/
PF-GNNIR-triangles.py
134 lines (118 loc) · 6.33 KB
/
PF-GNNIR-triangles.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
import os.path as osp
import time
import argparse
import torch
from torch.nn import Linear
from torch_geometric.data import DataLoader, DataListLoader
from torch_geometric.nn import DataParallel
from torch_geometric.datasets import TUDataset
import torch_geometric.transforms as T
from torch_geometric.nn.inits import *
from torch_geometric.utils import degree
from worker import train, test
from model import PFGNN_Net
from gnn_models import GNN, GNN_TRIANGLES
import torch.nn.functional as F
import numpy as np
import networkx as nx
np.set_printoptions(precision=5, suppress=True,linewidth=np.inf)
def main():
# Training settings
parser = argparse.ArgumentParser(description='PF-GNNIR')
parser.add_argument('--device', type=int, default=0,
help='which gpu to use if any (default: 0)')
parser.add_argument('--batch_size', type=int, default=60,
help='input batch size for training (default: 60)')
parser.add_argument('--epochs', type=int, default=300,
help='number of epochs to train (default: 300)')
parser.add_argument('--lr', type=float, default=0.001,
help='learning rate (default: 0.001)')
parser.add_argument('--factor', type=float, default=0.7)
parser.add_argument('--patience', type=float, default=20)
parser.add_argument('--min_lr', type=float, default=0.000001)
# parser.add_argument('--num_split', type=int, default=5,
# help='number of splits.')
parser.add_argument('--dim', type=int, default=150,
help='hidden dim. (default=150)')
parser.add_argument('--task_name', type=str, default="ZINC")
parser.add_argument('--parallel', type=bool, default=True,
help='run on multiple gpus (default: True)')
parser.add_argument('--depth', type=int, default=2,
help='number of IR iterations (default: 2)')
parser.add_argument('--num_particles', type=int, default=4,
help='number of particles of particle filter (default: 4)')
args = parser.parse_args()
class HandleNodeAttention(object):
def __call__(self, data):
data.attn = torch.softmax(data.x, dim=0).flatten()
data.x = None
return data
transform = T.Compose([HandleNodeAttention(), T.OneHotDegree(max_degree=14)])
path = osp.join(osp.dirname(osp.realpath(__file__)), 'data', 'TRIANGLES')
dataset = TUDataset(path, name='TRIANGLES', use_node_attr=True,
transform=transform)
train_dataset = dataset[:30000]
val_dataset = dataset[30000:35000]
test_dataset = dataset[35000:]
deg = torch.zeros(14, dtype=torch.long)
for data in train_dataset:
d = degree(data.edge_index[1], num_nodes=data.num_nodes, dtype=torch.long)
deg += torch.bincount(d, minlength=deg.numel())
class Net(torch.nn.Module):
def __init__(self, node_infeat, outdim, dim, depth, num_particles, gnn=GNN_TRIANGLES, deg=None):
super(Net, self).__init__()
self.dim = dim
self.linear = Linear(node_infeat, dim)
self.pfgnn = PFGNN_Net(outdim=outdim, dim=dim, depth=depth, num_particles=num_particles, gnn=gnn, deg=deg)
def reset_parameters(self):
self.linear.reset_parameters()
self.pfgnn.reset_parameters()
def forward(self, data):
out = F.relu(self.linear(data.x))
edge_attr = None
out, log_probs, batch_size = self.pfgnn(node_emb=out, edge_emb=edge_attr, data=data)
return out, log_probs, batch_size
device = torch.device("cuda:" + str(args.device)) if torch.cuda.is_available() else torch.device("cpu")
model = Net(node_infeat=dataset.num_features, outdim=int(dataset.num_classes+1),
dim=args.dim, depth=args.depth, num_particles=args.num_particles, gnn=GNN_TRIANGLES, deg=deg)
if args.parallel is True:
loader = DataListLoader
model = DataParallel(model).to(device)
else:
loader = DataLoader
model = model.to(device)
train_loader = loader(train_dataset, batch_size=args.batch_size, shuffle=True)
val_loader = loader(val_dataset, batch_size=args.batch_size)
test_loader = loader(test_dataset, batch_size=args.batch_size//2)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=3e-6)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min',
factor=args.factor, patience=args.patience,
min_lr=args.min_lr)
print("Total parameters: ", sum(p.numel() for p in model.parameters()))
torch.autograd.set_detect_anomaly(True)
start = None
best_val_acc = None
for epoch in range(1, args.epochs):
lr = scheduler.optimizer.param_groups[0]['lr']
start = time.time()
loss, l1, l2 = train(model=model, loader=train_loader, optimizer=optimizer,
device=device, parallel=args.parallel, regression=False)
train_correct, train_loss = test(model=model, loader=train_loader, device=device, parallel=args.parallel, regression=False)
val_correct, val_loss = test(model=model, loader=val_loader, device=device, parallel=args.parallel, regression=False)
test_correct, test_loss = test(model=model, loader=test_loader, device=device, parallel=args.parallel, regression=False)
scheduler.step(val_loss)
train_acc = train_correct.sum().item() / train_correct.size(0)
val_acc = val_correct.sum().item() / val_correct.size(0)
test_acc1 = test_correct[:5000].sum().item() / 5000
test_acc2 = test_correct[5000:].sum().item() / 5000
if best_val_acc is None or val_acc >= best_val_acc:
best_val_acc = val_acc
best_test_acc_S = test_acc1
best_test_acc_L = test_acc2
end = time.time()
print(('Epoch: {:03d}, Time: {:.3f}, LR: {:.5f}, Loss: {:.4f}, ValLoss: {:.4f}, TrainAcc: {:.3f}, ValAcc: {:.3f}, '
'TestAcc Orig : {:.3f}, TestAcc Large: {:.3f}, Best Orig : {:.3f}, Best Large: {:.3f}').format(
epoch, end-start, lr, loss, val_loss, train_acc,
val_acc, test_acc1, test_acc2, best_test_acc_S, best_test_acc_L))
if __name__ == "__main__":
main()