-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathgnn_model.py
308 lines (222 loc) · 11.3 KB
/
gnn_model.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
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
import torch
from torch_geometric.nn import MessagePassing
from torch_geometric.utils import add_self_loops, degree, softmax,remove_self_loops
from torch_geometric.nn import global_add_pool, global_mean_pool, global_max_pool, GlobalAttention, Set2Set
import torch.nn.functional as F
from torch_scatter import scatter_add
from torch_geometric.nn.inits import glorot, zeros
import numpy as np
import torch.nn as nn
from torch.nn import Parameter
num_atom_type = 121 #including the extra motif tokens and graph token
num_chirality_tag = 11 #degree
num_bond_type = 7 #including aromatic and self-loop edge, and extra motif tokens and graph token
num_bond_direction = 3
class GINConv(MessagePassing):
def __init__(self, emb_dim, aggr = "add"):
super(GINConv, self).__init__()
#multi-layer perceptron
self.mlp = torch.nn.Sequential(torch.nn.Linear(emb_dim, 2*emb_dim), torch.nn.ReLU(), torch.nn.Linear(2*emb_dim, emb_dim))
self.edge_embedding1 = torch.nn.Embedding(num_bond_type, emb_dim)
self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, emb_dim)
torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data)
torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data)
self.aggr = aggr
def forward(self, x, edge_index, edge_attr):
#add self loops in the edge space
edge_index = add_self_loops(edge_index, num_nodes = x.size(0))
#add features corresponding to self-loop edges.
self_loop_attr = torch.zeros(x.size(0), 2)
self_loop_attr[:,0] = 4
self_loop_attr = self_loop_attr.to(edge_attr.device).to(edge_attr.dtype)
edge_attr = torch.cat((edge_attr, self_loop_attr), dim = 0)
edge_embeddings = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1])
return self.propagate(edge_index[0], x=x, edge_attr=edge_embeddings)
def message(self, x_j, edge_attr):
return x_j + edge_attr
def update(self, aggr_out):
return self.mlp(aggr_out)
class GCNConv(MessagePassing):
def __init__(self, emb_dim, aggr = "add"):
super(GCNConv, self).__init__()
self.emb_dim = emb_dim
self.linear = torch.nn.Linear(emb_dim, emb_dim)
self.edge_embedding1 = torch.nn.Embedding(num_bond_type, emb_dim)
self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, emb_dim)
torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data)
torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data)
self.aggr = aggr
def norm(self, edge_index, num_nodes, dtype):
### assuming that self-loops have been already added in edge_index
edge_weight = torch.ones((edge_index.size(1), ), dtype=dtype,
device=edge_index.device)
row, col = edge_index
deg = scatter_add(edge_weight, row, dim=0, dim_size=num_nodes)
deg_inv_sqrt = deg.pow(-0.5)
deg_inv_sqrt[deg_inv_sqrt == float('inf')] = 0
return deg_inv_sqrt[row] * edge_weight * deg_inv_sqrt[col]
def forward(self, x, edge_index, edge_attr):
#add self loops in the edge space
edge_index, _ = add_self_loops(edge_index, num_nodes = x.size(0))
#add features corresponding to self-loop edges.
self_loop_attr = torch.zeros(x.size(0), 2)
self_loop_attr[:,0] = 4 #bond type for self-loop edge
self_loop_attr = self_loop_attr.to(edge_attr.device).to(edge_attr.dtype)
edge_attr = torch.cat((edge_attr, self_loop_attr), dim = 0)
edge_embeddings = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1])
norm = self.norm(edge_index, x.size(0), x.dtype)
x = self.linear(x)
return self.propagate(edge_index, x=x, edge_attr=edge_embeddings, norm=norm)
def message(self, x_j, edge_attr, norm):
return norm.view(-1, 1) * (x_j + edge_attr)
class GATConv(MessagePassing):
def __init__(self,
emb_dim,
heads=2,
negative_slope=0.2,
dropout=0.,
bias=True):
super(GATConv, self).__init__(node_dim=0, aggr='add') # "Add" aggregation.
self.in_channels = emb_dim
self.out_channels = emb_dim
self.edge_dim = emb_dim # new
self.heads = heads
self.negative_slope = negative_slope
self.dropout = dropout
self.edge_embedding1 = torch.nn.Embedding(num_bond_type, emb_dim)
self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, emb_dim)
torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data)
torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data)
self.weight = Parameter(torch.Tensor(emb_dim, heads * emb_dim)) # emb(in) x [H*emb(out)]
self.att = Parameter(torch.Tensor(1, heads, 2 * emb_dim + self.edge_dim)) # 1 x H x [2*emb(out)+edge_dim] # new
self.edge_update = Parameter(torch.Tensor(emb_dim + self.edge_dim, emb_dim)) # [emb(out)+edge_dim] x emb(out) # new
if bias:
self.bias = Parameter(torch.Tensor(emb_dim))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self):
glorot(self.weight)
glorot(self.att)
glorot(self.edge_update) # new
zeros(self.bias)
def forward(self, x, edge_index, edge_attr, size=None):
edge_attr = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1])
x = torch.mm(x, self.weight).view(-1, self.heads, self.out_channels)
if size is None and torch.is_tensor(x):
edge_index, _ = remove_self_loops(edge_index)
edge_index, _ = add_self_loops(edge_index, num_nodes=x.size(0))
self_loop_edges = torch.zeros(x.size(0), edge_attr.size(1)).to(edge_index.device)
edge_attr = torch.cat([edge_attr, self_loop_edges], dim=0)
return self.propagate(edge_index, x=x, edge_attr=edge_attr, size=size)
def message(self, x_i, x_j, size_i, edge_index_i, edge_attr):
edge_attr = edge_attr.unsqueeze(1).repeat(1, self.heads, 1)
x_j = torch.cat([x_j, edge_attr], dim=-1)
x_i = x_i.view(-1, self.heads, self.out_channels)
alpha = (torch.cat([x_i, x_j], dim=-1) * self.att).sum(dim=-1)
alpha = F.leaky_relu(alpha, self.negative_slope)
alpha = softmax(alpha, edge_index_i, num_nodes=size_i)
if self.training and self.dropout > 0:
alpha = F.dropout(alpha, p=self.dropout, training=True)
return x_j * alpha.view(-1, self.heads, 1)
def update(self, aggr_out):
aggr_out = aggr_out.mean(dim=1)
aggr_out = torch.mm(aggr_out, self.edge_update)
if self.bias is not None:
aggr_out = aggr_out + self.bias
return aggr_out
class GraphSAGEConv(MessagePassing):
def __init__(self, emb_dim, aggr = "mean"):
super(GraphSAGEConv, self).__init__()
self.emb_dim = emb_dim
self.linear = torch.nn.Linear(emb_dim, emb_dim)
self.edge_embedding1 = torch.nn.Embedding(num_bond_type, emb_dim)
self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, emb_dim)
torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data)
torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data)
self.aggr = aggr
def forward(self, x, edge_index, edge_attr):
#add self loops in the edge space
edge_index = add_self_loops(edge_index, num_nodes = x.size(0))
#add features corresponding to self-loop edges.
self_loop_attr = torch.zeros(x.size(0), 2)
self_loop_attr[:,0] = 4
self_loop_attr = self_loop_attr.to(edge_attr.device).to(edge_attr.dtype)
edge_attr = torch.cat((edge_attr, self_loop_attr), dim = 0)
edge_embeddings = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1])
x = self.linear(x)
return self.propagate(edge_index[0], x=x, edge_attr=edge_embeddings)
def message(self, x_j, edge_attr):
return x_j + edge_attr
def update(self, aggr_out):
return F.normalize(aggr_out, p = 2, dim = -1)
class GNN(torch.nn.Module):
"""
Args:
num_layer (int): the number of GNN layers
emb_dim (int): dimensionality of embeddings
JK (str): last, concat, max or sum.
max_pool_layer (int): the layer from which we use max pool rather than add pool for neighbor aggregation
drop_ratio (float): dropout rate
gnn_type: gin, gcn, graphsage, gat
Output:
node representations
"""
def __init__(self, num_layer, emb_dim, JK = "last", drop_ratio = 0, gnn_type = "gin"):
super(GNN, self).__init__()
self.num_layer = num_layer
self.drop_ratio = drop_ratio
self.JK = JK
if self.num_layer < 2:
raise ValueError("Number of GNN layers must be greater than 1.")
self.x_embedding1 = torch.nn.Embedding(num_atom_type, emb_dim)
self.x_embedding2 = torch.nn.Embedding(num_chirality_tag, emb_dim)
torch.nn.init.xavier_uniform_(self.x_embedding1.weight.data)
torch.nn.init.xavier_uniform_(self.x_embedding2.weight.data)
###List of MLPs
self.gnns = torch.nn.ModuleList()
for layer in range(num_layer):
if gnn_type == "gin":
self.gnns.append(GINConv(emb_dim, aggr = "add"))
elif gnn_type == "gcn":
self.gnns.append(GCNConv(emb_dim))
elif gnn_type == "gat":
self.gnns.append(GATConv(emb_dim))
elif gnn_type == "graphsage":
self.gnns.append(GraphSAGEConv(emb_dim))
# ###List of batchnorms
self.batch_norms = torch.nn.ModuleList()
for layer in range(num_layer):
self.batch_norms.append(torch.nn.BatchNorm1d(emb_dim))
def forward(self, *argv):
if len(argv) == 3:
x, edge_index, edge_attr = argv[0], argv[1], argv[2]
elif len(argv) == 1:
data = argv[0]
x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
else:
raise ValueError("unmatched number of arguments.")
x = self.x_embedding1(x[:,0]) + self.x_embedding2(x[:,1])
h_list = [x]
for layer in range(self.num_layer):
h = self.gnns[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.relu(h), self.drop_ratio, training = self.training)
h_list.append(h)
### Different implementations of Jk-concat
if self.JK == "concat":
node_representation = torch.cat(h_list, dim = 1)
elif self.JK == "last":
node_representation = h_list[-1]
elif self.JK == "max":
h_list = [h.unsqueeze_(0) for h in h_list]
node_representation = torch.max(torch.cat(h_list, dim = 0), dim = 0)[0]
elif self.JK == "sum":
h_list = [h.unsqueeze_(0) for h in h_list]
node_representation = torch.sum(torch.cat(h_list, dim = 0), dim = 0)[0]
return node_representation
if __name__ == "__main__":
pass