-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
195 lines (165 loc) · 8.5 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
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
import os
import gc
import argparse
import random
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch_geometric.utils as utils
from tqdm import tqdm
from utils import dataloader, early_stopping
from model import models
def set_env(seed):
# os.environ['CUDA_VISIBLE_DEVICES'] = '0, 1'
# os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
# os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':16:8'
os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
os.environ['PYTHONHASHSEED'] = str(seed)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# torch.backends.cudnn.enabled = False
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
# torch.use_deterministic_algorithms(True)
def get_parameters():
parser = argparse.ArgumentParser(description='ChebGibbsNet')
parser.add_argument('--enable_cuda', type=bool, default=True, help='enable or disable CUDA (default: True)')
parser.add_argument('--model_name', type=str, default='chebgibbsnet')
parser.add_argument('--dataset_name', type=str, default='cora', choices=['cora', 'citeseer', 'pubmed', \
'computers', 'photo', 'chameleon', 'squirrel', 'film', 'cornell', 'texas', 'wisconsin'])
parser.add_argument('--order', type=int, default=10, help='polynomial order (default: 10)')
parser.add_argument('--gibbs_type', type=str, default='jackson', choices=['none', 'dirichlet', \
'fejer', 'jackson', 'lanczos', 'lorentz', 'vekic', 'wang'], help='Gibbs damping factor type (default: jackson)')
parser.add_argument('--mu', type=int, default=3, help='mu for Lanczos (default: 3)')
parser.add_argument('--xi', type=float, default=4.0, help='xi for Lorentz (default: 4.0)')
parser.add_argument('--stigma', type=float, default=0.5, help='stigma for Vekic (default: 0.5)')
parser.add_argument('--heta', type=int, default=2, help='heta for Wang (default: 2)')
parser.add_argument('--act', type=str, default='smu', choices=['silu', 'gelu', 'mish', \
'tanhexp', 'sinsig', 'diracrelu', 'smu'], help='activation function (default: smu)')
parser.add_argument('--droprate_pre', type=float, default=0, help='dropout rate for Dropout before MLP (default: 0)')
parser.add_argument('--droprate_in', type=float, default=0, help='dropout rate for Dropout inside MLP (default: 0)')
parser.add_argument('--droprate_suf', type=float, default=0, help='dropout rate for Dropout after MLP (default: 0)')
parser.add_argument('--num_hid', type=int, default=64, help='the channel size of hidden layer feature (default: 64)')
parser.add_argument('--bs', type=int, default=1, help='batch size (default: 1)')
parser.add_argument('--lr', type=float, default=0.01, help='learning rate (default: 0.01)')
parser.add_argument('--weight_decay', type=float, default=0.0005, help='weight decay (default: 0.0005)')
parser.add_argument('--epochs', type=int, default=1000, help='epochs (default: 1000)')
parser.add_argument('--opt', type=str, default='adam', choices=['adam', 'adamw'], help='optimizer (default: adam)')
parser.add_argument('--patience', type=int, default=50, help='early stopping patience (default: 50)')
args = parser.parse_args()
print('Training configs: {}'.format(args))
# Running in Nvidia GPU (CUDA) or CPU
if args.enable_cuda and torch.cuda.is_available():
device = torch.device('cuda')
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
else:
device = torch.device('cpu')
gc.collect()
return args, device
def prepare_model(args, device):
dataset, data = dataloader.data_loader(args.dataset_name)
if hasattr(dataset, 'is_undirected'):
if data.edge_weight is None:
data.edge_index = utils.to_undirected(edge_index=data.edge_index, reduce='max')
else:
data.edge_index, data.edge_weight = utils.to_undirected(edge_index=data.edge_index,
edge_attr=data.edge_weight, num_nodes=data.x.size()[0],
reduce='max')
homophily = utils.homophily(edge_index=data.edge_index, y=data.y, method='node')
data, model = data.to(device), models.ChebGibbsNet(dataset, args, homophily).to(device)
if args.bs == 1:
if args.dataset_name in ['chameleon', 'cornell', 'film', 'squirrel', 'texas', \
'wisconsin', 'cora_ml', 'citeseer_dir', 'telegram']:
train_mask = data.train_mask[:, 0]
val_mask = data.val_mask[:, 0]
test_mask = data.test_mask[:, 0]
elif args.dataset_name == 'wikics':
train_mask = data.train_mask[:, 0]
val_mask = data.val_mask[:, 0]
test_mask = data.test_mask
else:
train_mask = data.train_mask
val_mask = data.val_mask
test_mask = data.test_mask
loss = nn.CrossEntropyLoss()
es = early_stopping.EarlyStopping(delta=0.0,
patience=args.patience,
verbose=True,
path="chebgibbsnet_" + args.gibbs_type + "_" + args.dataset_name + ".pt")
if args.opt == 'adam': # default
optimizer = optim.Adam(params=model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
elif args.opt == 'adamw':
optimizer = optim.AdamW(params=model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
else:
raise ValueError(f'ERROR: The {args.opt} optimizer is undefined.')
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.95)
return data, loss, train_mask, val_mask, test_mask, model, es, optimizer, scheduler
def run(args, model, optimizer, data, loss, train_mask, val_mask, test_mask, scheduler, es):
for epoch in tqdm(range(1, args.epochs+1)):
loss_train, acc_train = train(model, optimizer, data, loss, train_mask, scheduler)
loss_val, acc_val = val(model, data, loss, val_mask)
es(loss_val, model)
if es.early_stop:
print("Early stopping")
break
acc_train_list.append(acc_train)
acc_val_list.append(acc_val)
loss_test, acc_test = test(args, model, data, loss, test_mask)
acc_test_list.append(acc_test)
print(f'test acc: {acc_test * 100:.2f}%')
def train(model, optimizer, data, loss, mask, scheduler):
model.train()
optimizer.zero_grad()
out = model(data)
loss_train= loss(out[mask], data.y[mask])
acc_train = calc_accuracy(out, data, mask)
loss_train.backward()
optimizer.step()
# scheduler.step()
return loss_train, acc_train
@torch.no_grad()
def val(model, data, loss, mask):
model.eval()
out = model(data)
loss_val = loss(out[mask], data.y[mask])
acc_val = calc_accuracy(out, data, mask)
return loss_val, acc_val
@torch.no_grad()
def test(args, model, data, loss, mask):
model.load_state_dict(torch.load("chebgibbsnet_" + args.gibbs_type + "_" + args.dataset_name + ".pt"))
model.eval()
out = model(data)
loss_test = loss(out[mask], data.y[mask])
acc_test = calc_accuracy(out, data, mask)
return loss_test, acc_test
def calc_accuracy(out, data, mask):
preds = out.argmax(dim=-1)
acc = int((preds[mask] == data.y[mask]).sum()) / int(mask.sum())
return acc
if __name__ == '__main__':
seeds = [1, 42, 3407, 10076, 934890, 74512355, 124, 2132134, 43059, 2354367]
acc_train_list = []
acc_val_list = []
acc_test_list = []
args, device = get_parameters()
for i in range(len(seeds)):
set_env(seeds[i])
data, loss, train_mask, val_mask, test_mask, model, es, optimizer, scheduler = prepare_model(args, device)
run(args, model, optimizer, data, loss, train_mask, val_mask, test_mask, scheduler, es)
# set_env(42)
# args, device = get_parameters()
# data, loss, train_mask, val_mask, test_mask, model, es, optimizer, scheduler = prepare_model(args, device)
# run(args, model, optimizer, data, loss, train_mask, val_mask, test_mask, scheduler, es)
acc_train_mean = np.mean(acc_train_list)
acc_train_std = np.std(acc_train_list)
acc_val_mean = np.mean(acc_val_list)
acc_val_std = np.std(acc_val_list)
acc_test_mean = np.mean(acc_test_list)
acc_test_std = np.std(acc_test_list)
print(f'test acc mean: {acc_test_mean * 100:.2f}%')
print(f'test acc std: {acc_test_std * 100:.2f}%')