forked from boostcampaitech2/final-project-level3-nlp-10
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
407 lines (326 loc) · 16.4 KB
/
train.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
"""
Meta Pseudo Labeling이 구현된 코드입니다.
Reference: https://github.com/kekmodel/MPL-pytorch
"""
import os
import gc
import math
import random
import argparse
import pandas as pd
import numpy as np
from torch.cuda import default_stream
from sklearn.metrics import f1_score
import torch
from torch.optim.lr_scheduler import LambdaLR
import torch.nn.functional as F
from torch.utils.data.dataloader import DataLoader
from transformers import ElectraForSequenceClassification
from tokenizers import BertWordPieceTokenizer
import modeling
from utils import Config, set_seed, GOOGLE_APPLICATION_CREDENTIAL, MLFLOW_TRACKING_URI
from data import load_dataset, punctuation, punctuation2, tokenized_dataset
from tqdm import trange, tqdm
import mlflow
os.environ['GOOGLE_APPLICATION_CREDENTIALS']=GOOGLE_APPLICATION_CREDENTIAL
os.environ['MLFLOW_TRACKING_URI']=MLFLOW_TRACKING_URI
set_seed(42)
def seed_init_fn(x):
seed = 42 + x
np.random.seed(seed)
random.seed(seed)
torch.manual_seed(seed)
return
def get_cosine_schedule_with_warmup(optimizer,
num_warmup_steps,
num_training_steps,
num_wait_steps=0,
num_cycles=0.5,
last_epoch=-1):
def lr_lambda(current_step):
if current_step < num_wait_steps:
return 0.0
if current_step < num_warmup_steps + num_wait_steps:
return float(current_step) / float(max(1, num_warmup_steps + num_wait_steps))
progress = float(current_step - num_warmup_steps - num_wait_steps) / \
float(max(1, num_training_steps - num_warmup_steps - num_wait_steps))
return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
return LambdaLR(optimizer, lr_lambda, last_epoch)
def train(args, tokenizer, device) -> None:
config = Config(
dropout1=args.dropout1,
dropout2=args.dropout2,
label_smoothing=args.label_smoothing,
epochs=args.epochs,
embedding_dim=args.embedding_dim,
hidden_size=args.hidden_size)
# Print Hyperparameters
print(f'config : {config.__dict__}')
mlflow.log_params(config.__dict__)
# Train Dataset
df = pd.read_csv('labeled.csv')
p_df = pd.read_csv('twitch.csv')
eval_df = pd.read_csv('test2.csv')
# pseudo labeling할 데이터 중 2.8만개를 샘플로 사용합니다.
true_label = p_df[(p_df['none']<p_df['curse'])==True]
false_label = p_df.drop(true_label.index, axis=0).reset_index().drop(['index'], axis=1)
false_label = false_label.sample(frac=args.unlabeled_sample_frac, random_state=args.seed)
true_label = true_label.sample(frac=0.25)
true_label = true_label.append(false_label)
true_label = true_label.sample(frac=1, random_state=42).reset_index().drop(['index'], axis=1)
p_df = true_label
# weak augmentation
p_df = punctuation(p_df)
# strong augmentation
a_df = punctuation2(p_df['text'])
labels = list(df['label'])
eval_labels = list(eval_df['label'])
print(f'Test labels 0 : {eval_labels.count(0)}, 1 : {eval_labels.count(1)}')
df = tokenized_dataset(tokenizer, df)
p_df = tokenized_dataset(tokenizer, p_df)
a_df = tokenized_dataset(tokenizer, a_df)
eval_df = tokenized_dataset(tokenizer, eval_df)
dataset = load_dataset(df, labels)
p_dataset = load_dataset(p_df)
a_dataset = load_dataset(a_df)
eval_dataset = load_dataset(eval_df, eval_labels)
batch_size = 32
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True, worker_init_fn=seed_init_fn)
p_dataloader = DataLoader(p_dataset, batch_size=batch_size, shuffle=True, worker_init_fn=seed_init_fn)
a_dataloader = DataLoader(a_dataset, batch_size=batch_size, shuffle=True, worker_init_fn=seed_init_fn)
eval_dataloader = DataLoader(eval_dataset, batch_size=batch_size, shuffle=False)
# Load teacher model(pretrained), studentmodel
teacher = ElectraForSequenceClassification.from_pretrained('jiho0304/curseELECTRA')
vocab_size = args.vocab_size
print(f'vocab size = {vocab_size}')
student = modeling.Model(
vocab_size=vocab_size,
embedding_dim=config.embedding_dim,
hidden_size=config.hidden_size,
num_class=2,
dropout1=config.dropout1,
dropout2=config.dropout2)
student.to(device)
teacher.to(device)
# Set teacher, student's optimizer
criterion = torch.nn.CrossEntropyLoss(label_smoothing=config.label_smoothing)
optimizer_s = torch.optim.SGD(student.parameters(), lr=args.teacher_learning_rate)
optimizer_t = torch.optim.SGD(teacher.parameters(), lr=args.student_learning_rate)
scaler_s = torch.cuda.amp.GradScaler()
scaler_t = torch.cuda.amp.GradScaler()
scheduler_t = get_cosine_schedule_with_warmup(
optimizer=optimizer_t, num_warmup_steps=0, num_training_steps=len(p_dataloader))
scheduler_s = get_cosine_schedule_with_warmup(
optimizer=optimizer_s, num_warmup_steps=0, num_training_steps=len(p_dataloader))
# Meta Pseudo Labeling
print('------Start Training------')
torch.cuda.empty_cache()
gc.collect()
best_f1 = 0
prev_f1, patient = -1, 0
for step in trange(len(p_dataloader) * args.epochs):
teacher.train()
student.train()
labeled = next(iter(dataloader))
unlabeled = next(iter(p_dataloader))
a_labeled = next(iter(a_dataloader))
# labeled data
l_input = labeled['input_ids']
l_attention_mask = labeled['attention_mask']
targets = labeled['label'].to(device)
# reference의 strong augmentation을 augmentation한 unlabeled dataset으로 가정
a_input = a_labeled['input_ids']
a_attention_mask = a_labeled['attention_mask']
# reference의 weak augmentation을 unlabeled dataset으로 가정
u_input = unlabeled['input_ids']
u_attention_mask = unlabeled['attention_mask']
with torch.cuda.amp.autocast():
# teacher model에 먹일 input 구성 (labeled, augmention, unlabeled)
t_input_ids = torch.cat((l_input, a_input, u_input)).to(device)
t_attention_mask = torch.cat((l_attention_mask, a_attention_mask, u_attention_mask)).to(device)
t_logits = teacher(input_ids=t_input_ids, attention_mask=t_attention_mask)['logits']
t_logits_l = t_logits[:batch_size]
t_logits_a, t_logits_u = t_logits[batch_size:].chunk(2)
# teacher모델의 labeled data에 대한 loss
t_loss_l = criterion(t_logits_l, targets)
# augmentation을 통한 data의 label과 unlabeled data의 로스의 비교(unlabeled data로부터 증강되었기 때문)
soft_pseudo_label = torch.softmax(t_logits_u.detach()/args.temperature, dim=-1)
max_probs, hard_pseudo_label = torch.max(soft_pseudo_label, dim=-1)
mask = max_probs.ge(args.threshold).float()
t_loss_u = torch.mean( # KL.Div loss
-(soft_pseudo_label * torch.log_softmax(t_logits_a, dim=-1)).sum(dim=-1) * mask
)
weight_u = args.uda_lambda * min(1., (step+1)/args.uda_step) # lambda-u, uda_step
t_loss_uda = t_loss_l + weight_u * t_loss_u
# student model에 먹일 input 구성 (labeled, augmention)
s_input_ids = torch.cat((l_input, a_input)).to(device)
s_logits = student(s_input_ids)
s_logits = F.sigmoid(s_logits)
s_logits_l, s_logits_a = s_logits[:batch_size], s_logits[batch_size:]
# 업데이트 되지 않은 student 모델의 labeled data에 대한 로스값(labeled data에 대한 validation)
s_loss_l_old = F.cross_entropy(s_logits_l.detach(), targets)
# augmented data에 대해서 student가 학습
s_loss = criterion(s_logits_a, hard_pseudo_label)
scaler_s.scale(s_loss).backward()
scaler_s.step(optimizer_s)
scaler_s.update()
scheduler_s.step()
with torch.cuda.amp.autocast():
# 업데이트 된 student 모델의 labeled data에 대한 로스
with torch.no_grad():
s_logits_l = student(l_input.to(device))
s_logits_l = F.sigmoid(s_logits_l)
s_loss_l_new = F.cross_entropy(s_logits_l.detach(), targets)
# teacher coefficient : https://github.com/kekmodel/MPL-pytorch/issues/6
dot_product = s_loss_l_old - s_loss_l_new
# compute the teacher's gradient from student's feedback
_, hard_pseudo_label = torch.max(t_logits_a.detach(), dim=-1)
t_loss_mpl = dot_product * F.cross_entropy(t_logits_a, hard_pseudo_label)
t_loss = t_loss_uda + t_loss_mpl # t_loss_uda = t_loss_l + t_loss_unlabeled
scaler_t.scale(t_loss).backward()
scaler_t.step(optimizer_t)
scaler_t.update()
scheduler_t.step()
teacher.zero_grad()
student.zero_grad()
# step마다 Evalution 진행
if step > 0 and step % 10 == 0:
student.eval()
correct, loss = 0, 0
zero, one = 0, 0
prediction = []
with torch.no_grad():
for _, batch in tqdm(enumerate(eval_dataloader)):
data = batch['input_ids'].cuda()
labels = batch['label']
output = student(data)
predicted = torch.max(output,1)[1]
prediction += predicted.tolist()
zero += predicted.tolist().count(0)
one += predicted.tolist().count(1)
correct += (predicted==labels.cuda()).sum()
loss += F.cross_entropy(output, labels.cuda()).item()
eval_f1 = f1_score(eval_labels, prediction, average='macro')
print(f'Epoch: {step+1} | Train Loss : {loss/len(eval_dataloader):.5f} | Test Acc : {correct/len(eval_dataset):.5f} | Zero : {zero} | One : {one} | F1 : {eval_f1:.5f}')
if eval_f1 > best_f1:
# 가장 좋을 때의 모델을 저장합니다.
torch.save(student.state_dict(), f'./save/meta_pseudo/result_temp.pt')
best_f1 = eval_f1
if prev_f1 == eval_f1:
patient += 1
if patient == args.patient:
break
else:
patient = 0
prev_f1 = eval_f1
print(f'best f1 = {best_f1}')
def finetune(tokenizer, device):
"""
MPL이 적용된 student 모델을 다시 labeled data로 Fine tuning합니다.
"""
# Load datasets
df = pd.read_csv('labeled.csv')
eval_df = pd.read_csv('test2.csv')
labels = list(df['label'])
eval_labels = list(eval_df['label'])
df = punctuation(df)
df = tokenized_dataset(tokenizer, df)
eval_df = tokenized_dataset(tokenizer, eval_df)
dataset = load_dataset(df, labels)
eval_dataset = load_dataset(eval_df, eval_labels)
batch_size = args.batch_size
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
eval_dataloader = DataLoader(eval_dataset, batch_size=batch_size, shuffle=False)
# Load model
epochs = args.finetune_epochs
model = modeling.Model(
vocab_size=args.vocab_size,
embedding_dim=args.embedding_dim,
hidden_size=args.hidden_size,
num_class=args.num_classes,
dropout1=args.dropout1,
dropout2=args.dropout2
)
model.load_state_dict(torch.load('./save/meta_pseudo/result_temp.pt'))
model.to(device)
# Set criterion, optimizer, scheduler
criterion = torch.nn.CrossEntropyLoss(label_smoothing=args.label_smoothing)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.finetune_learning_rate)
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer=optimizer,
epochs = epochs,
max_lr=args.finetune_max_lr,
steps_per_epoch=len(dataloader),
pct_start=args.finetune_pct_start,
)
best_f1 = 0
for epoch in range(epochs):
running_loss = 0
model.train()
for i, labeled in enumerate(dataloader):
input = labeled['input_ids'].to(device)
label = labeled['label'].to(device)
output = model(input)
loss = criterion(output, label)
running_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
with torch.no_grad():
# Evalutaion
model.eval()
correct = 0
prediction = []
for j, batch in enumerate(eval_dataloader):
input = batch['input_ids'].cuda()
label = batch['label'].cuda()
output = model(input)
preds = output.argmax(-1)
prediction += preds.tolist()
correct += (preds==label).sum().item()
eval_acc = correct/len(eval_dataset)
f1 = f1_score(eval_labels, prediction, average='macro')
print(f'Epoch: {epoch+1} | Train Loss : {running_loss/len(dataloader):.5f} | Acc : {eval_acc:.5f} | F1 : {f1:.3f}')
mlflow.log_metric('train loss', running_loss/len(dataloader))
mlflow.log_metric('eval acc', eval_acc)
mlflow.log_metric('eval f1', f1)
if f1 > best_f1:
torch.save(model.state_dict(), f'./save/temp/result_{f1:.3f}.pt')
best_model = model.state_dict()
best_f1 = f1
# mlflow tracking model
mlflow.pytorch.log_model(best_model, 'model', registered_model_name="ToxicityText")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Meta Pseudo Labeling(Reference : https://github.com/kekmodel/MPL-pytorch)')
parser.add_argument('--dropout1', default=0.3, type=float, help='dropout embedding layer, linear layer')
parser.add_argument('--dropout2', default=0.4, type=float, help='dropout conv layer')
parser.add_argument('--teacher_learning_rate', default=1e-7, type=float, help='mpl teacher learning rate')
parser.add_argument('--student_learning_rate', default=1e-7, type=float, help='mpl student learning rate')
parser.add_argument('--label_smoothing', default=0, type=float, help='mpl traning label smoothing')
parser.add_argument('--embedding_dim', default=100, type=int, help='model embedding dimension')
parser.add_argument('--hidden_size', default=128, type=int, help='model hidden size')
parser.add_argument('--num_classes', default=2, type=int, help='number of classification')
parser.add_argument('--epochs', default=1, type=int, help='mpl trainig epochs')
parser.add_argument('--seed', default=42, type=int, help='random seed')
parser.add_argument('--vocab_size', default=30000, type=int, help='tokenizer vocab size')
parser.add_argument('--batch_size', default=32, type=int, help='mpl training batch size')
parser.add_argument('--unlabeled_sample_frac', default=0.025, type=float, help='unlabeled dataset sample ratio')
parser.add_argument('--temperature', default=0.9, type=float, help='pseudo label temperature')
parser.add_argument('--uda_lambda', default=1.0, type=float, help='pseudo label weight lambda')
parser.add_argument('--uda_step', default=1.0, type=float, help='pseudo label uda step')
parser.add_argument('--threshold', default=0.6, type=float, help='pseudo label threshold')
parser.add_argument('--patient', default=20, type=int, help='mpl early stopping patient')
parser.add_argument('--finetune_learning_rate', defulat=0.001, type=float, help='finetuning learning rate')
parser.add_argument('--finetune_epochs', defulat=10, type=int, help='finetuning epochs')
parser.add_argument('--finetune_max_lr', defulat=0.01, type=float, help='finetuning OneCyclelr scheduler max_lr')
parser.add_argument('--finetune_pct_start', defulat=0.1, type=float, help='finetuning OncCyclelr scheduler pct_start')
args = parser.parse_args()
# device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f'device = {device}')
# load tokenizer
tokenizer = BertWordPieceTokenizer('./vocab.txt', lowercase=False)
# MPL 수행 후 labeled data에 대해 finetuning을 시도합니다
train(args, tokenizer, device)
finetune(args, tokenizer, device)