-
Notifications
You must be signed in to change notification settings - Fork 154
/
model.py
203 lines (158 loc) · 7 KB
/
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
import __init__
import torch
from gcn_lib.sparse.torch_vertex import GENConv
from gcn_lib.sparse.torch_nn import norm_layer
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
import logging
class DeeperGCN(torch.nn.Module):
def __init__(self, args):
super(DeeperGCN, self).__init__()
self.num_layers = args.num_layers
self.dropout = args.dropout
self.block = args.block
self.checkpoint_grad = False
hidden_channels = args.hidden_channels
num_tasks = args.num_tasks
conv = args.conv
aggr = args.gcn_aggr
t = args.t
self.learn_t = args.learn_t
p = args.p
self.learn_p = args.learn_p
y = args.y
self.learn_y = args.learn_y
self.msg_norm = args.msg_norm
learn_msg_scale = args.learn_msg_scale
conv_encode_edge = args.conv_encode_edge
norm = args.norm
mlp_layers = args.mlp_layers
node_features_file_path = args.nf_path
self.use_one_hot_encoding = args.use_one_hot_encoding
# save gpu mem using gradient ckpt
if aggr not in ['add', 'max', 'mean'] and self.num_layers > 15:
self.checkpoint_grad = True
self.ckp_k = 9
print('The number of layers {}'.format(self.num_layers),
'Aggregation method {}'.format(aggr),
'block: {}'.format(self.block))
if self.block == 'res+':
print('LN/BN->ReLU->GraphConv->Res')
elif self.block == 'res':
print('GraphConv->LN/BN->ReLU->Res')
elif self.block == 'dense':
raise NotImplementedError('To be implemented')
elif self.block == "plain":
print('GraphConv->LN/BN->ReLU')
else:
raise Exception('Unknown block Type')
self.gcns = torch.nn.ModuleList()
self.layer_norms = torch.nn.ModuleList()
for layer in range(self.num_layers):
if conv == 'gen':
gcn = GENConv(hidden_channels, hidden_channels,
aggr=aggr,
t=t, learn_t=self.learn_t,
p=p, learn_p=self.learn_p,
y=y, learn_y=self.learn_y,
msg_norm=self.msg_norm, learn_msg_scale=learn_msg_scale,
encode_edge=conv_encode_edge, edge_feat_dim=hidden_channels,
norm=norm, mlp_layers=mlp_layers)
else:
raise Exception('Unknown Conv Type')
self.gcns.append(gcn)
self.layer_norms.append(norm_layer(norm, hidden_channels))
self.node_features = torch.load(node_features_file_path).to(args.device)
if self.use_one_hot_encoding:
self.node_one_hot_encoder = torch.nn.Linear(8, 8)
self.node_features_encoder = torch.nn.Linear(8 * 2, hidden_channels)
else:
self.node_features_encoder = torch.nn.Linear(8, hidden_channels)
self.edge_encoder = torch.nn.Linear(8, hidden_channels)
self.node_pred_linear = torch.nn.Linear(hidden_channels, num_tasks)
def forward(self, x, node_index, edge_index, edge_attr):
node_features_1st = self.node_features[node_index]
if self.use_one_hot_encoding:
node_features_2nd = self.node_one_hot_encoder(x)
# concatenate
node_features = torch.cat((node_features_1st, node_features_2nd), dim=1)
else:
node_features = node_features_1st
h = self.node_features_encoder(node_features)
edge_emb = self.edge_encoder(edge_attr)
if self.block == 'res+':
h = self.gcns[0](h, edge_index, edge_emb)
if self.checkpoint_grad:
for layer in range(1, self.num_layers):
h1 = self.layer_norms[layer-1](h)
h2 = F.relu(h1)
h2 = F.dropout(h2, p=self.dropout, training=self.training)
if layer % self.ckp_k != 0:
res = checkpoint(self.gcns[layer], h2, edge_index, edge_emb)
h = res + h
else:
h = self.gcns[layer](h2, edge_index, edge_emb) + h
else:
for layer in range(1, self.num_layers):
h1 = self.layer_norms[layer-1](h)
h2 = F.relu(h1)
h2 = F.dropout(h2, p=self.dropout, training=self.training)
h = self.gcns[layer](h2, edge_index, edge_emb) + h
h = F.relu(self.layer_norms[self.num_layers-1](h))
h = F.dropout(h, p=self.dropout, training=self.training)
return self.node_pred_linear(h)
elif self.block == 'res':
h = F.relu(self.layer_norms[0](self.gcns[0](h, edge_index, edge_emb)))
h = F.dropout(h, p=self.dropout, training=self.training)
for layer in range(1, self.num_layers):
h1 = self.gcns[layer](h, edge_index, edge_emb)
h2 = self.layer_norms[layer](h1)
h = F.relu(h2) + h
h = F.dropout(h, p=self.dropout, training=self.training)
return self.node_pred_linear(h)
elif self.block == 'dense':
raise NotImplementedError('To be implemented')
elif self.block == 'plain':
h = F.relu(self.layer_norms[0](self.gcns[0](h, edge_index, edge_emb)))
h = F.dropout(h, p=self.dropout, training=self.training)
for layer in range(1, self.num_layers):
h1 = self.gcns[layer](h, edge_index, edge_emb)
h2 = self.layer_norms[layer](h1)
h = F.relu(h2)
h = F.dropout(h, p=self.dropout, training=self.training)
return self.node_pred_linear(h)
else:
raise Exception('Unknown block Type')
def print_params(self, epoch=None, final=False):
if self.learn_t:
ts = []
for gcn in self.gcns:
ts.append(gcn.t.item())
if final:
print('Final t {}'.format(ts))
else:
logging.info('Epoch {}, t {}'.format(epoch, ts))
if self.learn_p:
ps = []
for gcn in self.gcns:
ps.append(gcn.p.item())
if final:
print('Final p {}'.format(ps))
else:
logging.info('Epoch {}, p {}'.format(epoch, ps))
if self.learn_y:
ys = []
for gcn in self.gcns:
ys.append(gcn.sigmoid_y.item())
if final:
print('Final sigmoid(y) {}'.format(ys))
else:
logging.info('Epoch {}, sigmoid(y) {}'.format(epoch, ys))
if self.msg_norm:
ss = []
for gcn in self.gcns:
ss.append(gcn.msg_norm.msg_scale.item())
if final:
print('Final s {}'.format(ss))
else:
logging.info('Epoch {}, s {}'.format(epoch, ss))