/
MVGRL_node.py
107 lines (87 loc) · 3.85 KB
/
MVGRL_node.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
import torch
import os.path as osp
import GCL.losses as L
import GCL.augmentors as A
import torch_geometric.transforms as T
from torch import nn
from tqdm import tqdm
from torch.optim import Adam
from GCL.eval import get_split, LREvaluator
from GCL.models import DualBranchContrast
from torch_geometric.nn import GCNConv
from torch_geometric.nn.inits import uniform
from torch_geometric.datasets import Planetoid
class GConv(nn.Module):
def __init__(self, input_dim, hidden_dim, num_layers):
super(GConv, self).__init__()
self.layers = torch.nn.ModuleList()
self.activation = nn.PReLU(hidden_dim)
for i in range(num_layers):
if i == 0:
self.layers.append(GCNConv(input_dim, hidden_dim))
else:
self.layers.append(GCNConv(hidden_dim, hidden_dim))
def forward(self, x, edge_index, edge_weight=None):
z = x
for conv in self.layers:
z = conv(z, edge_index, edge_weight)
z = self.activation(z)
return z
class Encoder(torch.nn.Module):
def __init__(self, encoder1, encoder2, augmentor, hidden_dim):
super(Encoder, self).__init__()
self.encoder1 = encoder1
self.encoder2 = encoder2
self.augmentor = augmentor
self.project = torch.nn.Linear(hidden_dim, hidden_dim)
uniform(hidden_dim, self.project.weight)
@staticmethod
def corruption(x, edge_index, edge_weight):
return x[torch.randperm(x.size(0))], edge_index, edge_weight
def forward(self, x, edge_index, edge_weight=None):
aug1, aug2 = self.augmentor
x1, edge_index1, edge_weight1 = aug1(x, edge_index, edge_weight)
x2, edge_index2, edge_weight2 = aug2(x, edge_index, edge_weight)
z1 = self.encoder1(x1, edge_index1, edge_weight1)
z2 = self.encoder2(x2, edge_index2, edge_weight2)
g1 = self.project(torch.sigmoid(z1.mean(dim=0, keepdim=True)))
g2 = self.project(torch.sigmoid(z2.mean(dim=0, keepdim=True)))
z1n = self.encoder1(*self.corruption(x1, edge_index1, edge_weight1))
z2n = self.encoder2(*self.corruption(x2, edge_index2, edge_weight2))
return z1, z2, g1, g2, z1n, z2n
def train(encoder_model, contrast_model, data, optimizer):
encoder_model.train()
optimizer.zero_grad()
z1, z2, g1, g2, z1n, z2n = encoder_model(data.x, data.edge_index)
loss = contrast_model(h1=z1, h2=z2, g1=g1, g2=g2, h1n=z1n, h2n=z2n)
loss.backward()
optimizer.step()
return loss.item()
def test(encoder_model, data):
encoder_model.eval()
z1, z2, _, _, _, _ = encoder_model(data.x, data.edge_index)
z = z1 + z2
split = get_split(num_samples=z.size()[0], train_ratio=0.1, test_ratio=0.8)
result = LREvaluator()(z, data.y, split)
return result
def main():
device = torch.device('cuda')
path = osp.join(osp.expanduser('~'), 'datasets')
dataset = Planetoid(path, name='Cora', transform=T.NormalizeFeatures())
data = dataset[0].to(device)
aug1 = A.Identity()
aug2 = A.PPRDiffusion(alpha=0.2)
gconv1 = GConv(input_dim=dataset.num_features, hidden_dim=512, num_layers=2).to(device)
gconv2 = GConv(input_dim=dataset.num_features, hidden_dim=512, num_layers=2).to(device)
encoder_model = Encoder(encoder1=gconv1, encoder2=gconv2, augmentor=(aug1, aug2), hidden_dim=512).to(device)
contrast_model = DualBranchContrast(loss=L.JSD(), mode='G2L').to(device)
optimizer = Adam(encoder_model.parameters(), lr=0.001)
with tqdm(total=200, desc='(T)') as pbar:
for epoch in range(1, 201):
loss = train(encoder_model, contrast_model, data, optimizer)
pbar.set_postfix({'loss': loss})
pbar.update()
test_result = test(encoder_model, data)
print(f'(E): Best test F1Mi={test_result["micro_f1"]:.4f}, F1Ma={test_result["macro_f1"]:.4f}')
if __name__ == '__main__':
main()