-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgnn_mol.py
165 lines (132 loc) · 6.13 KB
/
gnn_mol.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
import torch
import torch.nn as nn
from torch_geometric.nn import MessagePassing
import torch.nn.functional as F
from torch_geometric.nn import global_mean_pool, global_add_pool
from MolEncoders import AtomEncoder, BondEncoder
from torch_geometric.utils import degree
from torch_scatter import scatter_add
from torch_geometric.nn.inits import reset
import math
import pdb
nn_act = torch.nn.ReLU() #ReLU()
F_act = F.relu
class GINConv(MessagePassing):
def __init__(self, emb_dim):
super(GINConv, self).__init__(aggr = "add")
self.mlp = torch.nn.Sequential(torch.nn.Linear(emb_dim, 2*emb_dim), torch.nn.BatchNorm1d(2*emb_dim), nn_act, torch.nn.Linear(2*emb_dim, emb_dim))
self.eps = torch.nn.Parameter(torch.Tensor([0]))
self.bond_encoder = BondEncoder(emb_dim = emb_dim)
def forward(self, x, edge_index, edge_attr, edge_weight=None):
edge_embedding = self.bond_encoder(edge_attr)
out = self.mlp((1 + self.eps) *x + self.propagate(edge_index, x=x, edge_attr=edge_embedding, edge_weight=edge_weight))
return out
def message(self, x_j, edge_attr, edge_weight=None):
if edge_weight is not None:
mess = F_act((x_j + edge_attr) * edge_weight)
else:
mess = F_act(x_j + edge_attr)
return mess
# return F_act(x_j + edge_attr)
def update(self, aggr_out):
return aggr_out
class GINEConv(MessagePassing):
def __init__(self, emb_dim):
super(GINEConv, self).__init__(aggr = "add")
self.mlp = torch.nn.Sequential(torch.nn.Linear(emb_dim, 2*emb_dim),
torch.nn.BatchNorm1d(2*emb_dim),
torch.nn.ReLU(),
torch.nn.Linear(2*emb_dim, emb_dim))
self.eps = torch.nn.Parameter(torch.Tensor([0]))
self.bond_encoder = BondEncoder(emb_dim = emb_dim)
def forward(self, x, edge_index, edge_attr, edge_weight=None):
edge_embedding = self.bond_encoder(edge_attr)
out = self.mlp((1 + self.eps) *x + self.propagate(edge_index, x=x, edge_attr=edge_embedding, edge_weight=edge_weight))
return out
def message(self, x_j, edge_attr, edge_weight=None):
if edge_weight is not None:
mess = F.relu((x_j + edge_attr) * edge_weight)
else:
mess = F.relu(x_j + edge_attr)
return mess
def update(self, aggr_out):
return aggr_out
class GINMolHeadEncoder(torch.nn.Module):
def __init__(self, num_layer, emb_dim, drop_ratio=0.5, JK="last", residual=True):
super(GINMolHeadEncoder, self).__init__()
self.num_layer = num_layer
self.drop_ratio = drop_ratio
self.JK = JK
self.residual = residual
self.atom_encoder = AtomEncoder(emb_dim)
self.convs = torch.nn.ModuleList()
self.batch_norms = torch.nn.ModuleList()
for layer in range(num_layer):
self.convs.append(GINConv(emb_dim))
self.batch_norms.append(torch.nn.BatchNorm1d(emb_dim))
def forward(self, x, edge_index, edge_attr, batch):
h_list = [self.atom_encoder(x)]
for layer in range(self.num_layer):
h = self.convs[layer](h_list[layer], edge_index, edge_attr)
h = self.batch_norms[layer](h)
if layer == self.num_layer - 1:
h = F.dropout(h, self.drop_ratio, training = self.training)
else:
h = F.dropout(F_act(h), self.drop_ratio, training = self.training)
if self.residual:
h = h + h_list[layer]
h_list.append(h)
if self.JK == "last":
node_representation = h_list[-1]
elif self.JK == "sum":
node_representation = 0
for layer in range(self.num_layer + 1):
node_representation += h_list[layer]
return node_representation
class vGINMolHeadEncoder(torch.nn.Module):
def __init__(self, num_layer, emb_dim):
super(vGINMolHeadEncoder, self).__init__()
self.num_layer = num_layer
self.emb_dim = emb_dim
self.atom_encoder = AtomEncoder(emb_dim)
self.conv1 = GINEConv(emb_dim)
self.convs = nn.ModuleList([GINEConv(emb_dim) for _ in range(num_layer - 1)])
self.relu1 = nn.ReLU()
self.relus = nn.ModuleList(
[nn.ReLU() for _ in range(num_layer - 1)]
)
self.batch_norm1 = nn.BatchNorm1d(emb_dim)
self.batch_norms = nn.ModuleList([
nn.BatchNorm1d(emb_dim)
for _ in range(num_layer - 1)
])
self.dropout1 = nn.Dropout()
self.dropouts = nn.ModuleList([
nn.Dropout() for _ in range(num_layer - 1)
])
self.virtual_node_embedding = nn.Embedding(1, emb_dim)
self.virtual_mlp = nn.Sequential(*(
[nn.Linear(emb_dim, 2 * emb_dim),
nn.BatchNorm1d(2 * emb_dim), nn.ReLU()] +
[nn.Linear(2 * emb_dim, emb_dim),
nn.BatchNorm1d(emb_dim), nn.ReLU(),
nn.Dropout()]
))
self.virtual_pool = global_add_pool
def forward(self, x, edge_index, edge_attr, batch):
virtual_node_feat = self.virtual_node_embedding(torch.zeros(batch[-1].item() + 1).to(edge_index.dtype).to(edge_index.device))
x = self.atom_encoder(x)
post_conv = self.dropout1(self.relu1(self.batch_norm1(self.conv1(x, edge_index, edge_attr))))
for i, (conv, batch_norm, relu, dropout) in enumerate(
zip(self.convs, self.batch_norms, self.relus, self.dropouts)):
# --- Add global info ---
post_conv = post_conv + virtual_node_feat[batch]
post_conv = batch_norm(conv(post_conv, edge_index, edge_attr))
if i < len(self.convs) - 1:
post_conv = relu(post_conv)
post_conv = dropout(post_conv)
# --- update global info ---
if i < len(self.convs) - 1:
virtual_node_feat = self.virtual_mlp(self.virtual_pool(post_conv, batch) + virtual_node_feat)
# out_readout = self.readout(post_conv, batch)
return post_conv