsqueeze(3).expand(-1, -1, dim, dim)
ops[15] = sum_all.unsqueeze(-1).unsqueeze(-1).expand(-1, -1, dim, dim)
for i in range(1, 16):
op = ops[i]
if i == 5 or i == 15:
ops[i] = torch.divide(op, dim * dim)
else:
ops[i] = torch.divide(op, dim)
return torch.stack(ops[1:], dim=2)
def eset_ops_1_to_3(inputs):
N, D, m = inputs.shape
dim = inputs.shape[-1]
sum_all = inputs.sum(dim=-1)
x = inputs
ops = [None] * 5
ops[1] = sum_all.view(N, D, 1, 1, 1).expand(-1, -1, dim, dim, dim) / m
ops[2] = x.view(N, D, m, 1, 1).expand(-1, -1, -1, m, m)
ops[3] = x.view(N, D, 1, m, 1).expand(-1, -1, m, -1, m)
ops[4] = x.view(N, D, 1, 1, m).expand(-1, -1, m, m, -1)
return torch.stack(ops[1:], dim=2)
def eset_ops_3_to_3(inputs, normalize=True):
N, D, m, m, m = inputs.shape
dim = inputs.shape[-1]
# only care about marginalizing over each of the individual indices
sum_all = inputs.sum(dim=(-1, -2, -3))
sum_c1 = inputs.sum(dim=-1)
sum_c2 = inputs.sum(dim=-2)
sum_c3 = inputs.sum(dim=-3)
sum_c12 = inputs.sum(dim=(-1, -2))
sum_c13 = inputs.sum(dim=(-1, -3))
sum_c23 = inputs.sum(dim=(-2, -3))
# dont really care about diagonal
ops = [None] * 20
ops[1] = sum_all.view(N, D, 1, 1, 1).expand(-1, -1, dim, dim, dim) / (m * m * m)
ops[2] = sum_c1.unsqueeze(-1).expand(-1, -1, -1, -1, m) / (m)
ops[3] = sum_c1.unsqueeze(-2).expand(-1, -1, -1, m, -1) / (m)
ops[4] = sum_c1.unsqueeze(-3).expand(-1, -1, m, -1, -1) / (m)
ops[5] = sum_c2.unsqueeze(-1).expand(-1, -1, -1, -1, m) / (m)
ops[6] = sum_c2.unsqueeze(-2).expand(-1, -1, -1, m, -1) / (m)
ops[7] = sum_c2.unsqueeze(-3).expand(-1, -1, m, -1, -1) / (m)
ops[8] = sum_c3.unsqueeze(-1).expand(-1, -1, -1, -1, m) / m
ops[9] = sum_c3.unsqueeze(-2).expand(-1, -1, -1, m, -1) / m
ops[10] = sum_c3.unsqueeze(-3).expand(-1, -1, m, -1, -1) / m
ops[11] = sum_c12.view(N, D, m, 1, 1).expand(-1, -1, -1, m, m) / (m*m)
ops[12] = sum_c12.view(N, D, 1, m, 1).expand(-1, -1, m, -1, m) / (m*m)
ops[13] = sum_c12.view(N, D, 1, 1, m).expand(-1, -1, m, m, -1) / (m*m)
ops[14] = sum_c13.view(N, D, m, 1, 1).expand(-1, -1, -1, m, m) / (m*m)
ops[15] = sum_c13.view(N, D, 1, m, 1).expand(-1, -1, m, -1, m) / (m*m)
ops[16] = sum_c13.view(N, D, 1, 1, m).expand(-1, -1, m, m, -1) / (m*m)
ops[17] = sum_c23.view(N, D, m, 1, 1).expand(-1, -1, -1, m, m) / (m*m)
ops[18] = sum_c23.view(N, D, 1, m, 1).expand(-1, -1, m, -1, m) / (m*m)
ops[19] = sum_c23.view(N, D, 1, 1, m).expand(-1, -1, m, m, -1) / (m*m)
#if normalize:
# ops[1] = torch.divide(ops[1], dim * dim * dim)
# for d in range(2, 11):
# ops[d] = torch.divide(ops[d], dim)
# for d in range(11, 20):
# ops[d] = torch.divide(ops[d], dim * dim)
return torch.stack(ops[1:], dim=2)
def eset_ops_4_to_4(inputs, normalize=False):
N, D, m, _, _, _ = inputs.shape
sum_all = inputs.sum(dim=(-1, -2, -3, -4))
c1, c2, c3 = [], [], []
# collapse 1 dim
sum_c1 = inputs.sum(dim=-1)
sum_c2 = inputs.sum(dim=-2)
sum_c3 = inputs.sum(dim=-3)
sum_c4 = inputs.sum(dim=-4)
c1s = [sum_c1, sum_c2, sum_c3, sum_c4]
# collapse 2
sum_c12 = inputs.sum(dim=(-1, -2))
sum_c13 = inputs.sum(dim=(-1, -3))
sum_c14 = inputs.sum(dim=(-1, -4))
sum_c23 = inputs.sum(dim=(-2, -3))
sum_c24 = inputs.sum(dim=(-2, -4))
sum_c34 = inputs.sum(dim=(-3, -4))
c2s = [sum_c12, sum_c13, sum_c14, sum_c23, sum_c24, sum_c34]
# collapse 3
sum_c123 = inputs.sum(dim=(-1, -2, -3))
sum_c124 = inputs.sum(dim=(-1, -2, -4))
sum_c134 = inputs.sum(dim=(-1, -3, -4))
sum_c234 = inputs.sum(dim=(-2, -3, -4))
c3s = [sum_c123, sum_c124, sum_c134, sum_c234]
# broadcast
ops = []
ops.append(sum_all.view(N, D, 1, 1, 1, 1).expand(-1, -1, m, m, m, m))
# broadcast collapsed 1 dims
for c1 in c1s:
ops.append(c1.view(N, D, m, m, m, 1).expand(-1, -1, -1, -1, -1, m))
ops.append(c1.view(N, D, m, m, 1, m).expand(-1, -1, -1, -1, m, -1))
ops.append(c1.view(N, D, m, 1, m, m).expand(-1, -1, -1, m, -1, -1))
ops.append(c1.view(N, D, 1, m, m, m).expand(-1, -1, m, -1, -1, -1))
for c2 in c2s:
ops.append(c2.view(N, D, m, m, 1, 1).expand(-1, -1, -1, -1, m, m))
ops.append(c2.view(N, D, m, 1, m, 1).expand(-1, -1, -1, m, -1, m))
ops.append(c2.view(N, D, 1, m, m, 1).expand(-1, -1, m, -1, -1, m))
ops.append(c2.view(N, D, m, 1, 1, m).expand(-1, -1, -1, m, m, -1))
ops.append(c2.view(N, D, 1, m, 1, m).expand(-1, -1, m, -1, m, -1))
ops.append(c2.view(N, D, 1, 1, m, m).expand(-1, -1, m, m, -1, -1))
for c3 in c3s:
ops.append(c3.view(N, D, m, 1, 1, 1).expand(-1, -1, -1, m, m, m))
ops.append(c3.view(N, D, 1, m, 1, 1).expand(-1, -1, m, -1, m, m))
ops.append(c3.view(N, D, 1, 1, m, 1).expand(-1, -1, m, m, -1, m))
ops.append(c3.view(N, D, 1, 1, 1, m).expand(-1, -1, m, m, m, -1))
return torch.stack(ops, dim=2)
if __name__ == '__main__':
N = 32
D = 16
m = 2
x = torch.rand(N, D, m)
x = torch.rand(N, D, m)
x2 = torch.rand(N, D, m, m)
x3 = torch.rand(N, D, m, m, m)
x4 = torch.rand(N, D, m, m, m, m)
o = eops_1_to_1(x)
print('1->1 okay')
o2 = eops_1_to_2(x)
print('1->2 okay')
o1 = eops_2_to_1(x2)
print('2->1 okay')
o22 = eops_2_to_2(x2)
print('2->2 okay')
t33 = eset_ops_3_to_3(x3)
print(t33.shape)
t44 = eset_ops_4_to_4(x4)
print(t44.shape)