-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
181 lines (160 loc) · 7.46 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
from utils import to_undirected, remove_self_loops
from data_utils import eval_acc, eval_rocauc
from dataset import load_nc_dataset
from esgnn import ESGNN, Ir_Consistency_Loss
from parse import args
import torch.nn.functional as fn
import torch.optim as optim
import torch.nn as nn
import torch.autograd
import numpy as np
import torch
import dgl
import tempfile
import random
import time
# noinspection PyUnresolvedReferences
def set_rng_seed(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.deterministic = True
class EvalHelper:
# noinspection PyUnresolvedReferences
def __init__(self, args):
use_cuda = torch.cuda.is_available() and not args.cpu
dev = torch.device('cuda' if use_cuda else 'cpu')
# load dataset
dataset = load_nc_dataset(args.dataset, args.sub_dataset)
# load splits
if args.dataset in ['chameleon', 'squirrel', 'film', 'twitch-e', 'cora', 'citeseer', 'pubmed', 'polblogs']:
if args.dataset == "twitch-e":
assert args.sub_dataset in ['DE', 'ENGB', 'ES', 'FR', 'PTBR', 'RU', 'TW']
split_dic = np.load(f"{args.DATAPATH}split/LR{args.label_rate}/LR{args.label_rate}_{args.dataset}_{args.sub_dataset}_splits.npy",
allow_pickle=True)[args.split_index]
elif args.dataset == "etg_syn_hom":
split_dic = dataset.graph[f"LR{args.label_rate}_splits"][args.split_index]
else:
raise ValueError('Invalid method')
trn_idx, val_idx, tst_idx = np.array(split_dic["trn_idx"]), np.array(split_dic["val_idx"]), np.array(split_dic["tst_idx"])
assert len(set(trn_idx).intersection(val_idx)) == 0
assert len(set(trn_idx).intersection(tst_idx)) == 0
assert len(set(val_idx).intersection(tst_idx)) == 0
# pre-processing
if len(dataset.label.shape) == 1:
dataset.label = dataset.label.unsqueeze(1)
# n = dataset.graph['num_nodes']
c = dataset.label.max().item() + 1
d = dataset.graph['node_feat'].shape[1]
dataset.graph["edge_index"] = remove_self_loops(dataset.graph["edge_index"])[0]
dataset.graph['edge_index'] = to_undirected(dataset.graph['edge_index'])
# to-cuda
trn_idx = torch.from_numpy(trn_idx).to(dev)
val_idx = torch.from_numpy(val_idx).to(dev)
tst_idx = torch.from_numpy(tst_idx).to(dev)
dataset.label = dataset.label.to(dev)
dataset.graph['edge_index'] = dataset.graph['edge_index'].to(dev)
dataset.graph['node_feat'] = dataset.graph['node_feat'].to(dev)
edge = dataset.graph["edge_index"]
g = dgl.graph((edge[0], edge[1]))
model = ESGNN(g, d, args.hidden_channels, c, args.dropout, re_eps=args.re_eps, ir_eps=args.ir_eps, layer_num=args.num_layers).to(dev)
all_params = model.parameters()
self.ir_con_loss_fn = Ir_Consistency_Loss(g, args.hidden_channels // 2, args.dropout).to(dev)
all_params = list(all_params) + list(self.ir_con_loss_fn.parameters())
optmz = optim.Adam(all_params, lr=args.lr, weight_decay=args.reg)
# considering the class-imbalance problem in Twitch-DE
if args.rocauc or args.dataset == 'twitch-e':
loss_fn = nn.BCEWithLogitsLoss()
eval_fn = eval_rocauc
else:
loss_fn = nn.NLLLoss()
eval_fn = eval_acc
self.dataset = dataset
self.trn_idx, self.val_idx, self.tst_idx = trn_idx, val_idx, tst_idx
self.model, self.optmz = model, optmz
self.loss_fn, self.eval_fn = loss_fn, eval_fn
self.args = args
def before_loss(self, args, out):
# considering the class-imbalance problem in Twitch-DE
if args.rocauc or args.dataset == 'twitch-e':
true_label = fn.one_hot(self.dataset.label, self.dataset.label.max() + 1).type(out.dtype)
else:
true_label = self.dataset.label
out = fn.log_softmax(out, dim=1)
return out, true_label
def to_onehot(self, input):
oh_input = torch.zeros(input.shape[0], input.max() + 1).to(input.device)
val_idxes = torch.where(input >= 0)[0]
oh_input[val_idxes] = fn.one_hot(input[val_idxes].long(), input[val_idxes].long().max() + 1).to(oh_input.dtype)
return oh_input
def run_epoch(self, args):
self.model.train()
self.optmz.zero_grad()
re_logits, ir_logits, re_feat, ir_feat = self.model(self.dataset.graph["node_feat"])
# prediction loss
re_out, true_label = self.before_loss(args, re_logits)
pred_loss = self.loss_fn(re_out[self.trn_idx], true_label.squeeze(1)[self.trn_idx])
# Irrelevant Consistency Regularization
valtst_idx = torch.cat((self.val_idx, self.tst_idx), dim=0)
masked_pred = torch.zeros_like(re_logits)
masked_pred[self.trn_idx] = self.to_onehot(self.dataset.label.squeeze(1)).float()[self.trn_idx]
masked_pred[valtst_idx] = fn.softmax(re_logits[valtst_idx], dim=1)
ir_con_loss = self.ir_con_loss_fn(masked_pred, ir_feat)
loss = pred_loss + args.ir_con_lambda * ir_con_loss
loss.backward()
self.optmz.step()
print("epoch-loss: {:.4f}, pred-loss: {:.4f}, ir-con-loss: {:.4f}".format(loss.item(), pred_loss.item(), ir_con_loss.item()))
return loss.item()
def evaluate(self):
self.model.eval()
out, _, _, _ = self.model(self.dataset.graph["node_feat"])
trn_acc = self.eval_fn(self.dataset.label[self.trn_idx], out[self.trn_idx])
val_acc = self.eval_fn(self.dataset.label[self.val_idx], out[self.val_idx])
tst_acc = self.eval_fn(self.dataset.label[self.tst_idx], out[self.tst_idx])
return trn_acc, val_acc, tst_acc
# noinspection PyUnresolvedReferences
def train_and_eval(args):
# fix random initialization
set_rng_seed(args.rnd_seed)
# build model
agent = EvalHelper(args)
# trn and val
wait_cnt, best_epoch = 0, 0
best_val_acc = 0.0
best_model_sav = tempfile.TemporaryFile()
ct_ls = []
for t in range(args.nepoch):
cur_time = time.time()
agent.run_epoch(args)
ct_ls.append(time.time() - cur_time)
trn_acc, val_acc, tst_acc = agent.evaluate()
print("epoch: {}/{}, trn-acc={:.4f}%, val-acc={:.4f}%, tst-acc={:.4f}%".format(
t + 1, args.nepoch, trn_acc * 100, val_acc * 100, tst_acc * 100))
# training with early-stop
if val_acc > best_val_acc:
wait_cnt = 0
best_val_acc = val_acc
best_model_sav.close()
best_model_sav = tempfile.TemporaryFile()
torch.save(agent.model.state_dict(), best_model_sav)
best_epoch = t + 1
else:
wait_cnt += 1
if wait_cnt > args.early:
break
# final results
print("Load selected model ...")
best_model_sav.seek(0)
agent.model.load_state_dict(torch.load(best_model_sav))
trn_acc, val_acc, tst_acc = agent.evaluate()
print("trn-acc={:.4f}%, val-acc={:.4f}%, tst-acc={:.4f}%, avg-epoch-time={:.4f}".format(trn_acc * 100, val_acc * 100, tst_acc * 100, np.mean(ct_ls)))
return val_acc, tst_acc, best_epoch
def run(args):
val_acc, tst_acc, selected_epoch = train_and_eval(args)
print("val_acc={:.4f}%, tst_acc={:.4f}%, selected_epoch={}".format(val_acc * 100, tst_acc * 100, selected_epoch))
def main():
run(args)
if __name__ == '__main__':
main()