-
Notifications
You must be signed in to change notification settings - Fork 62
/
Copy pathmegnet.py
199 lines (169 loc) · 7.37 KB
/
megnet.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
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from utils import RBFExpansion
from torch_scatter import scatter
from torch.nn import Embedding
from typing import Union
from torch import Tensor
from torch_sparse import SparseTensor
from torch_geometric.nn.conv import MessagePassing
from torch_geometric.typing import Adj, OptTensor, PairTensor
from torch_geometric.utils.num_nodes import maybe_num_nodes
from torch_geometric.nn import Set2Set as set2set
class MEGConv(MessagePassing):
_alpha: OptTensor
def __init__(
self,
v_input_dim=16,
e_input_dim=100,
u_input_dim=2,
n1 = 64,
n2 = 32,
concat: bool = True,
dropout: float = 0.0,
bias: bool = True,
root_weight: bool = True,
**kwargs,
):
kwargs.setdefault('aggr', 'mean')
super(MEGConv, self).__init__(node_dim=0, **kwargs)
self.n1 = n1
self.n2 = n2
self.v_input_dim = v_input_dim
self.e_input_dim = e_input_dim
self.u_input_dim = u_input_dim
self.concat = concat
self.dropout = dropout
# Linear for v
self.linear_v1 = nn.Linear(v_input_dim + u_input_dim + n2, n1)
self.linear_v2 = nn.Linear(n1, n1)
self.linear_v3 = nn.Linear(n1, n2)
# Linear for e
self.linear_e1 = nn.Linear(2 * v_input_dim + e_input_dim + u_input_dim, n1)
self.linear_e2 = nn.Linear(n1, n1)
self.linear_e3 = nn.Linear(n1, n2)
# Linear for u
self.linear_u1 = nn.Linear(n2 + u_input_dim + n2, n1)
self.linear_u2 = nn.Linear(n1, n1)
self.linear_u3 = nn.Linear(n1, n2)
self.softplus = nn.Softplus()
self.reset_parameters()
def reset_parameters(self):
self.linear_v1.reset_parameters()
self.linear_v2.reset_parameters()
self.linear_v3.reset_parameters()
self.linear_e1.reset_parameters()
self.linear_e2.reset_parameters()
self.linear_e3.reset_parameters()
self.linear_u1.reset_parameters()
self.linear_u2.reset_parameters()
self.linear_u3.reset_parameters()
def forward(self, x: Union[Tensor, PairTensor], edge_index: Adj, batch=None, edge_attr: OptTensor = None, u_rep=None ,return_attention_weights=None):
v = x
e = edge_attr
u = u_rep # batch * k
# phi_e calculation
e_p = torch.cat((v[edge_index[0]], v[edge_index[1]], u[batch][edge_index[0]], e), dim=-1)
e_p = self.softplus(self.linear_e1(e_p))
e_p = self.softplus(self.linear_e2(e_p))
e_p = self.softplus(self.linear_e3(e_p))
# v_p calculation
edge_to_v = scatter(e_p, edge_index[1], dim=0, reduce="mean")
v_p = torch.cat((v, edge_to_v, u[batch]), dim=-1)
v_p = self.softplus(self.linear_v1(v_p))
v_p = self.softplus(self.linear_v2(v_p))
v_p = self.softplus(self.linear_v3(v_p))
# u_p calculation
ue = scatter(edge_to_v, batch, dim=0, reduce="mean")
uv = scatter(v_p, batch, dim=0, reduce="mean")
u_p = torch.cat((ue, uv, u), dim=-1)
u_p = self.softplus(self.linear_u1(u_p))
u_p = self.softplus(self.linear_u2(u_p))
u_p = self.softplus(self.linear_u3(u_p))
return v_p, e_p, u_p
class MEGNET(nn.Module):
"""megnet pyg implementation."""
def __init__(self,larger=1):
"""Set up megnet modules."""
super().__init__()
# self.embedding = self.embedding = Embedding(100, 16*larger)
self.embedding = nn.Linear(92, 16*larger)
self.batchsize = 64
self.rbf = RBFExpansion(
vmin=0,
vmax=5.0,
bins=100,
lengthscale=0.5,
)
self.softplus = nn.Softplus()
self.meglayer1 = MEGConv(v_input_dim=32*larger, e_input_dim=32*larger, u_input_dim=32*larger)
self.meglayer2 = MEGConv(v_input_dim=32*larger, e_input_dim=32*larger, u_input_dim=32*larger)
self.meglayer3 = MEGConv(v_input_dim=32*larger, e_input_dim=32*larger, u_input_dim=32*larger)
self.ffv0 = nn.Sequential(nn.Linear(16*larger, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffv1 = nn.Sequential(nn.Linear(32*larger, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffv2 = nn.Sequential(nn.Linear(32*larger, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffe0 = nn.Sequential(nn.Linear(100, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffe1 = nn.Sequential(nn.Linear(32*larger, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffe2 = nn.Sequential(nn.Linear(32*larger, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffu0 = nn.Sequential(nn.Linear(2, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffu1 = nn.Sequential(nn.Linear(32*larger, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.ffu2 = nn.Sequential(nn.Linear(32*larger, 64*larger), nn.Softplus(), nn.Linear(64*larger, 32*larger), nn.Softplus())
self.node_linear = nn.Linear(32*larger, 16*larger)
self.node_s2s = set2set(in_channels=16*larger, processing_steps=3)
self.edge_linear = nn.Linear(32*larger, 16*larger)
self.edge_s2s = set2set(in_channels=16*larger, processing_steps=3)
self.fc_out = nn.Sequential(nn.Linear(96*larger, 32*larger), nn.Softplus(), nn.Linear(32*larger, 16*larger), nn.Softplus(), nn.Linear(16*larger, 9))
def forward(self, data, test=False) -> torch.Tensor:
z = data.x
# z = z.squeeze(-1).long()
# calculate v 16
v = self.embedding(z)
# calculate e 100
e = torch.norm(data.edge_attr, dim=1)
e = self.rbf(e)
# calculate u batch*2
u = torch.zeros(self.batchsize, 2).float().to(z.device)
if test:
u = torch.zeros(1, 2).float().to(z.device)
v = self.ffv0(v)
e = self.ffe0(e)
u = self.ffu0(u)
# meg layer one
v1, e1, u1 = self.meglayer1(x=v, edge_index=data.edge_index, batch=data.batch, edge_attr=e, u_rep=u)
v1 = v1 + v
e1 = e1 + e
u1 = u1 + u
v_t = v1
e_t = e1
u_t = u1
# fc
v1 = self.ffv1(v1)
e1 = self.ffe1(e1)
u1 = self.ffu1(u1)
# meg layer two
v1, e1, u1 = self.meglayer2(x=v1, edge_index=data.edge_index, batch=data.batch, edge_attr=e1, u_rep=u1)
v1 = v1 + v_t
e1 = e1 + e_t
u1 = u1 + u_t
v_t = v1
e_t = e1
u_t = u1
# fc
v1 = self.ffv2(v1)
e1 = self.ffe2(e1)
u1 = self.ffu2(u1)
# meg layer three
v1, e1, u1 = self.meglayer3(x=v1, edge_index=data.edge_index, batch=data.batch, edge_attr=e1, u_rep=u1)
v1 = v1 + v_t
e1 = e1 + e_t
u1 = u1 + u_t
# set2set
v1 = self.node_linear(v1)
e1 = self.edge_linear(e1)
node_vec = self.node_s2s(v1, data.batch)
edge_vec = self.edge_s2s(e1, data.batch[data.edge_index[1]])
final_vec = torch.cat((node_vec, edge_vec, u1), dim=-1)
out = self.fc_out(final_vec)
return out