-
Notifications
You must be signed in to change notification settings - Fork 1
/
custom_gnn.py
336 lines (283 loc) · 14 KB
/
custom_gnn.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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
import random
from utils import accuracy, load_data, pathsGen, consis_loss, load_large_dataset, load_syn_cora, load_Ogbn
import torch
import torch.nn.functional as F
from models import PathWeightModel
import argparse
import numpy as np
import torch.optim as optim
import time
parser = argparse.ArgumentParser()
# globla settings
parser.add_argument('--only_test', type=bool, default=False)
parser.add_argument('--para_name', type=str, default='cora')
parser.add_argument('--pw_adj_name', type=str, default='ogbn-arxiv')
parser.add_argument('--dataset', type=str, default='ogbn-arxiv') #syn-cora ogbn-arxiv CoraFull CoauthorCS CoauthorPhysics AminerCS AmazonComputers AmazonPhoto
parser.add_argument('--homophily_ratio_name', type=str, default='h0.80-r1')
parser.add_argument('--use_label_rate', type=bool, default=False)
parser.add_argument('--num_example_per_class', type=float, default=2)
parser.add_argument('--path', type=str, default='./data/')
parser.add_argument('--cuda',type=bool, default=True, help='ables CUDA training.')
parser.add_argument('--seed', type=int, default=42, help='Random seed.')
parser.add_argument('--cuda_device', type=int, default=1, help='cuda device')
parser.add_argument('--patience', type=int, default=300, help='patience')
parser.add_argument('--schedule_patience', type=int, default=50, help='schedule_patience')
parser.add_argument('--epochs', type=int, default=10000, help='Number of epochs to train.')
parser.add_argument('--lam_pw_emd', type=float, default=10.0, help='tradeoff between \\hat{F} and \\dot{F}.') #citeseer 5 10; cora 0.1 0.5;
# triplet loss settings
parser.add_argument('--use_triple', type=int, default=0, help='1 or 0')# 50 135 128 256 for citeseer cora pubmed nell
parser.add_argument('--samp_neg', type=int, default=5000, help='negative pairs of triple_loss.')
parser.add_argument('--samp_pos', type=int, default=15000, help='positive pairs of triple_loss.')
parser.add_argument('--lam_tri', type=float, default=.01, help='the weight of triple loss') # 3=85.1, 74.3,
parser.add_argument('--lam_tri_lstm', type=float, default=1., help='the weight of triple loss')
parser.add_argument('--margin', type=float, default=.1, help='margin between negative samples')
parser.add_argument('--pesudo_ratio', type=float, default=1., help='number of pesudo labels for triple loss')
# lr
parser.add_argument('--learning_rate', type=float, default=0.01)
parser.add_argument('--weight_decay', type=float, default=5e-4, help='Weight decay (L2 loss on parameters).')
# batch size
parser.add_argument('--batch_size', type=int, default=300)
# consistency loss settings
parser.add_argument('--use_consis', type=bool, default=True)
parser.add_argument('--T', type=float, default=0.5, help='temperature')
parser.add_argument('--K', type=int, default=4, help='number of batch per epoch')
parser.add_argument('--lam_u', type=int, default=1.,
help='tradeoff between sup loss and unsup loss, loss=sup_loss + lam_u * unsup_loss')
#Path Attention module settings
parser.add_argument('--embedding_dim', type=int, default=512)
parser.add_argument('--window_size', type=int, default=8) # cora3,10; citeseer 456
parser.add_argument('--path_length', type=int, default=8)
parser.add_argument('--lstm_hidden_units', type=int, default=128)
# MLP settings
parser.add_argument('--nhid', type=int, default=128, help='Number of hidden units.')
parser.add_argument('--dropout_pw', type=float, default=0.6, help='Dropout rate (1 - keep probability).')
parser.add_argument('--dropout_adj', type=float, default=0.5, help='Dropout rate (1 - keep probability).')
parser.add_argument('--dropout_enc', type=float, default=0.6, help='Dropout rate (1 - keep probability).')
parser.add_argument('--dropout_input', type=float, default=0.5, help='Dropout rate (1 - keep probability).')
parser.add_argument('--dropout_hidden', type=float, default=0.5, help='Dropout rate (1 - keep probability).')
parser.add_argument('--alpha', type=float, default=0.2, help='Alpha for the leaky_relu.')
parser.add_argument('--order', type=int, default=8, help='multi-hop gnn order') # cora 8 10, citeseer 4 5
args = parser.parse_args()
args.cuda = args.cuda and torch.cuda.is_available()
torch.cuda.set_device(args.cuda_device)
dataset = args.dataset
# np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
# device = 'cuda:{}'.format(args.cuda_device)
# load data
if args.dataset in ['CoraFull', 'CoauthorCS', 'AmazonComputers', 'AmazonPhoto']:
graph, labels, adj, features, idx_val, idx_test, idx_train = load_large_dataset(root=args.path, name=args.dataset)
elif args.dataset in ['cora', 'citeseer', 'pubmed']:
graph, labels, adj, features, idx_val, idx_test, idx_train, _ = load_data(args.dataset, args.path+args.dataset)
elif args.dataset in ['ogbn-arxiv']:
graph, labels, adj, features, idx_val, idx_test, idx_train = load_Ogbn(dataset, root=args.path)
elif args.dataset in ['syn-cora', 'syn-product']:
seed = random.randint(0,200)
graph, labels, adj, features, idx_val, idx_test, idx_train = load_syn_cora(args.dataset, args.path, args.homophily_ratio_name, seed)
node_num = len(graph)
nclass = int(labels.max()) + 1
embedding_dim =args.embedding_dim
print('init model ...')
model = PathWeightModel(
features.shape[1],
args.nhid,
nclass,
args.lstm_hidden_units,
node_num,
args.order,
embedding_dim,
args.lam_pw_emd,
args.alpha,
args.dropout_pw,
args.dropout_adj,
args.dropout_enc,
args.dropout_input,
args.dropout_hidden
)
optimizer = optim.Adam(model.parameters(),
lr=args.learning_rate, weight_decay=args.weight_decay)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.98)
if args.cuda:
model.cuda()
features = features.cuda()
adj = adj.cuda()
labels = labels.cuda()
idx_train = idx_train.cuda()
idx_val = idx_val.cuda()
idx_test = idx_test.cuda()
print('complete init model')
device = next(model.parameters()).device
print('device =', device)
iter = pathsGen(node_num, args.batch_size, graph, args.path_length, args.window_size)
# plot path weight values
def triple_loss(features,pw_adj, logs,samp_neg, samp_pos, margin=args.margin):
global model
label_num = int(args.pesudo_ratio * features.shape[0])
embeddings = model.encode(features*1.0)
idx = torch.tensor(np.random.randint(0, features.shape[0], label_num), dtype=torch.long, device=device)
ps = [torch.exp(p)[idx] for p in logs]
sum_p = 0.
for p in ps:
sum_p = sum_p + p
avg_p = sum_p / len(ps)
out = avg_p.argmax(dim=-1).view(-1,1)
indicator_adj = (out == out.t())
negs = torch.where((~indicator_adj).triu(1))
poss = torch.where(indicator_adj.triu(1))
samp_neg = samp_neg if samp_neg <= negs[0].shape[0] else negs[0].shape[0]
samp_pos = samp_pos if samp_pos <= poss[0].shape[0] else poss[0].shape[0]
samp_neg = np.random.randint(0,negs[0].shape[0], samp_neg)
samp_pos = np.random.randint(0, poss[0].shape[0], samp_pos)
loss_neg = torch.mean(F.pairwise_distance(embeddings[idx[negs[0][samp_neg]]], embeddings[idx[negs[1][samp_neg]]]))
loss_pos = torch.mean(F.relu(margin - F.pairwise_distance(embeddings[idx[poss[0][samp_pos]]], embeddings[idx[poss[1][samp_pos]]])))
loss = 0.
loss = loss_pos + loss_neg
# pw_emd = torch.matmul(pw_adj, embeddings)
pw_emd = torch.matmul(pw_adj, embeddings*1.0)
# pw_emd = torch.cat((gnn_emd, pw_emd), dim=1)
loss_pos_lstm = torch.mean(F.pairwise_distance(pw_emd[idx[poss[0][samp_pos]]], pw_emd[idx[poss[1][samp_pos]]]))
loss_neg_lstm = torch.mean(
F.relu(margin - F.pairwise_distance(pw_emd[idx[negs[0][samp_neg]]], pw_emd[idx[negs[1][samp_neg]]])))
loss_lstm = loss_pos_lstm + loss_neg_lstm
all_loss = loss * args.lam_tri + loss_lstm * args.lam_tri_lstm
return all_loss
loss_val_list = []
loss_train_list = []
acc_val_list = []
acc_train_list = []
acc_best = 0.
loss_best = 100.
loss_mn = np.inf
acc_mx = 0.0
bad_counter = 1
best_epoch = 0
def train():
t_total = time.time()
global iter, best_epoch, acc_best, loss_best, bad_counter, loss_mn, acc_mx
model.train()
for epoch in range(args.epochs):
t = time.time()
X_list = []
for _ in range(args.K):
try:
data = next(iter)
except StopIteration:
iter = pathsGen(node_num, args.batch_size, graph, args.path_length, args.window_size)
data = next(iter)
X_list.append(data)
# foward
out_list=[]
loss_train = 0.
pw_adj = 0
for k in range(args.K):
out, pw = model(features, adj, *(X_list[k]))
# embeddings.append(pw_embedding)
out_list.append(out)
loss_train += F.nll_loss(out_list[k][idx_train], labels[idx_train])
pw_adj += pw
# normalize
rowsum = pw_adj.sum(1)
r_inv = torch.pow(rowsum, -1).view(-1)
r_inv[torch.isinf(r_inv)] = 0.
pw_adj.mul_(r_inv.t())
loss_train = loss_train / args.K
#pw_adj /= args.K
cons_loss = 0.
if args.use_consis:
cons_loss = consis_loss(out_list, args.T) * args.lam_u
loss_train += cons_loss
tri_loss = 0.
if args.use_triple:
tri_loss += triple_loss(features, pw_adj, out_list, args.samp_neg, args.samp_pos, args.margin)
loss_train += tri_loss
# bp
optimizer.zero_grad()
loss_train.backward()
optimizer.step()
# eval
model.eval()
with torch.no_grad():
# batch = next(pathsGen(node_num, args.test_batch_size, graph, args.path_length, args.window_size))
batch = data
output, _ = model(features, adj, *batch, pw_adj=pw_adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val])
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_val = accuracy(output[idx_val], labels[idx_val])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_val_list.append(loss_val.item())
loss_train_list.append(loss_train.item())
acc_val_list.append(acc_val.item())
acc_train_list.append(acc_train.item())
spend = time.time() - t
print('Epoch:{:04d}'.format(epoch + 1),
'tri:{:4f}'.format(tri_loss),
'consis:{:4f}'.format(cons_loss),
'loss_train:{:.4f}'.format(loss_train.item()),
'acc_train:{:.4f}'.format(acc_train.item()),
'loss_val:{:.4f}'.format(loss_val.item()),
'acc_val:{:.4f}'.format(acc_val.item()),
'acc_best:{:.4f}'.format(acc_best),
'loss_best:{:.4f}'.format(loss_best),
'bad_count:{:03d}'.format(bad_counter),
'time:{:.4f}s'.format(spend),
'remain_time:{:.4f}h'.format((args.epochs-epoch)*spend/3600),
),
if loss_val_list[-1] <= loss_mn or acc_val_list[-1] >= acc_mx: # or epoch < 400:
if acc_val_list[-1] >= acc_best : #and loss_val_list[-1] <= loss_best: #
loss_best = loss_val_list[-1]
acc_best = acc_val_list[-1]
best_epoch = epoch
if epoch >= 0:
torch.save(pw_adj, "./checkpoint/"+dataset+'.pth')
torch.save(model.state_dict(), "./checkpoint/"+dataset + '.pkl')
loss_mn = np.min((loss_val_list[-1], loss_mn))
acc_mx = np.max((acc_val_list[-1], acc_mx))
bad_counter = 1
else:
bad_counter += 1
if bad_counter % args.schedule_patience == 0 and scheduler.get_lr()[-1] >= 1e-4:
scheduler.step()
print('schedule learning rate to',scheduler.get_lr()[-1])
if bad_counter == args.patience:
print('Early stop! Min loss: ', loss_mn, ', Max accuracy: ', acc_mx)
print('Early stop model validation loss: ', loss_best, ', accuracy: ', acc_best)
print("Optimization Finished!")
print("Total time elapsed: {:.4f}s".format(time.time() - t_total))
break
model.train()
def test(pw_adj_name):
model.eval()
start_time = time.time()
with torch.no_grad():
X = next(pathsGen(node_num, 0, graph, args.path_length, args.window_size))
pw_adj = torch.load("./checkpoint/"+ pw_adj_name+'.pth', map_location=device)
output, _ = model(features, adj, *X, pw_adj=pw_adj)
end_time = time.time()
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
loss_val = F.nll_loss(output[idx_val], labels[idx_val])
acc_test = accuracy(output[idx_test], labels[idx_test])
acc_val = accuracy(output[idx_val], labels[idx_val])
acc_train = accuracy(output[idx_train], labels[idx_train])
print("Test set results:",
"loss_test= {:.4f}".format(loss_test.data.item()),
"loss_train= {:.4f}".format(loss_train.data.item()),
"acc_test= {:.4f}".format(acc_test.item()),
"acc_train= {:.4f}".format(acc_train.item()),
"test_time= {:.4f}".format(end_time - start_time))
return loss_test.item(), loss_val.item(), loss_train.item(), acc_test.item(), acc_val.item(), acc_train.item()
if not args.only_test:
train()
para_name = args.dataset
pw_adj_name = args.dataset
else:
para_name = args.para_name
pw_adj_name = args.pw_adj_name
## test
# Restore best model
print('Loading {}'.format("./checkpoint/"+para_name))
model.load_state_dict(torch.load("./checkpoint/"+ para_name + '.pkl', map_location=device))
model.eval()
test(pw_adj_name)