-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
185 lines (163 loc) · 8.38 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
import os
import numpy as np
import random
import torch
import torch.nn as nn
from data_aug import get_aug_data
from transformers import WEIGHTS_NAME, CONFIG_NAME
from transformers import AutoConfig
from transformers import AutoModelForSequenceClassification,AutoTokenizer, get_linear_schedule_with_warmup
from data import process_train, make_batch, split_and_load_dataset, shuffle
from utils import get_logger
from optims import ChildTuningAdamW
from sklearn.metrics import classification_report
logger = get_logger('./logs', __name__)
def set_seed(seed, deterministic=False):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
if deterministic:
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def load_model_and_tokenizer(model_path, device, p_dropout=0.1, n_labels=None):
logger.info("loading model and tokenizer from {} ...".format(model_path))
tokenizer = AutoTokenizer.from_pretrained(model_path)
config = AutoConfig.from_pretrained(model_path)
# if n_labels:
# config.num_labels = n_labels
if hasattr(config, 'hidden_dropout_prob'):
config.hidden_dropout_prob = p_dropout
if hasattr(config, 'classifier_dropout'):
config.classifier_dropout = p_dropout
if hasattr(config, 'attention_probs_dropout_prob'):
config.attention_probs_dropout_prob = p_dropout
# model = AutoModelForSequenceClassification.from_config(config)
model = AutoModelForSequenceClassification.from_pretrained(model_path, config=config)
#freeze paras
# for name, para in model.named_parameters():
# if name.find('embedding') != -1:
# para.requires_grad = False
# for para in model.parameters():
# para.requires_grad = False
# for name, para in model.named_parameters():
# if name.find('layer') != -1:
# para.requires_grad = False
if n_labels:
hidden_size = model.config.hidden_size
if hasattr(model.classifier, 'out_proj'):
model.classifier.out_proj = nn.Linear(in_features=hidden_size, out_features=n_labels, bias=True)
else:
model.classifier = nn.Linear(in_features=hidden_size, out_features=n_labels, bias=True)
#model.classifier = Highway(in_features=hidden_size, out_features=n_labels, func=F.relu, bias=True)
model.config.num_labels = n_labels
model.to(device)
return model, tokenizer
def cal_performance(preds, labels, score_type='recall'):
# print(preds, labels)
# pred_flat = np.argmax(np.array(preds), axis=-1).flatten()
# labels_flat = np.array(labels).flatten()
#
# print(pred_flat, labels_flat)
report = classification_report(labels, preds, zero_division=0, output_dict=True)
acc = report['accuracy']
#score = report['weighted avg']['f1-score']
if score_type == 'macro-f1':
score = report['macro avg']['f1-score']
else:
score = report['macro avg']['recall']
return acc, score
def train_epoch(model, criterion, optim, scheduler, train_loader, val_loader, epoch, train_log_interval=10, val_internal=50, val_res=None, save_dir=None, device=0):
model.train()
len_iter = len(train_loader)
n_step = 0
val_acces, val_fscores, val_losses = [], [], []
for i, batch in enumerate(train_loader, start=1):
optim.zero_grad()
input_ids, attention_mask, labels = make_batch(batch, device)
outputs = model(input_ids, attention_mask=attention_mask)
loss = criterion(outputs.logits, labels)
loss.backward()
optim.step()
n_step += 1
if scheduler:
scheduler.step()
if i % train_log_interval == 0:
logger.info("epoch: %d [%d/%d], loss: %.6f, lr: %.8f, steps: %d" %
(epoch, i, len_iter, loss.item(), optim.param_groups[0]["lr"], n_step + len_iter * (epoch-1)))
if i % val_internal == 0:
acc, score, loss = val_epoch(model, criterion, val_loader, save_dir, val_res, device)
val_acces.append(acc)
val_fscores.append(score)
val_losses.append(loss)
return val_acces, val_fscores, val_losses
def val_epoch(model, criterion, val_loader, save_dir, val_res, device):
model.eval()
total_eval_loss = 0
preds, Labels = [], []
with torch.no_grad():
for batch in val_loader:
input_ids, attention_mask, labels = make_batch(batch, device)
outputs = model(input_ids, attention_mask=attention_mask)
loss = criterion(outputs.logits, labels)
total_eval_loss += loss.item()
batch_preds = torch.argmax(outputs.logits, dim=-1).detach().cpu().tolist()
label_ids = labels.to('cpu').numpy().tolist()
preds.extend(batch_preds)
Labels.extend(label_ids)
avg_val_loss =total_eval_loss / len(val_loader)
acc, score = cal_performance(preds, Labels)
if save_dir:
if score > max(val_res) and acc > 0.76 :
save_model(model, save_dir)
val_res.append(score)
logger.info("Valid | acc: %.4f, score: %.4f, global optim: %.4f, loss: %.4f" % (acc, score, max(val_res), avg_val_loss))
return acc, score, avg_val_loss
def save_model(model, save_dir):
output_model_file = os.path.join(save_dir, WEIGHTS_NAME)
output_config_file = os.path.join(save_dir, CONFIG_NAME)
torch.save(model.state_dict(), output_model_file)
model.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(save_dir)
logger.info('Model has save to %s' % save_dir)
def train(model, criterion, optim, scheduler, train_loader, val_loader, n_epoch, save_dir, device):
val_res = [0]
for i in range(1, n_epoch + 1):
train_epoch(model, criterion, optim, scheduler, train_loader, val_loader, save_dir=save_dir, epoch=i, train_log_interval=10, val_internal=20, val_res=val_res, device=device)
val_epoch(model, criterion, val_loader, save_dir, val_res, device)
if __name__ == '__main__':
device = torch.device("cuda:3" if torch.cuda.is_available() else "cpu")
data, n_labels, cnt = process_train('./train_data.csv') #Counter({6: 17549, 2: 6515, 1: 5651, 5: 4138, 3: 2534, 4: 432})
#data, n_labels = get_aug_data('./eda_to_aug.csv')
criterion = nn.CrossEntropyLoss()
seeds = [222]#[111]#, 222, 333, 444, 555]#, 666, 777, 888, 999, 1111]#, 2222, 3333, 4444, 5555]#
for seed in seeds:
set_seed(seed)
n_epoch = 3
model_name = 'bert'
if model_name == 'bert':
model, tokenizer = load_model_and_tokenizer('./bert-base-chinese', device, p_dropout=0.1, n_labels=n_labels) #p_drop 0.1
elif model_name == 'roberta':
model, tokenizer = load_model_and_tokenizer('./roberta-base-finetuned-dianping-chinese', device,
p_dropout=0.05,
n_labels=n_labels)
elif model_name == 'electra':
model, tokenizer = load_model_and_tokenizer('./chinese-electra-180g-large-discriminator', device,
p_dropout=0.3,
n_labels=n_labels)
elif model_name == 'roberta-jd':
model, tokenizer = load_model_and_tokenizer('./roberta-base-finetuned-jd-full-chinese', device,
p_dropout=0.05,
n_labels=n_labels)
save_dir = './checkpoints/{}/{}'.format(model_name, seed)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
*_, train_loader, val_loader = split_and_load_dataset(data, tokenizer, max_len=84, batch_size=32, test_size=0.1) # 0.05
optim = torch.optim.AdamW(model.parameters(), lr=2e-5) # 5e-6
# optim = ChildTuningAdamW(model.parameters(), lr=2e-5, mode='ChildTuning-D', reserve_p=0.1,
# model=model, criterion=criterion, train_loader=train_loader, device=device)
scheduler = get_linear_schedule_with_warmup(optim, num_warmup_steps=400, num_training_steps=len(train_loader) * n_epoch)
train(model, criterion, optim, scheduler, train_loader, val_loader, n_epoch=n_epoch, save_dir=save_dir, device=device)
del model, tokenizer, optim, scheduler, train_loader, val_loader
torch.cuda.empty_cache()
data = shuffle(data)