-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
153 lines (131 loc) · 5.81 KB
/
main.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
import sys, argparse
import numpy as np
from tqdm import tqdm
import torch
import torch.nn.functional as F
from torch_geometric import seed_everything
from models import *
from utils import *
from loader import loader
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, default='Cora',
choices=['Cora', 'Citeseer', 'Pubmed', 'DBLP', 'Computers', 'Photo', 'CS', 'Physics',
'Chameleon', 'Squirrel', 'Texas', 'Wisconsin', 'Cornell', 'Film', 'Penn94', 'Cornell5'],
help="The dataset to be used.")
parser.add_argument('--seed', type=int, default=1234,
help="random seed")
parser.add_argument('--alt', type=str, default='adj',
choices=['adj', 'adj_norm'],
help="Decompose the original adj or the normalized adj.")
parser.add_argument('--backbone', type=str, default='APPNP',
choices=['APPNP'],
help="The backbone GNN works with the ALT.")
parser.add_argument('--K', type=int, default=5,
help="Propagation step of APPNP")
parser.add_argument('--alpha', type=float, default=0.1,
help="Restart prob of APPNP")
parser.add_argument('--hidden_unit', type=int, default=32,
help="Hidden dim")
parser.add_argument('--amp', type=float, default=0.5,
help="The parameter of the high-pass filters in the augmenter.")
parser.add_argument('--wd', type=float, default=5e-5,
help="Weight decay.")
parser.add_argument('--lr', type=float, default=0.01,
help="Learning rate.")
parser.add_argument('--MLPA', type=float, default=0.001,
help="The ratio of signals from the MLP(A) part.")
parser.add_argument('--device', type=str, default='3',
choices=['cpu', '0', '1', '2', '3'],
help="The GPU device to be used.")
parser.add_argument('--run', type=int, default=10,
help="The # of test runs.")
args = parser.parse_args()
if args.device == 'cpu':
device = 'cpu'
else:
device = 'cuda:' + args.device
torch.autograd.set_detect_anomaly(True)
seed = args.seed
seed_everything(seed)
def train(model, augmenter, optimizer, optimizer_aug, data):
model.train()
augmenter.train()
optimizer.zero_grad()
optimizer_aug.zero_grad()
edge_weight = augmenter(data)
out = model(data, edge_weight)
loss = F.nll_loss(out[data.train_idx], data.y[data.train_idx])
loss.backward()
optimizer.step()
optimizer_aug.step()
@torch.no_grad()
def test(model, augmenter, data):
model.eval()
augmenter.eval()
edge_weight = augmenter(data)
out, accs = model(data, edge_weight), []
for _, idx in data('train_idx', 'val_idx', 'test_idx'):
acc = accuracy(out[idx], data.y[idx])
accs.append(acc.cpu())
return accs
""" Experimental settings """
hidden_unit = args.hidden_unit
""" Load dataset """
data = loader(args.dataset)
if args.dataset in ['Penn94', 'Cornell5']:
all_idx = torch.nonzero(data.y != -1).flatten().tolist()
num_nodes = len(all_idx)
else:
num_nodes = data.y.shape[0]
all_idx = np.array(range(num_nodes))
num_features = data.x.shape[1]
num_classes = data.y.max().item()+1
data.A = torch_geometric.utils.to_dense_adj(data.edge_index)[0]
data = data.to(device)
print("Device: {}".format(args.device))
print("ALT: {}".format(args.alt))
print("Backbone GNN: {}".format(args.backbone))
print("Propagation Steps: {}".format(args.K))
print("Dataset: {}".format(args.dataset))
print("amp: {}".format(args.amp))
print("wd: {}".format(args.wd))
print("lr: {}".format(args.lr))
print("hidden_unit: {}".format(hidden_unit))
print("MLP_A: {}".format(args.MLPA))
final_tests = []
dataset_split = [0.2, 0.2, 0.6]
for i in range(args.run):
print("Run: {}".format(i))
""" For 'Cora', 'Citeseer', 'Pubmed' we use their default split. For the others, we randomly split them. """
if args.dataset not in ['Cora', 'Citeseer', 'Pubmed']:
np.random.shuffle(all_idx)
data.train_idx = torch.LongTensor(all_idx[:int(num_nodes*dataset_split[0])])
data.val_idx = torch.LongTensor(all_idx[int(num_nodes*dataset_split[0]):int(num_nodes*dataset_split[0])+int(num_nodes*dataset_split[1])])
data.test_idx = torch.LongTensor(all_idx[int(num_nodes*dataset_split[0])+int(num_nodes*dataset_split[1]):])
else:
data.train_idx = data.train_mask.nonzero(as_tuple=True)[0]
data.val_idx = data.val_mask.nonzero(as_tuple=True)[0]
data.test_idx = data.test_mask.nonzero(as_tuple=True)[0]
""" To decompose the original adj or normalized adj """
if args.alt == 'adj':
model = DualGNN(num_features, hidden_unit, num_classes, args.backbone, MLPA=args.MLPA, A_dim=data.A.shape[0]).to(device)
decompose_norm = False
elif args.alt == 'adj_norm':
model = DualGNN_norm(num_features, hidden_unit, num_classes, args.backbone, MLPA=args.MLPA, A_dim=data.A.shape[0]).to(device)
decompose_norm = True
augmenter = Augmenter(num_features, hidden_unit, amp=args.amp, decompose_norm=decompose_norm).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.wd)
optimizer_aug = torch.optim.Adam(augmenter.parameters(), lr=args.lr, weight_decay=args.wd)
best_test_acc = 0
best_val_acc = 0
for epoch in tqdm(range(1, 201)):
train(model, augmenter, optimizer, optimizer_aug, data)
if epoch % 10 == 0:
train_acc, val_acc, test_acc = test(model, augmenter, data)
if val_acc > best_val_acc:
best_val_acc = val_acc
best_test_acc = test_acc
final_tests.append(best_test_acc)
print("Mean: {:.4f}, Std: {:.4f}".format(np.array(final_tests).mean(axis=0),
np.array(final_tests).std(axis=0)))
print()