-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
167 lines (154 loc) · 6.85 KB
/
train.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
from __future__ import division
from __future__ import print_function
import datetime
import json
import logging
import os
import pickle
import time
import numpy as np
import optimizers
import torch
from config import parser
from models.base_models import NCModel, LPModel, MDModel
from utils.data_utils import load_data
from utils.train_utils import get_dir_name, format_metrics
torch.autograd.set_detect_anomaly(True)
# import os
# os.environ['CUDA_VISIBLE_DEVICES'] = '1'
# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# os.environ["CUDA_VISIBLE_DEVICES"] = "2"
def train(args):
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if int(args.double_precision):
torch.set_default_dtype(torch.float64)
if int(args.cuda) >= 0:
torch.cuda.manual_seed(args.seed)
args.device = 'cuda:' + str(args.cuda) if int(args.cuda) >= 0 else 'cpu'
args.patience = args.epochs if not args.patience else int(args.patience)
logging.getLogger().setLevel(logging.INFO)
if args.save:
if not args.save_dir:
dt = datetime.datetime.now()
date = f"{dt.year}_{dt.month}_{dt.day}"
models_dir = os.path.join(os.environ['LOG_DIR'], args.task, date)
save_dir = get_dir_name(models_dir)
else:
save_dir = args.save_dir
logging.basicConfig(level=logging.INFO,
handlers=[
logging.FileHandler(os.path.join(save_dir, 'log.txt')),
logging.StreamHandler()
])
logging.info(f'Using: {args.device}')
logging.info("Using seed {}.".format(args.seed))
# Load data
data = load_data(args, os.path.join(os.environ['DATAPATH'], args.dataset))
args.n_nodes, args.feat_dim = data['features'].shape
if args.task == 'nc':
Model = NCModel
args.n_classes = int(data['labels'].max() + 1)
logging.info(f'Num classes: {args.n_classes}')
elif args.task == 'md':
Model = MDModel
args.eval_freq = args.epochs + 1
else:
args.nb_false_edges = len(data['train_edges_false'])
args.nb_edges = len(data['train_edges'])
if args.task == 'lp':
Model = LPModel
else:
Model = RECModel
# No validation for reconstruction task
args.eval_freq = args.epochs + 1
if not args.lr_reduce_freq:
args.lr_reduce_freq = args.epochs
# Model and optimizer
model = Model(args)
logging.info(str(model))
# print(model.parameters())
curvature_lr = 1e-4
if args.model == 'HGCN' and args.task == 'md':
pararms = [{'params': model.encoder.layers[0].linear.weight}, {'params': model.encoder.layers[0].linear.bias}, {'params': model.encoder.layers[0].c_in, 'lr':curvature_lr},{'params': model.encoder.layers[0].c_out, 'lr':curvature_lr} ]
optimizer = getattr(optimizers, args.optimizer)(pararms, lr=args.lr,weight_decay=args.weight_decay)
else:
optimizer = getattr(optimizers, args.optimizer)(params=model.parameters(), lr=args.lr,weight_decay=args.weight_decay)
# for g in optimizer
lr_scheduler = torch.optim.lr_scheduler.StepLR(
optimizer,
step_size=int(args.lr_reduce_freq),
gamma=float(args.gamma)
)
tot_params = sum([np.prod(p.size()) for p in model.parameters()])
logging.info(f"Total number of parameters: {tot_params}")
if args.cuda is not None and int(args.cuda) >= 0:
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda)
model = model.to(args.device)
for x, val in data.items():
if torch.is_tensor(data[x]):
data[x] = data[x].to(args.device)
# Train model
t_total = time.time()
counter = 0
best_val_metrics = model.init_metric_dict()
best_test_metrics = None
best_emb = None
for epoch in range(args.epochs):
t = time.time()
model.train()
optimizer.zero_grad()
embeddings = model.encode(data['features'], data['adj_train_norm'])
train_metrics = model.compute_metrics(embeddings, data, 'train')
train_metrics['loss'].backward()
if args.grad_clip is not None:
max_norm = float(args.grad_clip)
all_params = list(model.parameters())
for param in all_params:
torch.nn.utils.clip_grad_norm_(param, max_norm)
optimizer.step()
lr_scheduler.step()
if (epoch + 1) % args.log_freq == 0:
logging.info(" ".join(['Epoch: {:04d}'.format(epoch + 1),
'lr: {}'.format(lr_scheduler.get_lr()[0]),
format_metrics(train_metrics, 'train'),
'time: {:.4f}s'.format(time.time() - t)
]))
if (epoch + 1) % args.eval_freq == 0:
model.eval()
embeddings = model.encode(data['features'], data['adj_train_norm'])
val_metrics = model.compute_metrics(embeddings, data, 'val')
if (epoch + 1) % args.log_freq == 0:
logging.info(" ".join(['Epoch: {:04d}'.format(epoch + 1), format_metrics(val_metrics, 'val')]))
if model.has_improved(best_val_metrics, val_metrics):
best_test_metrics = model.compute_metrics(embeddings, data, 'test')
best_emb = embeddings.cpu()
if args.save:
np.save(os.path.join(save_dir, 'embeddings.npy'), best_emb.detach().numpy())
best_val_metrics = val_metrics
counter = 0
else:
counter += 1
if counter == args.patience and epoch > args.min_epochs:
logging.info("Early stopping")
break
logging.info("Optimization Finished!")
logging.info("Total time elapsed: {:.4f}s".format(time.time() - t_total))
if not best_test_metrics:
model.eval()
best_emb = model.encode(data['features'], data['adj_train_norm'])
best_test_metrics = model.compute_metrics(best_emb, data, 'test')
logging.info(" ".join(["Val set results:", format_metrics(best_val_metrics, 'val')]))
logging.info(" ".join(["Test set results:", format_metrics(best_test_metrics, 'test')]))
if args.save:
np.save(os.path.join(save_dir, 'embeddings.npy'), best_emb.cpu().detach().numpy())
if hasattr(model.encoder, 'att_adj'):
filename = os.path.join(save_dir, args.dataset + '_att_adj.p')
pickle.dump(model.encoder.att_adj.cpu().to_dense(), open(filename, 'wb'))
print('Dumped attention adj: ' + filename)
json.dump(vars(args), open(os.path.join(save_dir, 'config.json'), 'w'))
torch.save(model.state_dict(), os.path.join(save_dir, 'model.pth'))
logging.info(f"Saved model in {save_dir}")
if __name__ == '__main__':
args = parser.parse_args()
train(args)