/
GBT.py
97 lines (77 loc) · 3.31 KB
/
GBT.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
import torch
import os.path as osp
import GCL.losses as L
import GCL.augmentors as A
import torch_geometric.transforms as T
from tqdm import tqdm
from torch.optim import Adam
from GCL.eval import get_split, LREvaluator
from GCL.models.contrast_model import WithinEmbedContrast
from torch_geometric.nn import GCNConv
from torch_geometric.datasets import WikiCS
from pl_bolts.optimizers import LinearWarmupCosineAnnealingLR
class GConv(torch.nn.Module):
def __init__(self, input_dim, hidden_dim):
super(GConv, self).__init__()
self.act = torch.nn.PReLU()
self.bn = torch.nn.BatchNorm1d(2 * hidden_dim, momentum=0.01)
self.conv1 = GCNConv(input_dim, 2 * hidden_dim, cached=False)
self.conv2 = GCNConv(2 * hidden_dim, hidden_dim, cached=False)
def forward(self, x, edge_index, edge_weight=None):
z = self.conv1(x, edge_index, edge_weight)
z = self.bn(z)
z = self.act(z)
z = self.conv2(z, edge_index, edge_weight)
return z
class Encoder(torch.nn.Module):
def __init__(self, encoder, augmentor):
super(Encoder, self).__init__()
self.encoder = encoder
self.augmentor = augmentor
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)
z = self.encoder(x, edge_index, edge_weight)
z1 = self.encoder(x1, edge_index1, edge_weight1)
z2 = self.encoder(x2, edge_index2, edge_weight2)
return z, z1, z2
def train(encoder_model, contrast_model, data, optimizer):
encoder_model.train()
optimizer.zero_grad()
_, z1, z2 = encoder_model(data.x, data.edge_index, data.edge_attr)
loss = contrast_model(z1, z2)
loss.backward()
optimizer.step()
return loss.item()
def test(encoder_model, data):
encoder_model.eval()
z, _, _ = encoder_model(data.x, data.edge_index, data.edge_attr)
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', 'WikiCS')
dataset = WikiCS(path, transform=T.NormalizeFeatures())
data = dataset[0].to(device)
aug1 = A.Compose([A.EdgeRemoving(pe=0.5), A.FeatureMasking(pf=0.1)])
aug2 = A.Compose([A.EdgeRemoving(pe=0.5), A.FeatureMasking(pf=0.1)])
gconv = GConv(input_dim=dataset.num_features, hidden_dim=256).to(device)
encoder_model = Encoder(encoder=gconv, augmentor=(aug1, aug2)).to(device)
contrast_model = WithinEmbedContrast(loss=L.BarlowTwins()).to(device)
optimizer = Adam(encoder_model.parameters(), lr=5e-4)
scheduler = LinearWarmupCosineAnnealingLR(
optimizer=optimizer,
warmup_epochs=400,
max_epochs=4000)
with tqdm(total=4000, desc='(T)') as pbar:
for epoch in range(1, 4001):
loss = train(encoder_model, contrast_model, data, optimizer)
scheduler.step()
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()