-
Notifications
You must be signed in to change notification settings - Fork 1
/
main_un.py
336 lines (289 loc) · 12.3 KB
/
main_un.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 os
from collections import Counter
import torch.nn.utils.prune as prune
import torch
from sklearn.metrics import accuracy_score
from sklearn.model_selection import GridSearchCV, StratifiedKFold
from sklearn.svm import SVC
from torch.autograd import Variable
from tqdm import tqdm
from GCL.models import DualBranchContrast
import GCL.losses as L
import GCL.augmentors as A
from torch.nn import Linear, ReLU
import torch_geometric.transforms as T
from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DataLoader
from torch_geometric.utils import degree, to_dense_batch
from CGT_model.CGT import CGT
from aug.DropPath import DropPath
import random
import numpy as np
import torch.nn.functional as F
from datetime import datetime
import warnings
from aug.gumble import gumbel_softmax
from util import save_accs, gene_arg
warnings.filterwarnings("ignore")
now = datetime.now()
now = now.strftime("%m_%d-%H_%M_%S")
transform = T.AddRandomWalkPE(walk_length=20, attr_name='pe')
args = gene_arg()
class NormalizedDegree(object):
def __init__(self, mean, std):
self.mean = mean
self.std = std
def __call__(self, data):
deg = degree(data.edge_index[0], dtype=torch.float)
deg = (deg - self.mean) / self.std
data.x = deg.view(-1, 1)
return data
def svc_classify(fold, x, y, search):
kf = StratifiedKFold(n_splits=fold, shuffle=True, random_state=42)
accuracies = []
for train_index, test_index in tqdm(kf.split(x, y), total=kf.get_n_splits()):
x_train, x_test = x[train_index], x[test_index]
y_train, y_test = y[train_index], y[test_index]
if search:
params = {'C': [0.001, 0.01, 0.1, 1, 10, 100, 1000]}
classifier = GridSearchCV(SVC(), params, cv=5, scoring='accuracy', verbose=0)
else:
classifier = SVC(C=10)
classifier.fit(x_train, y_train)
accuracies.append(accuracy_score(y_test, classifier.predict(x_test)))
return np.mean(accuracies), np.std(accuracies)
def load_data(args, transform):
data_name = args.dataset + '-pe'
dataset = TUDataset(os.path.join(args.data_root, data_name),
name=args.dataset
)
if dataset.data.x is None:
if "REDDIT" not in args.dataset:
max_degree = 0
degs = []
for data in dataset:
degs += [degree(data.edge_index[0], dtype=torch.long)]
max_degree = max(max_degree, degs[-1].max().item())
if max_degree < 1000:
dataset.transform = T.OneHotDegree(max_degree)
else:
deg = torch.cat(degs, dim=0).to(torch.float)
mean, std = deg.mean().item(), deg.std().item()
dataset.transform = NormalizedDegree(mean, std)
else:
feature_dim = 0
degrees = []
for g in dataset:
feature_dim = max(feature_dim, degree(g.edge_index[0]).max().item())
degrees.extend(degree(g.edge_index[0]).tolist())
MAX_DEGREES = 400
oversize = 0
for d, n in Counter(degrees).items():
if d > MAX_DEGREES:
oversize += n
# print(f"N > {MAX_DEGREES}, #NUM: {oversize}, ratio: {oversize/sum(degrees):.8f}")
feature_dim = min(feature_dim, MAX_DEGREES)
feature_dim += 1
for i, g in enumerate(dataset):
degrees = degree(g.edge_index[0])
degrees[degrees > MAX_DEGREES] = MAX_DEGREES
degrees = torch.Tensor([int(x) for x in degrees.numpy().tolist()])
feat = F.one_hot(degrees.to(torch.long), num_classes=int(feature_dim)).float()
g.x = feat
dataset[i] = g
num_tasks = dataset.num_classes
num_features = dataset.num_features
num_dataset = len(dataset)
all_loader = DataLoader(dataset, batch_size=args.batch_size)
return dataset, all_loader, num_tasks, num_features, num_dataset
datasets, all_loader, num_tasks, num_features, num_dataset = load_data(args, transform)
device = torch.device('cuda:{}'.format(args.devices) if torch.cuda.is_available() else 'cpu')
model = CGT(fea_dim=num_features, channels=args.channels, num_layers=args.num_layers,
num_tasks=num_tasks, num_features=num_features).to(device)
model_new = CGT(fea_dim=num_features, channels=args.channels, num_layers=args.num_layers,
num_tasks=num_tasks, num_features=num_features).to(device)
contrast_model = DualBranchContrast(loss=L.InfoNCE(tau=0.2), mode='G2G').to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-5)
# MLP
projector = torch.nn.Sequential(
Linear(args.channels, args.channels),
ReLU(inplace=True),
Linear(args.channels, args.channels)).to(device)
def data_aug(x, batch, pe, edge_index, method):
x1, edge_index1 = x, edge_index
x2, edge_index2 = x, edge_index
ratio = args.aug_ratio
if method == "TokenMask":
h, mask = to_dense_batch(x, batch)
B = h.shape[0]
node_num = h.shape[1]
features = h.shape[2]
node_score = model.gumbel(h)
node_score = node_score.reshape(B, node_num)
node_mask = gumbel_softmax(node_score, device=device, rate=1 - ratio, tau=1, hard=True)
node_mask[:, 0] = 1.
node_mask = node_mask.expand(features, -1, -1).permute(1, 2, 0)
h = h * node_mask
x2 = h[mask]
elif method == "FeatureMask":
aug1 = A.FeatureMasking(pf=ratio)
aug2 = A.FeatureMasking(pf=ratio)
x1, edge_index1, _ = aug1(x, edge_index)
x2, edge_index2, _ = aug2(x, edge_index)
elif method == 'PEMask':
mask_num = int(pe.size(1) * ratio)
mask_col = random.sample(range(pe.size(1)), mask_num)
mask = torch.ones_like(pe)
mask[:, mask_col] = 0
pe = pe * mask
elif method == 'MAE':
pretrain_aug = torch.load(f"mae_model/{args.dataset}.pt")
pretrain_aug.to(device)
pretrain_aug.eval()
_, _, gene_x = pretrain_aug(x1, edge_index1)
num_nodes = gene_x.shape[0] * 0.2
ran_idx = torch.randperm(gene_x.shape[0])[:int(num_nodes)]
x1[ran_idx] = gene_x[ran_idx]
return x1, edge_index1, x2, edge_index2, pe
def model_aug(method):
ratio = args.aug_ratio
model_new.load_state_dict(model.state_dict())
if method == "Gaussian":
std = 0.01
mean = 0.0
for param in model_new.parameters():
if param.requires_grad:
noise = torch.normal(mean=mean, std=std, size=param.size()).to(device)
param.data.add_(noise)
elif method == "DropWeight":
parameters_to_prune = []
for name, module in model_new.named_modules():
if not list(module.children()):
for param_name, param in module.named_parameters():
if param_name == "weight":
parameters_to_prune.append((module, param_name))
for module, param_name in parameters_to_prune:
prune.l1_unstructured(module, param_name, amount=ratio)
prune.remove(module, param_name)
elif method == "DropPath":
model_new.drop_path_prob = ratio
model_new.drop_path = DropPath(model_new.drop_path_prob)
elif method == "DropHead":
for module in model_new.convs:
module.attn_drophead = ratio
def train_data(loader, data_method, model_method):
model.train()
total_loss = 0
for data in loader:
data = data.to(device)
optimizer.zero_grad()
x1, edge_index1, x2, edge_index2, pe = data_aug(data.x, data.batch, data.pe, data.edge_index, data_method)
_, g1, g2 = model(data.x, x1, x2, edge_index1, edge_index2, pe, data.edge_index, data.edge_attr,
data.batch)
g1, g2 = [projector(g) for g in [g1, g2]]
train_loss = contrast_model(g1=g1, g2=g2, batch=data.batch)
train_loss.backward()
total_loss += train_loss.item()
optimizer.step()
return total_loss
def train_model(loader, data_method, model_method):
model.train()
total_loss = 0
model_aug(model_method)
for data in loader:
data = data.to(device)
optimizer.zero_grad()
g1, _, _ = model(data.x, data.x, data.x, data.edge_index, data.edge_index, data.pe, data.edge_index,
data.edge_attr, data.batch)
g2, _, _ = model_new(data.x, data.x, data.x, data.edge_index, data.edge_index, data.pe, data.edge_index,
data.edge_attr, data.batch)
g1, g2 = [projector(g) for g in [g1, g2]]
g2 = Variable(g2.detach().data, requires_grad=False)
train_loss = contrast_model(g1=g1, g2=g2, batch=data.batch)
train_loss.backward()
total_loss += train_loss.item()
optimizer.step()
return total_loss
def train_cross(loader, data_method, model_method):
model.train()
total_loss = 0
model_aug(model_method)
for data in loader:
data = data.to(device)
optimizer.zero_grad()
x1, edge_index1, x2, edge_index2, pe = data_aug(data.x, data.batch, data.pe, data.edge_index, data_method)
_, g1, g2 = model(data.x, x1, x2, edge_index1, edge_index2, pe, data.edge_index, data.edge_attr,
data.batch)
_, g3, g4 = model_new(data.x, x1, x2, edge_index1, edge_index2, pe, data.edge_index, data.edge_attr,
data.batch)
g1, g2, g3, g4 = [projector(g) for g in [g1, g2, g3, g4]]
g3 = Variable(g3.detach().data, requires_grad=False)
g4 = Variable(g4.detach().data, requires_grad=False)
train_loss = contrast_model(g1=g1, g2=g2, batch=data.batch)
train_loss += contrast_model(g1=g3, g2=g4, batch=data.batch)
train_loss += contrast_model(g1=g1, g2=g3, batch=data.batch)
train_loss += contrast_model(g1=g2, g2=g4, batch=data.batch)
train_loss.backward()
total_loss += train_loss.item()
optimizer.step()
return total_loss
@torch.no_grad()
def test():
model.eval()
x = []
y = []
for data in all_loader:
data = data.to(device)
optimizer.zero_grad()
g, _, _ = model(data.x, data.x, data.x, data.edge_index, data.edge_index, data.pe, data.edge_index,
data.edge_attr, data.batch)
x.append(g)
y.append(data.y)
X = torch.cat(x, dim=0)
Y = torch.cat(y, dim=0)
X = X.detach().cpu().numpy()
Y = Y.detach().cpu().numpy()
acc, std = svc_classify(args.fold, X, Y, search=True)
return acc, std
if __name__ == "__main__":
run_name = f"{args.dataset}"
best_loss = 1e9
accs = []
args.save_path = f"exps/{run_name}-{now}"
os.makedirs(os.path.join(args.save_path, str(args.seed)), exist_ok=True)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available():
# cudnn.deterministic = True
torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
if args.aug == "data":
train = train_data
print("Start data aug pretrain")
elif args.aug == "model":
train = train_model
print("Start model aug pretrain")
elif args.aug == "cross":
train = train_cross
print("Start cross aug pretrain")
pretrained_bar = tqdm(total=args.epochs, position=0)
for epoch in range(1, args.epochs + 1):
loss = train(all_loader, args.data_method, args.model_method)
# val_acc = test()
# state_dict = {"model": model.state_dict(), "optimizer": optimizer.state_dict(), "epoch": epoch}
if loss < best_loss:
best_loss = loss
state_dict = {"model": model.state_dict(), "optimizer": optimizer.state_dict(), "epoch": epoch}
torch.save(state_dict, os.path.join(args.save_path, str(args.seed), "best_model.pt"))
pretrained_bar.set_postfix({'loss': loss})
pretrained_bar.set_description(f'Epoch {epoch}')
pretrained_bar.update(1)
pretrained_bar.close()
state_dict = torch.load(os.path.join(args.save_path, str(args.seed), "best_model.pt"))
model.load_state_dict(state_dict["model"])
# unsupervised test
acc, std = test()
print('Acc: {:.4f} ± {:.4f}'.
format(acc, std))
save_accs(args, acc, std)