-
Notifications
You must be signed in to change notification settings - Fork 15
/
main.py
447 lines (369 loc) · 16.5 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
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
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
import sys
import os
import torch
import random
import numpy as np
from tqdm import tqdm
from torch.autograd import Variable
from torch.nn.parameter import Parameter
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import math
import pdb
from DGCNN_embedding import DGCNN
from mlp_dropout import MLPClassifier
from sklearn import metrics
from gcn import *
#from tensorboard_logger import configure,log_value
import gpu
from setproctitle import *
import time
np.set_printoptions(threshold=np.inf)
torch.set_printoptions(precision=2,threshold=float('inf'))
sys.path.append('%s/pytorch_structure2vec-master/s2v_lib' % os.path.dirname(os.path.realpath(__file__)))
from embedding import EmbedMeanField, EmbedLoopyBP
from util import cmd_args, load_data,create_process_name
args=cmd_args
if args.init_from!='':
tmp=args.init_from
state_dict=torch.load(args.init_from)
args=state_dict['args']
args.init_from=tmp
pname=create_process_name()
setproctitle(pname)
if not os.path.exists(args.savedir):
os.makedirs(args.savedir)
if not os.path.exists(args.logdir):
os.makedirs(args.logdir)
save_path=os.path.join(args.savedir,pname)
log_path=os.path.join(args.logdir,pname+'.txt')
if os.path.exists(save_path):
os.system('rm -rf '+save_path)
os.makedirs(save_path)
if not args.print:
f=open(log_path,'a+')
sys.stderr=f
sys.stdout=f
train_graphs, test_graphs = load_data()
gpu.find_idle_gpu(args.gpu)
class Classifier(nn.Module):
def __init__(self):
super(Classifier, self).__init__()
self.rank_loss=args.rank_loss
self.model=args.model
self.eps=args.eps
if args.pool=='mean':
self.pool=self.mean_pool
elif args.pool=='max':
self.pool=self.max_pool
if self.model=='gcn':
self.num_layers=args.gcn_layers
self.gcns=nn.ModuleList()
x_size=args.input_dim
for _ in range(self.num_layers):
self.gcns.append(GCNBlock(x_size,args.hidden_dim,args.bn,args.gcn_res,args.gcn_norm,args.dropout,args.relu))
x_size=args.hidden_dim
self.mlp=MLPClassifier(args.hidden_dim,args.mlp_hidden,args.num_class,args.mlp_layers,args.dropout)
else:
self.margin=args.margin
self.agcn_res=args.agcn_res
self.single_loss=args.single_loss
self.num_layers=args.num_layers
if args.arch==1:
assert args.gcn_layers%self.num_layers==0
gcn_layer_list=[args.gcn_layers//self.num_layers]*self.num_layers
elif args.arch==2:
gcn_layer_list=[args.gcn_layers]+[1]*(self.num_layers-1)
self.agcns=nn.ModuleList()
x_size=args.input_dim
for i in range(args.num_layers):
self.agcns.append(AGCNBlock(args,x_size,args.hidden_dim,gcn_layer_list[i],args.dropout,args.relu))
x_size=self.agcns[-1].pass_dim
if args.model=='diffpool':
args.diffpool_k=int(math.ceil(args.diffpool_k*args.percent))
self.mlps=nn.ModuleList()
if not args.concat:
for i in range(args.num_layers):
self.mlps.append(MLPClassifier(input_size=args.hidden_dim, hidden_size=args.mlp_hidden, num_class=args.num_class,num_layers=args.mlp_layers,dropout=args.dropout))
else:
self.mlps=MLPClassifier(input_size=args.hidden_dim*self.num_layers, hidden_size=args.mlp_hidden, num_class=args.num_class,num_layers=args.mlp_layers,dropout=args.dropout)
def PrepareFeatureLabel(self, batch_graph):
batch_size = len(batch_graph)
labels = torch.LongTensor(batch_size)
max_node_num = 0
for i in range(batch_size):
labels[i] = batch_graph[i].label
max_node_num = max(max_node_num, batch_graph[i].num_nodes)
#print('tags:',batch_graph[i].node_tags)
masks = torch.zeros(batch_size, max_node_num)
adjs = torch.zeros(batch_size, max_node_num, max_node_num)
if batch_graph[0].node_tags is not None:
node_tag_flag = True
batch_node_tag = torch.zeros(batch_size, max_node_num, args.feat_dim)
else:
node_tag_flag = False
if batch_graph[0].node_features is not None:
node_feat_flag = True
batch_node_feat = torch.zeros(batch_size, max_node_num, args.attr_dim)
else:
node_feat_flag = False
for i in range(batch_size):
cur_node_num = batch_graph[i].num_nodes
if node_tag_flag == True:
tmp_tag_idx = torch.LongTensor(batch_graph[i].node_tags).view(-1, 1)
tmp_node_tag = torch.zeros(cur_node_num, args.feat_dim)
tmp_node_tag.scatter_(1, tmp_tag_idx, 1)
batch_node_tag[i, 0:cur_node_num] = tmp_node_tag
#node attribute feature
if node_feat_flag == True:
tmp_node_fea = torch.from_numpy(batch_graph[i].node_features).type('torch.FloatTensor')
batch_node_feat[i, 0:cur_node_num] = tmp_node_fea
#adjs
adjs[i, 0:cur_node_num, 0:cur_node_num] = batch_graph[i].adj
#masks
masks[i,0:cur_node_num] = 1
#cobime the two kinds of node feature
if node_feat_flag == True:
node_feat = batch_node_feat.clone()
if node_feat_flag and node_tag_flag:
# concatenate one-hot embedding of node tags (node labels) with continuous node features
node_feat = torch.cat([batch_node_tag.type_as(node_feat), node_feat], 1)
elif node_feat_flag == False and node_tag_flag == True:
node_feat = batch_node_tag
elif node_feat_flag == True and node_tag_flag == False:
pass
else:
node_feat = torch.ones(batch_size,max_node_num,1) # use all-one vector as node features
if args.mode == 'gpu':
node_feat = node_feat.cuda()
labels = labels.cuda()
adjs = adjs.cuda()
masks = masks.cuda()
return node_feat, labels, adjs, masks
def forward(self,batch_graph,is_print=False):
'''
node_feat: FloatTensor, [batch,max_node_num,input_dim]
labels: LongTensor, [batch]
adj: FloatTensor, [batch,max_node_num,max_node_num]
mask: FloatTensor, [batch,max_node_num]
'''
node_feat, labels, adj,mask = self.PrepareFeatureLabel(batch_graph)
# print('node_feat:',node_feat.type(),node_feat.shape,node_feat)
# print('labels:',labels.type(),labels.shape,labels)
# print('adj:',labels.type(),adj.shape,adj)
# print('mask:',labels.type(),mask.shape,mask)
if self.model=='gcn':
return self.gcn_forward(node_feat,labels,adj,mask)
else:
return self.agcn_forward(node_feat,labels,adj,mask,is_print=is_print)
def mean_pool(self,x,mask):
return x.sum(dim=1)/(self.eps+mask.sum(dim=1,keepdim=True))
@staticmethod
def max_pool(x,mask):
#output: [batch,x.shape[2]]
m=(mask-1)*1e10
r,_=(x+m.unsqueeze(2)).max(dim=1)
return r
def gcn_forward(self,node_feat,labels,adj,mask):
X=node_feat
vis=[]
for i in range(self.num_layers):
X=self.gcns[i](X,adj,mask)
if args.save_feat and not self.training:
vis.append(X.cpu())
embed=self.pool(X,mask)
if args.save_feat and not self.training:
vis.append(mask.cpu())
vis.append(embed.cpu())
vis.append(labels.cpu())
vis=vis[::-1]
logits,_,loss,acc=self.mlp(embed,labels)
return logits,loss,acc,acc,None,None,vis
def agcn_forward(self,node_feat,labels,adj,mask,is_print=False):
# node_feat, labels = self.PrepareFeatureLabel(batch_graph)
cls_loss=node_feat.new_zeros(self.num_layers)
rank_loss=node_feat.new_zeros(self.num_layers-1)
X=node_feat
p_t=[]
pred_logits=0
visualize_tools=[]
visualize_tools1=[labels.cpu()]
embeds=0
concats=[]
layer_acc=[]
for i in range(self.num_layers):
embed,X,adj,mask,visualize_tool=self.agcns[i](X,adj,mask,is_print=is_print)
embeds=embeds+embed
concats.append(embed)
visualize_tools.append(visualize_tool)
if args.save_feat and not self.training:
visualize_tools1.append([embed.cpu(),X.cpu(),mask.cpu()])
if args.concat:
continue
if not self.agcn_res:
logits,softmax_logits,loss,acc=self.mlps[i](embed,labels)
else:
logits,softmax_logits,loss,acc=self.mlps[i](embeds,labels)
layer_acc.append(acc)
pred_logits=pred_logits+softmax_logits
cls_loss[i]=loss
if args.concat:
logits,softmax_logits,loss,acc=self.mlps(torch.cat(concats,dim=1),labels)
pred_logits=softmax_logits
pred=pred_logits.data.max(1)[1]
return logits,loss,acc,acc,None,visualize_tools,visualize_tools1
pred=pred_logits.data.max(1)[1]
avg_acc = pred.eq(labels.data.view_as(pred)).cpu().sum().item() / float(labels.size()[0])
if is_print:
if self.training:
print('training sample loss')
else:
print('test sample loss')
print('cls_loss:',cls_loss)
print('rank_loss:',rank_loss)
if self.single_loss:
cls_loss=cls_loss[-1]
avg_acc=acc
if self.rank_loss:
loss=cls_loss.mean()+rank_loss.mean()
else:
loss=cls_loss.mean()
return logits,loss,acc,avg_acc,layer_acc,visualize_tools,visualize_tools1
def loop_dataset(g_list, classifier, sample_idxes, epoch,optimizer=None, bsize=50):
total_loss = []
total_layer_acc = []
total_iters = (len(sample_idxes) + (bsize - 1) * (optimizer is None)) // bsize
pbar = range(total_iters)
# pbar = tqdm(range(total_iters), unit='batch')
all_targets = []
all_scores = []
n_samples = 0
visual_pos=[int(x) for x in args.sample.strip().split(',')]
vis1=[]
for pos in pbar:
selected_idx = sample_idxes[pos * bsize : (pos + 1) * bsize]
batch_graph = [g_list[idx] for idx in selected_idx]
targets = [g_list[idx].label for idx in selected_idx]
all_targets += targets
if (not classifier.training) and (pos in visual_pos) and args.model!='gcn':
print('=======================test minibatch:',pos,'==================================')
logits, loss, acc,avg_acc,layer_acc,visualize_tools,visualize_tools1 = classifier(batch_graph,is_print=(pos in visual_pos))
vis1.append(visualize_tools1)
all_scores.append(logits[:, 1].detach()) # for binary classification
if epoch%args.save_freq==0 and (not classifier.training) and args.save and (pos in visual_pos) and args.model!='gcn':
visualize_tools=list(zip(*visualize_tools))
visualize_tools=[[x.detach().cpu().numpy() for x in y] for y in visualize_tools]
np.save(os.path.join(save_path,'sample%03d_epoch%03d.npy'%(pos,epoch)),[batch_graph[0].g,batch_graph[0].node_tags]+visualize_tools)
if optimizer is not None:
optimizer.zero_grad()
loss.backward()
if args.clip:
torch.nn.utils.clip_grad_norm_(classifier.parameters(),args.max_grad_norm)
optimizer.step()
loss = loss.data.cpu().numpy()
# pbar.set_description('loss: %0.5f acc: %0.5f' % (loss, acc) )
total_loss.append( np.array([loss, acc,avg_acc]) * len(selected_idx))
if not args.concat and args.model!='gcn':
total_layer_acc.append( np.array(layer_acc) * len(selected_idx))
n_samples += len(selected_idx)
if optimizer is None:
assert n_samples == len(sample_idxes)
total_loss = np.array(total_loss)
avg_loss = np.sum(total_loss, 0) / n_samples
if not args.concat and args.model!='gcn':
total_layer_acc= np.array(total_layer_acc)
avg_layer_acc = np.sum(total_layer_acc, 0) / n_samples
all_scores = torch.cat(all_scores).cpu().data.numpy()
# np.savetxt('test_scores.txt', all_scores) # output test predictions
all_targets = np.array(all_targets)
fpr, tpr, _ = metrics.roc_curve(all_targets, all_scores, pos_label=1)
auc = metrics.auc(fpr, tpr)
avg_loss = np.concatenate((avg_loss, [auc]))
if not args.concat and args.model!='gcn':
return avg_loss,avg_layer_acc,vis1
else:
return avg_loss,[-1]*args.num_layers,vis1
def main():
print(args)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
print('# train: %d, # test: %d' % (len(train_graphs), len(test_graphs)))
classifier = Classifier()
print(classifier)
for n,p in classifier.named_parameters():
print(n,p.type(),p.shape)
if args.mode == 'gpu':
classifier = classifier.cuda()
optimizer = optim.Adam(classifier.parameters(), lr=args.lr)
train_idxes = list(range(len(train_graphs)))
best_loss = None
best_acc=float('-inf')
best_avg_acc=float('-inf')
best_overall_acc=float('-inf')
start_epoch=0
best_epoch=0
best_avg_epoch=0
if args.init_from!='':
classifier.load_state_dict(state_dict['model_state_dict'])
optimizer.load_state_dict(state_dict['optim_state_dict'])
start_epoch=state_dict['epoch']
best_overall_acc=state_dict['best_overall_acc']
dummy_idxes=list(range(len(train_graphs)))
for _ in range(start_epoch):
random.shuffle(dummy_idxes)
p=0
for epoch in range(start_epoch+1,args.epochs):
if args.decay==1 and epoch==args.epochs-args.patient:
for pg in optimizer.param_groups:
tmp=pg['lr']=pg['lr']*0.1
print('===>>lr decay to %f'%tmp)
if args.decay==2 and p>=args.patient:
for pg in optimizer.param_groups:
tmp=pg['lr']=pg['lr']*0.1
print('===>>lr decay to %f'%tmp)
p=0
start_time=time.time()
random.shuffle(train_idxes)
classifier.train()
avg_loss,avg_layer_acc,vis = loop_dataset(train_graphs, classifier, train_idxes, epoch,optimizer=optimizer,bsize=args.bsize)
if not args.printAUC:
avg_loss[3] = 0.0
print('=====>average training of epoch %d: loss %.5f acc %.5f avg_acc %.5f auc %.5f' % (epoch, avg_loss[0], avg_loss[1], avg_loss[2],avg_loss[3]))
# log_value('train acc',avg_loss[1],epoch)
classifier.eval()
test_loss,test_layer_acc,vis = loop_dataset(test_graphs, classifier, list(range(len(test_graphs))),epoch,bsize=args.test_bsize)
if best_acc<test_loss[1]:
best_acc=test_loss[1]
best_epoch=epoch
if best_avg_acc<test_loss[2]:
best_avg_acc=test_loss[2]
best_avg_epoch=epoch
p=0
else:
p+=1
if max(best_acc,best_avg_acc)>best_overall_acc:
best_overall_acc=max(best_acc,best_avg_acc)
torch.save({'model_state_dict':classifier.state_dict(),
'optim_state_dict':optimizer.state_dict(),
'args':args,
'epoch':epoch,
'best_overall_acc':best_overall_acc},
os.path.join(save_path,'best_model.pth'))
torch.save(vis,os.path.join(save_path,'best_feature.pth'))
if not args.printAUC:
test_loss[3] = 0.0
s='=====>average test of epoch %d: loss %.5f acc %.5f avg_acc %.5f best acc %.5f(%d) %.5f(%d) time:%.0fs'%(epoch, test_loss[0], test_loss[1],test_loss[2], best_acc,best_epoch,best_avg_acc,best_avg_epoch,time.time()-start_time)
for k in range(args.num_layers):
s+=' layer%d %.4f'%(k,test_layer_acc[k])
print(s)
if args.model=='agcn' and args.tau>0:
for k in range(classifier.num_layers):
print('layer%d: tau=%.5f, lamda1=%.5f lamda2=%.5f'%(k,classifier.agcns[k].tau.item(),classifier.agcns[k].lamda1.item(),classifier.agcns[k].lamda2.item()))
with open(os.path.join(args.logdir,'acc_results.txt'), 'a+') as f:
f.writelines(pname+': '+'%.4f(%d) %.4f(%d)\n'%(best_acc,best_epoch,best_avg_acc,best_avg_epoch))
if __name__ == '__main__':
main()
if not args.print:
f.close()