-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_bert_gcn.py
345 lines (299 loc) · 15.3 KB
/
train_bert_gcn.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
import torch as th
from transformers import AutoModel, AutoTokenizer
import torch.nn.functional as F
from utils import *
import dgl
import torch.utils.data as Data
from ignite.engine import Events, create_supervised_evaluator, create_supervised_trainer, Engine
from ignite.metrics import Accuracy, Loss
from torchmetrics.classification import MultilabelAccuracy, MultilabelPrecision, MultilabelRecall, MultilabelF1Score
from sklearn.metrics import accuracy_score
import numpy as np
import os
import shutil
import argparse
import sys
import logging
from datetime import datetime
from torch.optim import lr_scheduler
from model import BertGCN, BertGAT
parser = argparse.ArgumentParser()
parser.add_argument('--max_length', type=int, default=128, help='the input length for bert') # bert의 input 최대 길이(128)
parser.add_argument('--batch_size', type=int, default=64) # 배치 크기(64)
parser.add_argument('-m', '--m', type=float, default=0.7, help='the factor balancing BERT and GCN prediction') # BERT와 GCN의 예측을 조절하는 trade-off lambda(0.7)
parser.add_argument('--nb_epochs', type=int, default=30) # 에폭(50)
parser.add_argument('--bert_init', type=str, default='roberta-base',
choices=['roberta-base', 'roberta-large', 'bert-base-uncased', 'bert-large-uncased']) # BERT checkpoint
parser.add_argument('--pretrained_bert_ckpt', default=None) # 사전학습한 BERT의 checkpoint 불러오기(이걸해야 BertGCN의 성능이 유의미하다고 한 듯)
parser.add_argument('--dataset', default='psysym', choices=['20ng', 'R8', 'R52', 'ohsumed', 'mr','psysym']) # 데이터셋 선택(20ng)
parser.add_argument('--checkpoint_dir', default=None, help='checkpoint directory, [bert_init]_[gcn_model]_[dataset] if not specified') # BertGCN 모델을 저장할 위치. 이름은 기본 설정이 [bert_init]_[gcn_model]_[dataset]
parser.add_argument('--gcn_model', type=str, default='gcn', choices=['gcn', 'gat']) # GCN을 사용할지, GAT를 사용할지 선택(gcn)
parser.add_argument('--gcn_layers', type=int, default=2) # GCN 레이어 수(2)
parser.add_argument('--n_hidden', type=int, default=200, help='the dimension of gcn hidden layer, the dimension for gat is n_hidden * heads') # GNN이 연산할 때 feature의 차원(200)
parser.add_argument('--heads', type=int, default=8, help='the number of attentionn heads for gat') # GAT 사용할 때만 쓸듯?
parser.add_argument('--dropout', type=float, default=0.5) # dropout(0.5)
parser.add_argument('--gcn_lr', type=float, default=1e-3) # gcn의 learning rate(1e-3)
parser.add_argument('--bert_lr', type=float, default=1e-5) # bert의 learning rate(1e-5)
args = parser.parse_args()
max_length = args.max_length
batch_size = args.batch_size
m = args.m
nb_epochs = args.nb_epochs
bert_init = args.bert_init
pretrained_bert_ckpt = args.pretrained_bert_ckpt
dataset = args.dataset
checkpoint_dir = args.checkpoint_dir
gcn_model = args.gcn_model
gcn_layers = args.gcn_layers
n_hidden = args.n_hidden
heads = args.heads
dropout = args.dropout
gcn_lr = args.gcn_lr
bert_lr = args.bert_lr
if checkpoint_dir is None: # 경로 따로 지정안하면 경로와 파일명이 기본으로 결정됨
ckpt_dir = './checkpoint/{}_{}_{}'.format(bert_init, gcn_model, dataset)
else: # 아니면 따로 저장
ckpt_dir = checkpoint_dir
os.makedirs(ckpt_dir, exist_ok=True) # 경로 생성
shutil.copy(os.path.basename(__file__), ckpt_dir)
# log 남기기
sh = logging.StreamHandler(sys.stdout)
sh.setFormatter(logging.Formatter('%(message)s'))
sh.setLevel(logging.INFO)
fh = logging.FileHandler(filename=os.path.join(ckpt_dir, 'training.log'), mode='w')
fh.setFormatter(logging.Formatter('%(message)s'))
fh.setLevel(logging.INFO)
logger = logging.getLogger('training logger')
logger.addHandler(sh)
logger.addHandler(fh)
logger.setLevel(logging.INFO)
# device 설정
cpu = th.device('cpu')
gpu = th.device('cuda:0')
logger.info('arguments:')
logger.info(str(args))
logger.info('checkpoints will be saved in {}'.format(ckpt_dir))
# Model
# Data Preprocess
adj, features, y_train, y_val, y_test, train_mask, val_mask, test_mask, train_size, test_size = load_corpus(dataset)
'''
adj: n*n sparse adjacency matrix
y_train, y_val, y_test: n*c matrices
train_mask, val_mask, test_mask: n-d bool array
'''
# compute number of real train/val/test/word nodes and number of classes
nb_node = features.shape[0] # 노드 개수
nb_train, nb_val, nb_test = train_mask.sum(), val_mask.sum(), test_mask.sum() # train, valid, test 데이터 수
nb_word = nb_node - nb_train - nb_val - nb_test # word 수
nb_class = y_train.shape[1] # 예측을 수행할 클래스 수
# instantiate model according to class number
if gcn_model == 'gcn': # gcn 모델을 입력받은 경우
model = BertGCN(nb_class=nb_class, pretrained_model=bert_init, m=m, gcn_layers=gcn_layers,
n_hidden=n_hidden, dropout=dropout) # 예측을 수행할 클래스 수, 시작할 bert checkpoint, gcn 레이어 수, 히든 차원, dropout
else: # gat 모델을 입력받은 경우
model = BertGAT(nb_class=nb_class, pretrained_model=bert_init, m=m, gcn_layers=gcn_layers,
heads=heads, n_hidden=n_hidden, dropout=dropout)
if pretrained_bert_ckpt is not None: # 이 task로 학습시켜둔 BERT 모델이 있는 경우
ckpt = th.load(pretrained_bert_ckpt, map_location=gpu) # 학습시켜둔 모델 불러오기
model.bert_model.load_state_dict(ckpt['bert_model']) # bert model의 가중치 가져오기
model.classifier.load_state_dict(ckpt['classifier']) # bert model의 분류기 가중치 가져오기
# load documents and compute input encodings
corpse_file = './data/corpus/' + dataset +'_shuffle.txt' # 데이터 본문 전체
with open(corpse_file, 'r') as f:
text = f.read()
text = text.replace('\\', '')
text = text.split('\n')
def encode_input(text, tokenizer):
input = tokenizer(text, max_length=max_length, truncation=True, padding='max_length', return_tensors='pt')
# print(input.keys())
return input.input_ids, input.attention_mask # 토큰화
input_ids, attention_mask = encode_input(text, model.tokenizer) # 모델의 토크나이저로 토큰화
input_ids = th.cat([input_ids[:-nb_test], th.zeros((nb_word, max_length), dtype=th.long), input_ids[-nb_test:]]) # (train+word+test, max_length): 각 document의 토큰화 결과
attention_mask = th.cat([attention_mask[:-nb_test], th.zeros((nb_word, max_length), dtype=th.long), attention_mask[-nb_test:]]) # (train+word+test, max_length): 각 document의 마스킹 결과
# transform one-hot label to class ID for pytorch computation
y = y_train + y_test + y_val # 전체 라벨링 결과를 모두 더해줌(자리가 이미 정해져있으므로 순서 상관 없음)
#y_train = y_train.argmax(axis=1) # 학습 데이터의 라벨의 클래스 인덱스 배열(multi label classification에서 수정 필요)
#y = y.argmax(axis=1) # 전체 데이터의 라벨의 클래스 인덱스 배열(multi label classification에서 수정 필요)
# document mask used for update feature
doc_mask = train_mask + val_mask + test_mask # word 제외 모든 document
# build DGL Graph
adj_norm = normalize_adj(adj + sp.eye(adj.shape[0])) # 인접행렬 자기자신에 1 더해준 후 정규화
g = dgl.from_scipy(adj_norm.astype('float32'), eweight_name='edge_weight') # adj_norm을 edge weight로하는 그래프 생성
g.ndata['input_ids'], g.ndata['attention_mask'] = input_ids, attention_mask # (train+word+test, max_length)
g.ndata['label'], g.ndata['train'], g.ndata['val'], g.ndata['test'] = \
th.LongTensor(y), th.FloatTensor(train_mask), th.FloatTensor(val_mask), th.FloatTensor(test_mask) # (train+word+test, ) # g.ndata['label']: 모든 document의 클래스 인덱스
g.ndata['label_train'] = th.LongTensor(y_train) # (real_train_size, ) # g.ndata['label_train']: 학습 document의 클래스 인덱스
g.ndata['cls_feats'] = th.zeros((nb_node, model.feat_dim))
logger.info('graph information:')
logger.info(str(g))
# create index loader
train_idx = Data.TensorDataset(th.arange(0, nb_train, dtype=th.long))
val_idx = Data.TensorDataset(th.arange(nb_train, nb_train + nb_val, dtype=th.long))
test_idx = Data.TensorDataset(th.arange(nb_node-nb_test, nb_node, dtype=th.long))
doc_idx = Data.ConcatDataset([train_idx, val_idx, test_idx]) # 전체 인덱스
# 배치 크기만큼 인덱스 개수를 나눔
idx_loader_train = Data.DataLoader(train_idx, batch_size=batch_size, shuffle=True)
idx_loader_val = Data.DataLoader(val_idx, batch_size=batch_size)
idx_loader_test = Data.DataLoader(test_idx, batch_size=batch_size)
idx_loader = Data.DataLoader(doc_idx, batch_size=batch_size, shuffle=True) # 전체 데이터셋에 대해 나눔
# Training
def update_feature(): # 한 에폭이 끝날 때 마다 수행됨
global model, g, doc_mask
# no gradient needed, uses a large batchsize to speed up the process
dataloader = Data.DataLoader(
Data.TensorDataset(g.ndata['input_ids'][doc_mask], g.ndata['attention_mask'][doc_mask]),
batch_size=1024
) # 전체 데이터셋에 대해 1024 사이즈로 배치를 나눠 데이터 로드
with th.no_grad():
model = model.to(gpu)
model.eval()
cls_list = []
for i, batch in enumerate(dataloader): # 모든 배치의 데이터셋에 대해 cls_feat을 업데이트
input_ids, attention_mask = [x.to(gpu) for x in batch]
output = model.bert_model(input_ids=input_ids, attention_mask=attention_mask)[0][:, 0]
cls_list.append(output.cpu())
cls_feat = th.cat(cls_list, axis=0)
g = g.to(cpu)
g.ndata['cls_feats'][doc_mask] = cls_feat
return g # 그래프 출력
optimizer = th.optim.Adam([
{'params': model.bert_model.parameters(), 'lr': bert_lr},
{'params': model.classifier.parameters(), 'lr': bert_lr},
{'params': model.gcn.parameters(), 'lr': gcn_lr},
], lr=1e-3
)
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[30], gamma=0.1)
threshold = 0.5
class MaskedLoss(th.nn.Module):
def __init__(self, loss_weighting="mean", pos_weight=None, loss_type='bce', focal_gamma=2.):
super(MaskedLoss, self).__init__()
self.loss_weighting = loss_weighting
self.pos_weight = pos_weight
self.loss_type = loss_type
self.focal_gamma = focal_gamma
def forward(self, logits, labels, masks=None):
# treat unlabeled samples(-1) as implict negative (0.)
labels2 = th.clamp_min(labels, 0.)
losses = F.binary_cross_entropy_with_logits(logits, labels2, reduction='none', pos_weight=self.pos_weight)
if self.loss_type == 'focal':
pt = th.exp(-losses) # prevents nans when probability 0
losses = (1-pt) ** self.focal_gamma * losses
if masks is not None:
masked_losses = th.masked_select(losses, masks)
if self.loss_weighting == 'mean':
return masked_losses.mean()
elif self.loss_weighting == 'geo_mean':
masked_losses = masked_losses.mean(0) # loss per task
return th.exp(th.mean(th.log(masked_losses)))
else:
if self.loss_weighting == 'mean':
return losses.mean()
elif self.loss_weighting == 'geo_mean':
losses = losses.mean(0)
return th.exp(torch.mean(th.log(losses)))
def train_step(engine, batch): # 매 배치(일부 document에 대한 feature에 관한 것)마다 반복하는 학습 과정
global model, g, optimizer
model.train()
model = model.to(gpu)
g = g.to(gpu)
optimizer.zero_grad()
(idx, ) = [x.to(gpu) for x in batch]
optimizer.zero_grad()
train_mask = g.ndata['train'][idx].type(th.BoolTensor)
y_pred = model(g, idx)[train_mask] # 학습 데이터에 대해 BertGCN으로 예측한 idx의 값
y_true = g.ndata['label_train'][idx][train_mask] # 학습 데이터 target으로 저장되어 있던 값들 가져오기
#loss = F.nll_loss(y_pred, y_true) # 학습 데이터의 loss계산
y_mask = (y_true>-1)
loss_fn = MaskedLoss()
loss = loss_fn(y_pred, y_true,masks=y_mask)
loss.backward()
optimizer.step() # 파라미터 업데이트
g.ndata['cls_feats'].detach_()
train_loss = loss.item()
with th.no_grad():
if train_mask.sum() > 0:
y_true = y_true.detach().cpu()
y_pred = y_pred.detach().cpu()
#y_pred = y_pred.argmax(axis=1).detach().cpu()
#train_acc = accuracy_score(y_true, y_pred)
metric_fn = MultilabelAccuracy(threshold=threshold, num_labels=nb_class)
y_true[y_true==-1] = 0 # missing label
train_acc = metric_fn(y_pred, y_true)
else:
train_acc = 1
return train_loss, train_acc
trainer = Engine(train_step)
@trainer.on(Events.EPOCH_COMPLETED) # 한 에폭이 끝날 때마다 실행
def reset_graph(trainer):
scheduler.step()
update_feature() # g.ndata['cls_feats'] 중 주어진 인덱스에 대한 값들만 업데이트
th.cuda.empty_cache()
def test_step(engine, batch):
global model, g
with th.no_grad():
model.eval()
model = model.to(gpu)
g = g.to(gpu)
(idx, ) = [x.to(gpu) for x in batch]
y_pred = model(g, idx)
y_true = g.ndata['label'][idx]
return y_pred, y_true
def thresholded_output_transform(output):
y_pred, y = output
y[y==-1] = 0
result = th.zeros_like(y_pred, dtype=th.float16)
result[y_pred >= threshold] = 1
y_pred = result
#print(y_pred, y)
return y_pred, y
def thresholded_output_transform_ml(output):
y_pred, y = output
y_mask = (y>-1)
'''
y[y==-1] = 0
result = th.zeros_like(y_pred, dtype=th.float16)
result[y_pred >= threshold] = 1
y_pred = result
'''
#print(y_pred, y)
return y_pred, y, {'masks':y_mask}
evaluator = Engine(test_step)
metrics={
'accuracy': Accuracy(is_multilabel=True, output_transform=thresholded_output_transform),
'nll': Loss(MaskedLoss(), output_transform=thresholded_output_transform_ml),
}
for n, f in metrics.items():
f.attach(evaluator, n)
@trainer.on(Events.EPOCH_COMPLETED)
def log_training_results(trainer):
evaluator.run(idx_loader_train)
metrics = evaluator.state.metrics
train_acc, train_nll = metrics["accuracy"], metrics["nll"]
evaluator.run(idx_loader_val)
metrics = evaluator.state.metrics
val_acc, val_nll = metrics["accuracy"], metrics["nll"]
evaluator.run(idx_loader_test)
metrics = evaluator.state.metrics
test_acc, test_nll = metrics["accuracy"], metrics["nll"]
logger.info(
"Epoch: {} Train acc: {:.4f} loss: {:.4f} Val acc: {:.4f} loss: {:.4f} Test acc: {:.4f} loss: {:.4f}"
.format(trainer.state.epoch, train_acc, train_nll, val_acc, val_nll, test_acc, test_nll)
)
if val_acc > log_training_results.best_val_acc:
logger.info("New checkpoint")
th.save(
{
'bert_model': model.bert_model.state_dict(),
'classifier': model.classifier.state_dict(),
'gcn': model.gcn.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': trainer.state.epoch,
},
os.path.join(
ckpt_dir, 'checkpoint.pth'
)
)
log_training_results.best_val_acc = val_acc
log_training_results.best_val_acc = 0
g = update_feature()
trainer.run(idx_loader, max_epochs=nb_epochs)