-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_opv.py
171 lines (146 loc) · 7.44 KB
/
train_opv.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
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
import time
import argparse
import torch
import torch.nn as nn
import torch_geometric.transforms as T
from torch_geometric.loader import DataLoader
from models import MHNN, GNN_2D, MHNNS
from datasets import OPVHGraph, OPVGraph, OneTarget
from utils import Logger, seed_everything
@torch.no_grad()
def evaluate(args, model, loader, std=None):
model.eval()
err = 0.0
# for MAE
for batch in loader:
batch = batch.to(args.device)
out = model(batch)
if std is not None:
err += (out * std - batch.y * std).abs().sum().item()
else:
err += (out - batch.y).abs().sum().item()
mae = err / len(loader.dataset)
return mae
if __name__ == '__main__':
print('Task start time:')
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
start_time = time.time()
parser = argparse.ArgumentParser(description='OCELOT training')
# Dataset arguments
parser.add_argument('--data_dir', type=str, required=True)
parser.add_argument('--target', type=int, default=0, help='target of dataset')
# Training hyperparameters
parser.add_argument('--runs', default=3, type=int)
parser.add_argument('--seed', default=0, type=int)
parser.add_argument('--device', type=int, default=0)
parser.add_argument('--epochs', default=300, type=int)
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--lr', default=0.0001, type=float)
parser.add_argument('--min_lr', default=0.000001, type=float)
parser.add_argument('--wd', default=0.0, type=float)
parser.add_argument('--clip_gnorm', default=None, type=float)
parser.add_argument('--log_steps', type=int, default=1)
# Model hyperparameters
parser.add_argument('--method', default='mhnn', help='model type')
parser.add_argument('--All_num_layers', default=3, type=int, help='number of basic blocks')
parser.add_argument('--MLP1_num_layers', default=2, type=int, help='layer number of mlps')
parser.add_argument('--MLP2_num_layers', default=2, type=int, help='layer number of mlp2')
parser.add_argument('--MLP3_num_layers', default=2, type=int, help='layer number of mlp3')
parser.add_argument('--MLP4_num_layers', default=2, type=int, help='layer number of mlp4')
parser.add_argument('--MLP_hidden', default=64, type=int, help='hidden dimension of mlps')
parser.add_argument('--output_num_layers', default=2, type=int)
parser.add_argument('--output_hidden', default=64, type=int)
parser.add_argument('--aggregate', default='mean', choices=['sum', 'mean'])
parser.add_argument('--normalization', default='ln', choices=['bn', 'ln', 'None'])
parser.add_argument('--activation', default='relu', choices=['Id', 'relu', 'prelu'])
parser.add_argument('--dropout', default=0.0, type=float)
args = parser.parse_args()
print(args)
device = f'cuda:{args.device}' if torch.cuda.is_available() else 'cpu'
device = torch.device(device)
# load dataset and normalize targets to mean = 0 and std = 1
if args.target in [0, 1, 2, 3]:
args.polymer = False
elif args.target in [4, 5, 6, 7]:
args.polymer = True
else:
raise Exception('Invalid target value!')
transform = T.Compose([OneTarget(target=args.target)])
if args.method == 'mhnn':
train_dataset = OPVHGraph(root=args.data_dir, polymer=args.polymer, partition='train', transform=transform)
valid_dataset = OPVHGraph(root=args.data_dir, polymer=args.polymer, partition='valid', transform=transform)
test_dataset = OPVHGraph(root=args.data_dir, polymer=args.polymer, partition='test', transform=transform)
else:
train_dataset = OPVGraph(root=args.data_dir, polymer=args.polymer, partition='train', transform=transform)
valid_dataset = OPVGraph(root=args.data_dir, polymer=args.polymer, partition='valid', transform=transform)
test_dataset = OPVGraph(root=args.data_dir, polymer=args.polymer, partition='test', transform=transform)
# Normalize targets to mean = 0 and std = 1.
mean = train_dataset.data.y.mean(dim=0, keepdim=True)
std = train_dataset.data.y.std(dim=0, keepdim=True)
train_dataset.data.y = (train_dataset.data.y - mean) / std
valid_dataset.data.y = (valid_dataset.data.y - mean) / std
test_dataset.data.y = (test_dataset.data.y - mean) / std
mean, std = mean[:, args.target].item(), std[:, args.target].item()
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True)
valid_loader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False)
# load logger
logger = Logger(args.runs, args)
for run in range(args.runs):
# set global seed for this run
seed = args.seed + run
seed_everything(seed=seed, workers=True)
print(f'\nRun No. {run+1}:')
print(f'Seed: {seed}\n')
# initialize model etc.
if args.method == 'mhnn':
model = MHNNS(1, args)
elif args.method in ['gin', 'gcn', 'gat', 'gatv2']:
model = GNN_2D(1, gnn_type=args.method, drop_ratio=args.dropout)
else:
raise ValueError(f'Undefined model name: {args.method}')
model = model.to(args.device)
print("# Params:", sum(p.numel() for p in model.parameters() if p.requires_grad))
loss_fn = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.wd)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
factor=0.7,
patience=5,
min_lr=args.min_lr)
# training
best_val_mae = None
for epoch in range(1, 1 + args.epochs):
model.train()
loss_all = 0.0
lr = scheduler.optimizer.param_groups[0]['lr']
for data in train_loader:
data = data.to(args.device)
optimizer.zero_grad()
out = model(data)
loss = loss_fn(out, data.y)
loss.backward()
loss_all += loss.item() * data.num_graphs
if args.clip_gnorm is not None:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_gnorm)
optimizer.step()
loss_all /= len(train_loader.dataset)
valid_mae = evaluate(args, model, valid_loader, std=std)
scheduler.step(valid_mae)
if best_val_mae is None or valid_mae < best_val_mae:
test_mae = evaluate(args, model, test_loader, std=std)
best_val_mae = valid_mae
logger.add_result(run, [loss_all, valid_mae, test_mae])
if epoch % args.log_steps == 0:
print(f'Run: {run + 1:02d}, '
f'Epoch: {epoch:02d}, '
f'lr: {lr:.6f}, '
f'Loss: {loss_all:.6f}, '
f'Valid MAE: {valid_mae:.6f}, '
f'Test MAE: {test_mae:.6f}')
logger.print_statistics(run)
logger.print_statistics()
print('Task end time:')
print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
end_time = time.time()
print('Total time taken: {} s.'.format(int(end_time - start_time)))