-
Notifications
You must be signed in to change notification settings - Fork 0
/
eq_models.py
290 lines (256 loc) · 10.3 KB
/
eq_models.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
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
import pdb
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from models import MLP
from equivariant_layers import ops_2_to_1, ops_1_to_2, ops_1_to_1, ops_2_to_2, set_ops_3_to_3, set_ops_4_to_4
from equivariant_layers_expand import eops_1_to_1, eops_1_to_2, eops_2_to_1, eops_2_to_2, eset_ops_3_to_3, eset_ops_4_to_4, eset_ops_1_to_3
class Eq1to1(nn.Module):
def __init__(self, in_dim, out_dim, ops_func=None):
super(Eq1to1, self).__init__()
self.basis_dim = 2
self.out_dim = out_dim
self.in_dim = in_dim
self.coefs = nn.Parameter(torch.normal(0, np.sqrt(2. / (in_dim + out_dim + self.basis_dim)), (in_dim, out_dim, self.basis_dim)))
self.bias = nn.Parameter(torch.zeros(1, out_dim, 1))
if ops_func is None:
self.ops_func = ops_1_to_1
else:
self.ops_func = ops_func
def forward(self, inputs):
ops = self.ops_func(inputs)
output = torch.einsum('dsb, nibd->nis', self.coefs, ops)
output = output + self.bias
return output
class Eq2to1(nn.Module):
def __init__(self, in_dim, out_dim):
super(Eq2to1, self).__init__()
self.basis_dim = 5
self.out_dim = out_dim
self.in_dim = in_dim
self.coefs = nn.Parameter(torch.normal(0, np.sqrt(2. / (in_dim + out_dim + self.basis_dim)), (in_dim, out_dim, self.basis_dim)))
self.bias = nn.Parameter(torch.zeros(1, out_dim, 1))
def forward(self, inputs):
'''
inputs: N x D x m x m
Returns: N x D x m
'''
ops = ops_2_to_1(inputs)
output = torch.einsum('dsb,ndbi->nsi', self.coefs, ops)
output = output + self.bias
return output
class Eq1to2(nn.Module):
def __init__(self, in_dim, out_dim):
super(Eq1to2, self).__init__()
self.basis_dim = 5
self.out_dim = out_dim
self.in_dim = in_dim
self.coefs = nn.Parameter(torch.normal(0, np.sqrt(2. / (in_dim + out_dim + self.basis_dim)), (in_dim, out_dim, self.basis_dim)))
# diag bias, all bias, mat diag bias
self.bias = nn.Parameter(torch.zeros(1, out_dim, 1, 1))
def forward(self, inputs):
ops = eops_1_to_2(inputs)
output = torch.einsum('dsb,ndbij->nsij', self.coefs, ops)
output = output + self.bias
return output
class Eq2to2(nn.Module):
def __init__(self, in_dim, out_dim, ops_func=eops_2_to_2):
super(Eq2to2, self).__init__()
self.basis_dim = 15
self.out_dim = out_dim
self.in_dim = in_dim
self.coefs = nn.Parameter(torch.normal(0, np.sqrt(2. / (in_dim + out_dim + self.basis_dim)), (in_dim, out_dim, self.basis_dim)))
self.bias = nn.Parameter(torch.zeros(1, out_dim, 1, 1))
self.diag_bias = nn.Parameter(torch.zeros(1, out_dim, 1, 1))
self.diag_eyes = {}
self.diag_eye = None #torch.eye(n).unsqueeze(0).unsqueeze(0).to(device)
if ops_func is None:
self.ops_func = ops_2_to_2
else:
self.ops_func = ops_func
def forward(self, inputs):
ops = self.ops_func(inputs)
output = torch.einsum('dsb,ndbij->nsij', self.coefs, ops)
n = output.shape[-1]
if n not in self.diag_eyes:
device = self.diag_bias.device
diag_eye = torch.eye(n).unsqueeze(0).unsqueeze(0).to(device)
diag_eye = torch.eye(n).unsqueeze(0).unsqueeze(0).to(device)
self.diag_eyes[n] = diag_eye
diag_eye = self.diag_eyes[n]
diag_bias = diag_eye.multiply(self.diag_bias)
output = output + self.bias + diag_bias
return output
class Eq1to2Net(nn.Module):
def __init__(self, in_dim, hid_dim, out_dim, layer_dims):
self.layers = nn.ModuleList(
[Eq1to2(in_dim, out_dim)] + \
[Eq2to2(out_dim, out_dim) for _ in range(1, len(layer_dims))]
)
self.fc_out = nn.Linear(hid_dim, out_dim)
def forward(self, inputs):
'''
inputs: N x D x m
outputs: N x D x m x m
'''
output = inputs
for l in self.layers:
inputs = F.relu(l(outputs))
output = self.fc_out(output)
return output
class Net1to1(nn.Module):
def __init__(self, layers, out_dim, out_model='linear', ops_func=None):
super(Net1to1, self).__init__()
self.layers = nn.ModuleList([Eq1to1(din, dout, ops_func) for din, dout in layers])
if out_model == 'linear':
self.out_net = nn.Linear(layers[-1][-1], out_dim)
else:
self.out_net = MLP(layers[-1][-1], out_dim)
def forward(self, x):
for layer in self.layers:
x = F.relu(layer(x))
x = x.permute(0, 2, 1)
output = self.out_net(x)
return output
class Net2to2(nn.Module):
def __init__(self, layers, out_dim, out_model='Linear', ops_func=None, **kwargs):
'''
layers: list of tuples (dim_in, dim_out)
out_dim: output dimension
n: size of input n \times n tensor
'''
super(Net2to2, self).__init__()
self.layers = nn.ModuleList([Eq2to2(din, dout, ops_func) for din, dout in layers])
if out_model == 'Linear':
self.out_net = nn.Linear(layers[-1][-1], out_dim)
elif out_model=='MLP':
self.out_net = MLP(layers[-1][0], kwargs['mlp_hid_dim'], out_dim)
def forward(self, x):
'''
x: N x d x m x m
Returns: N x m x m x out_dim
'''
for layer in self.layers:
x = F.relu(layer(x))
x = x.permute(0, 2, 3, 1)
output = self.out_net(x)
return output
class Eq1to3(nn.Module):
def __init__(self, in_dim, out_dim, ops_func=None):
super(Eq1to3, self).__init__()
self.basis_dim = 4
self.in_dim = in_dim
self.out_dim = out_dim
self.coefs = nn.Parameter(torch.normal(0, np.sqrt(2.0 / (in_dim + out_dim + self.basis_dim)),
(in_dim, out_dim, self.basis_dim)))
self.bias = nn.Parameter(torch.zeros(1, out_dim, 1, 1, 1))
if ops_func is None:
self.ops_func = eset_ops_1_to_3
else:
self.ops_func = ops_func
def forward(self, x):
ops = self.ops_func(x)
output = torch.einsum('dsb,ndbijk->nsijk', self.coefs, ops) # in/out/basis, batch/in/basis/ijk
output = output + self.bias
return output
class SetEq3to3(nn.Module):
def __init__(self, in_dim, out_dim, ops_func=None):
super(SetEq3to3, self).__init__()
self.basis_dim = 19
self.in_dim = in_dim
self.out_dim = out_dim
self.coefs = nn.Parameter(torch.normal(0, np.sqrt(2.0 / (in_dim + out_dim + self.basis_dim)),
(in_dim, out_dim, self.basis_dim)))
self.bias = nn.Parameter(torch.zeros(1, out_dim, 1, 1, 1))
if ops_func is None:
self.ops_func = set_ops_3_to_3
else:
self.ops_func = ops_func
def forward(self, x):
ops = self.ops_func(x)
output = torch.einsum('dsb,ndbijk->nsijk', self.coefs, ops) # in/out/basis, batch/in/basis/ijk
output = output + self.bias
return output
class SetEq4to4(nn.Module):
def __init__(self, in_dim, out_dim, ops_func):
super(SetEq4to4, self).__init__()
self.basis_dim = 69
self.in_dim = in_dim
self.out_dim = out_dim
self.coefs = nn.Parameter(torch.normal(0, np.sqrt(2.0 / (in_dim + out_dim + self.basis_dim)),
(in_dim, out_dim, self.basis_dim)))
self.bias = nn.Parameter(torch.zeros(1, out_dim, 1, 1, 1, 1))
if ops_func is None:
self.ops_func = set_ops_4_to_4
else:
self.ops_func = ops_func
def forward(self, x):
ops = self.ops_func(x)
output = torch.einsum('dsb,ndbijkl->nsijkl', self.coefs, ops) # in/out/basis, batch/in/basis/ijk
output = output + self.bias
return output
class SetNet3to3(nn.Module):
def __init__(self, layers, out_dim, out_model='Linear', ops_func=None):
super(SetNet3to3, self).__init__()
self.layers = nn.ModuleList([SetEq3to3(din, dout, ops_func) for din, dout in layers])
if out_model == 'Linear':
self.out_net = nn.Linear(layers[-1][1], out_dim)
else:
self.out_net = MLP(layers[-1][1], out_dim)
def forward(self, x):
'''
x: tensor of size Batch x feature x n x n x n
Return: tensor of size Batch x n x n x n
'''
for layer in self.layers:
x = F.relu(layer(x))
x = x.permute(0, 2, 3, 4, 1)
output = self.out_net(x)
return output
class SetNet4to4(nn.Module):
def __init__(self, layers, out_dim, out_model='Linear', ops_func=None):
super(SetNet4to4, self).__init__()
self.layers = nn.ModuleList([SetEq4to4(din, dout, ops_func) for din, dout in layers])
if out_model == 'Linear':
self.out_net = nn.Linear(layers[-1][1], out_dim)
else:
self.out_net = MLP(layers[-1][1], out_dim)
def forward(self, x):
'''
x: tensor of size Batch x feature x n x n x n
Return: tensor of size Batch x n x n x n
'''
for layer in self.layers:
x = F.relu(layer(x))
x = x.permute(0, 2, 3, 4, 5, 1)
output = self.out_net(x)
return output
if __name__ == '__main__':
N = 10
d_in = 5
d_hid = 3
d_out = 1
m = 3
x4 = torch.rand(N, d_in, m, m, m, m)
x3 = torch.rand(N, d_in, m, m, m)
x2 = torch.rand(N, d_in, m, m)
x1 = torch.rand(N, d_in, m)
m12a = Eq1to2(d_in, d_hid)
m12b = Eq1to2(d_hid, d_out)
m21a = Eq2to1(d_in, d_hid)
m21b = Eq2to1(d_hid, d_out)
out_dim = 4
layers = [(d_in, 6), (6, 3), (3, out_dim)]
net = Net2to2(layers, out_dim)
print(m12b(m21a(x2)).shape, f'expect {N} x 1 x {m} x {m}')
print(m21b(m12a(x1)).shape, f'expect {N} x 1 x {m}')
print(net(x2).shape, 'expect dim:', f'{N} x {m} x {m} x {out_dim}')
n11 = Net1to1(layers, out_dim)
x11 = torch.rand(N, m, d_in)
print(n11(x11), '1->1')
print('done 11')
m33 = SetEq3to3(d_in, d_hid)
print(m33(x3).shape, f'expect {N} x {d_hid} x {m} x {m} x {m}')
m44 = SetEq4to4(d_in, d_hid)
print(m44(x4).shape, f'expect {N} x {d_hid} x {m} x {m} x {m}, {m}')