-
Notifications
You must be signed in to change notification settings - Fork 0
/
InputEncoder.py
129 lines (123 loc) · 5.18 KB
/
InputEncoder.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
import torch
import torch.nn as nn
from Emb import MultiEmbedding
from utils import MLP
from basis_layers import rbf_class_mapping
from typing import Final
EPS = 1e-4
class QInputEncoder(nn.Module):
LambdaBound: Final[float]
laplacian: Final[bool]
useea: Final[bool]
sqrtlambda: Final[bool]
def __init__(self, featdim, hiddim, LambdaBound=1e-4, **kwargs) -> None:
super().__init__()
self.LambdaBound = LambdaBound
self.featdim = featdim
self.useea = False
self.sqrtlambda = kwargs["sqrtlambda"]
if kwargs["dataset"] in ["pepfunc", "pepstruct"]:
self.xemb = MultiEmbedding(hiddim, [18, 4, 8, 8, 6, 2, 7, 3, 3],
**kwargs["xemb"])
self.edgeEmb = MultiEmbedding(hiddim, [5, 5, 5], **kwargs["xemb"])
elif kwargs["dataset"].startswith("qm9"):
self.xemb = nn.Linear(11, hiddim)
self.edgeEmb = nn.Linear(4, hiddim)
elif kwargs["dataset"] == "PTC_MR":
self.xemb = nn.Linear(18, hiddim)
self.edgeEmb = nn.Linear(4, hiddim)
elif kwargs["dataset"] == "PROTEINS":
self.xemb = nn.Linear(4, hiddim)
self.useea = True
elif kwargs["dataset"] == "MUTAG":
self.xemb = nn.Linear(7, hiddim)
self.edgeEmb = nn.Linear(4, hiddim)
elif kwargs["dataset"] == "DD":
self.xemb = nn.Linear(89, hiddim)
self.useea = True
elif kwargs["dataset"] == "IMDB-BINARY":
self.xemb = nn.Linear(1, hiddim)
self.useea = True
elif kwargs["dataset"] == "ogbg-molhiv":
self.xemb = MultiEmbedding(hiddim, kwargs["xembdims"],
**kwargs["xemb"])
tmp = kwargs["xemb"].copy()
self.tedgeEmb = MultiEmbedding(featdim, [100, 100, 100], **tmp)
self.edgeEmb = lambda x: self.tedgeEmb(x.to(torch.long))
elif kwargs["dataset"] in ["zinc", "zinc-full"]:
self.xemb = MultiEmbedding(hiddim, [40], **kwargs["xemb"])
self.edgeEmb = MultiEmbedding(hiddim, [20], **kwargs["xemb"])
self.useea = True
elif kwargs["dataset"] == "pascalvocsp":
self.xemb = nn.Sequential(nn.Linear(14, hiddim))
self.edgeEmb = nn.Sequential(nn.Linear(2, hiddim))
else:
raise NotImplementedError
self.LambdaEmb = rbf_class_mapping[kwargs["lexp"]](
hiddim, **kwargs["basic"], **kwargs["lambdaemb"])
self.degreeEmb = MultiEmbedding(
hiddim, [100], **kwargs["xemb"]) if kwargs["degreeemb"] else None
self.normA = kwargs["normA"]
self.laplacian = kwargs["laplacian"]
self.decompnoise = kwargs["decompnoise"]
self.use_pos = kwargs["use_pos"]
if self.use_pos:
self.distEmb = rbf_class_mapping[kwargs["lexp"]](hiddim, **kwargs["basic"], **kwargs["lambdaemb"])
def setnoiseratio(self, ratio):
self.decompnoise = ratio
def forward(self, A, X, nodemask, pos=None):
'''
A (b, n, n, d)
'''
if self.useea:
A = A.squeeze(-1).to(torch.float)
else:
eA = self.edgeEmb(A)
A = torch.any(A != 0, dim=-1).to(torch.float)
D = torch.sum(A, dim=-1) # (#graph, N)
if self.laplacian:
L = torch.diag_embed(D) - A
else:
L = A # (#graph, N, N)
if self.normA:
tD = torch.clamp_min(D, 1) # (# graph, N, N)
tD = torch.rsqrt_(tD)
L = tD.unsqueeze(1) * L * tD.unsqueeze(2)
if self.training:
N = L.shape[1]
perm = torch.randperm(N, device=L.device)
L = L[:, perm][:, :, perm]
invperm = torch.empty_like(perm)
invperm[perm] = torch.arange(N, device=perm.device)
Lambda, U = torch.linalg.eigh(L)
U = U[:, invperm] # (#graph, N, M)
else:
Lambda, U = torch.linalg.eigh(L)
if self.laplacian:
Lambda = Lambda[:, 1:]
U = U[:, :, 1:]
X = self.xemb(X)
if self.degreeEmb is not None:
X *= self.degreeEmb(D.to(torch.long).unsqueeze(-1))
X.masked_fill_(nodemask.unsqueeze(-1), 0)
Lambdamask = torch.abs(
Lambda) < self.LambdaBound # (#graph, M) # mask zero frequency
U.masked_fill_(Lambdamask.unsqueeze(1), 0)
U.masked_fill_(nodemask.unsqueeze(-1), 0)
if self.sqrtlambda:
Lambda = torch.sqrt(torch.relu_(Lambda))
if self.training:
Lambda += self.decompnoise * torch.randn_like(Lambda)
U += self.decompnoise * torch.randn_like(U)
LambdaEmb = self.LambdaEmb(Lambda) # (#graph, M, d2)
if self.use_pos:
dist = torch.norm(pos.unsqueeze(1) - pos.unsqueeze(2), dim=-1)
dist = self.distEmb(dist)
eA = eA * dist
if self.useea:
U = U.unsqueeze(-1)
else:
eA.masked_fill_(nodemask.unsqueeze(1).unsqueeze(-1), 0)
eA.masked_fill_(nodemask.unsqueeze(2).unsqueeze(-1), 0)
U = torch.einsum("bnmd,bml->bnld", eA, U)
return LambdaEmb, Lambdamask, U, X, nodemask