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
|
import argparse
import os.path as osp
import torch
import torch.nn.functional as F
from torch_geometric.datasets import Flickr
from torch_geometric.loader import GraphSAINTRandomWalkSampler
from torch_geometric.nn import GraphConv
from torch_geometric.typing import WITH_TORCH_SPARSE
from torch_geometric.utils import degree
if not WITH_TORCH_SPARSE:
quit("This example requires 'torch-sparse'")
path = osp.join(osp.dirname(osp.realpath(__file__)), '..', 'data', 'Flickr')
dataset = Flickr(path)
data = dataset[0]
row, col = data.edge_index
data.edge_weight = 1. / degree(col, data.num_nodes)[col] # Norm by in-degree.
parser = argparse.ArgumentParser()
parser.add_argument('--use_normalization', action='store_true')
args = parser.parse_args()
loader = GraphSAINTRandomWalkSampler(data, batch_size=6000, walk_length=2,
num_steps=5, sample_coverage=100,
save_dir=dataset.processed_dir,
num_workers=4)
class Net(torch.nn.Module):
def __init__(self, hidden_channels):
super().__init__()
in_channels = dataset.num_node_features
out_channels = dataset.num_classes
self.conv1 = GraphConv(in_channels, hidden_channels)
self.conv2 = GraphConv(hidden_channels, hidden_channels)
self.conv3 = GraphConv(hidden_channels, hidden_channels)
self.lin = torch.nn.Linear(3 * hidden_channels, out_channels)
def set_aggr(self, aggr):
self.conv1.aggr = aggr
self.conv2.aggr = aggr
self.conv3.aggr = aggr
def forward(self, x0, edge_index, edge_weight=None):
x1 = F.relu(self.conv1(x0, edge_index, edge_weight))
x1 = F.dropout(x1, p=0.2, training=self.training)
x2 = F.relu(self.conv2(x1, edge_index, edge_weight))
x2 = F.dropout(x2, p=0.2, training=self.training)
x3 = F.relu(self.conv3(x2, edge_index, edge_weight))
x3 = F.dropout(x3, p=0.2, training=self.training)
x = torch.cat([x1, x2, x3], dim=-1)
x = self.lin(x)
return x.log_softmax(dim=-1)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = Net(hidden_channels=256).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
def train():
model.train()
model.set_aggr('add' if args.use_normalization else 'mean')
total_loss = total_examples = 0
for data in loader:
data = data.to(device)
optimizer.zero_grad()
if args.use_normalization:
edge_weight = data.edge_norm * data.edge_weight
out = model(data.x, data.edge_index, edge_weight)
loss = F.nll_loss(out, data.y, reduction='none')
loss = (loss * data.node_norm)[data.train_mask].sum()
else:
out = model(data.x, data.edge_index)
loss = F.nll_loss(out[data.train_mask], data.y[data.train_mask])
loss.backward()
optimizer.step()
total_loss += loss.item() * data.num_nodes
total_examples += data.num_nodes
return total_loss / total_examples
@torch.no_grad()
def test():
model.eval()
model.set_aggr('mean')
out = model(data.x.to(device), data.edge_index.to(device))
pred = out.argmax(dim=-1)
correct = pred.eq(data.y.to(device))
accs = []
for _, mask in data('train_mask', 'val_mask', 'test_mask'):
accs.append(correct[mask].sum().item() / mask.sum().item())
return accs
for epoch in range(1, 51):
loss = train()
accs = test()
print(f'Epoch: {epoch:02d}, Loss: {loss:.4f}, Train: {accs[0]:.4f}, '
f'Val: {accs[1]:.4f}, Test: {accs[2]:.4f}')
|
