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
|
import os.path as osp
import torch
import torch.nn.functional as F
from ogb.nodeproppred import PygNodePropPredDataset
import torch_geometric.transforms as T
from torch_geometric.nn import MaskLabel, TransformerConv
from torch_geometric.utils import index_to_mask
root = osp.join(osp.dirname(osp.realpath(__file__)), '..', 'data', 'OGB')
dataset = PygNodePropPredDataset('ogbn-arxiv', root, T.ToUndirected())
class UniMP(torch.nn.Module):
def __init__(self, in_channels, num_classes, hidden_channels, num_layers,
heads, dropout=0.3):
super().__init__()
self.label_emb = MaskLabel(num_classes, in_channels)
self.convs = torch.nn.ModuleList()
self.norms = torch.nn.ModuleList()
for i in range(1, num_layers + 1):
if i < num_layers:
out_channels = hidden_channels // heads
concat = True
else:
out_channels = num_classes
concat = False
conv = TransformerConv(in_channels, out_channels, heads,
concat=concat, beta=True, dropout=dropout)
self.convs.append(conv)
in_channels = hidden_channels
if i < num_layers:
self.norms.append(torch.nn.LayerNorm(hidden_channels))
def forward(self, x, y, edge_index, label_mask):
x = self.label_emb(x, y, label_mask)
for conv, norm in zip(self.convs, self.norms):
x = norm(conv(x, edge_index)).relu()
return self.convs[-1](x, edge_index)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
data = dataset[0].to(device)
data.y = data.y.view(-1)
model = UniMP(dataset.num_features, dataset.num_classes, hidden_channels=64,
num_layers=3, heads=2).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0005)
split_idx = dataset.get_idx_split()
train_mask = index_to_mask(split_idx['train'], size=data.num_nodes)
val_mask = index_to_mask(split_idx['valid'], size=data.num_nodes)
test_mask = index_to_mask(split_idx['test'], size=data.num_nodes)
def train(label_rate=0.65): # How many labels to use for propagation.
model.train()
propagation_mask = MaskLabel.ratio_mask(train_mask, ratio=label_rate)
supervision_mask = train_mask ^ propagation_mask
optimizer.zero_grad()
out = model(data.x, data.y, data.edge_index, propagation_mask)
loss = F.cross_entropy(out[supervision_mask], data.y[supervision_mask])
loss.backward()
optimizer.step()
return float(loss)
@torch.no_grad()
def test():
model.eval()
propagation_mask = train_mask
out = model(data.x, data.y, data.edge_index, propagation_mask)
pred = out[val_mask].argmax(dim=-1)
val_acc = int((pred == data.y[val_mask]).sum()) / pred.size(0)
propagation_mask = train_mask | val_mask
out = model(data.x, data.y, data.edge_index, propagation_mask)
pred = out[test_mask].argmax(dim=-1)
test_acc = int((pred == data.y[test_mask]).sum()) / pred.size(0)
return val_acc, test_acc
for epoch in range(1, 501):
loss = train()
val_acc, test_acc = test()
print(f'Epoch: {epoch:03d}, Loss: {loss:.4f}, Val: {val_acc:.4f}, '
f'Test: {test_acc:.4f}')
|
