-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_pipeline.py
222 lines (202 loc) · 10.9 KB
/
train_pipeline.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
from utils import *
from transformers import BertTokenizer, AdamW, get_scheduler
from tqdm import tqdm
from pymodel.model import BERT_Model
import torch
import os
import argparse
from random import seed
class Trainer:
def __init__(self, config):
self.config = config
self.fix_seed(config.seed)
print(config.__dict__)
self.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
self.tokenizer = BertTokenizer.from_pretrained(config.pretrained_model)
self.train_dataloader = init_dataloader(config.train_path, config, "train", self.tokenizer)
self.valid_dataloader = init_dataloader(config.dev_path, config, "dev", self.tokenizer)
for i in self.valid_dataloader.dataset.data:
print(i)
break
self.test_dataloader = init_dataloader(config.test_path, config, "testtt", self.tokenizer)
self.model = BERT_Model(config)
# self.model = SoftMaskedBert4Csc(cfg=config, device=self.device, tokenizer=self.tokenizer)
# self.model.load_state_dict(torch.load('/home/lixiang/Spell/exps/best/test_new_61.pt'))
# print("model is loaded")
self.model.to(self.device)
self.optimizer = AdamW(self.model.parameters(), lr=config.lr)
self.scheduler = self.set_scheduler()
self.best_score = {"valid-c": 0, "valid-s": 0}
self.best_epoch = {"valid-c": 0, "valid-s": 0}
def fix_seed(self, seed_num):
torch.manual_seed(seed_num)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
seed(seed_num)
def set_scheduler(self):
num_epochs = self.config.num_epochs
num_training_steps = num_epochs * len(self.train_dataloader)
lr_scheduler = get_scheduler(
"linear",
optimizer=self.optimizer,
num_warmup_steps=0,
num_training_steps=num_training_steps
)
return lr_scheduler
def __forward_prop(self, dataloader, back_prop=False):
collected_outputs = []
prob_outputs = []
final_out = []
softmax = torch.nn.Softmax(-1)
f = open(self.config.save_path + '.txt', "w")
for id, batch in tqdm(enumerate(dataloader)):
f.write(str(id + 1) + ':' + '\n')
batch = {k: v.to(self.device) for k, v in batch.items()}
for repeat in range(10):
loss, logits = self.model(input_ids=batch['input_ids'],
attention_mask=batch['attention_mask'],
token_type_ids=batch['token_type_ids'],
pinyin_ids=batch['pinyin_ids']
)
outputs = torch.argmax(logits, dim=-1).cpu()
outputs_prob = torch.max(softmax(logits), dim=-1)[0].cpu()
outputs_prob_k, outputs_index_k = torch.topk(softmax(logits), k=3, dim=-1)
# for outputs_i, prob_i in zip(outputs, outputs_prob):
# collected_outputs.append(outputs_i)
# prob_outputs.append(prob_i)
for prob_i in outputs_prob:
prob_outputs.append(prob_i)
for outputs_line, prob_line, batch_line, prob_k, index_k in zip(outputs, outputs_prob,
batch['input_ids'], outputs_prob_k,
outputs_index_k):
max_prob = 0
threshold = 0.0
for n, (outputs_i, prob_i, batch_i, prob_k_i, index_k_i) in enumerate(
zip(outputs_line, prob_line, batch_line, prob_k, index_k)):
outputs_i = outputs_i.item()
prob_i = prob_i.item()
batch_i = batch_i.item()
prob_k_i = prob_k_i
index_k_i = index_k_i
if outputs_i != batch_i and prob_i > max_prob and batch_i not in [0, 101, 102]:
max_prob = prob_i
max_index = n
max_output = outputs_i
second_prob, third_prob = prob_k_i[1].item(), prob_k_i[2].item()
second_output, third_output = index_k_i[1].item(), index_k_i[2].item()
if max_prob > threshold:
origin = batch_line[max_index].clone()
batch_line[max_index] = max_output
f.write('repeat:' +
str(repeat) + ', ' +
str(max_index) + ', ' +
self.tokenizer.decode(origin) + 'to' +
self.tokenizer.decode(max_output) + ' ' + str(max_prob) + ' ' +
self.tokenizer.decode(second_output) + ' ' + str(second_prob) + ' ' +
self.tokenizer.decode(third_output) + ' ' + str(third_prob) +
'\n')
f.write('\n')
for line in batch['input_ids']:
final_out.append([i.item() for i in line])
return None, final_out, prob_outputs
def __forward_prop_all(self, dataloader, back_prop=True):
loss_sum = 0
steps = 0
collected_outputs = []
prob_outputs = []
softmax = torch.nn.Softmax(-1)
for batch in tqdm(dataloader):
batch = {k: v.to(self.device) for k, v in batch.items()}
loss, logits = self.model(input_ids=batch['input_ids'],
attention_mask=batch['attention_mask'],
token_type_ids=batch['token_type_ids'],
trg_ids=batch['trg_ids'],
pinyin_ids=batch['pinyin_ids'])
# det_loss, cor_loss, prob, logits, sequence_output = self.model(input_ids=batch['input_ids'],
# attention_mask=batch['attention_mask'],
# cor_labels=batch['trg_ids'],
# )
# loss = det_loss + cor_loss
loss_sum += loss.item()
if back_prop:
loss.backward()
# 对抗训练
loss_adv, _ = self.model(input_ids=batch['input_ids'],
attention_mask=batch['attention_mask'],
token_type_ids=batch['token_type_ids'],
trg_ids=batch['trg_ids'],
pinyin_ids=batch['pinyin_ids'])
loss_adv.backward() # 反向传播，并在正常的grad基础上，累加对抗训练的梯度
self.optimizer.step()
self.scheduler.step()
self.optimizer.zero_grad()
else:
outputs = torch.argmax(logits, dim=-1).cpu()
outputs_prob = torch.max(softmax(logits), dim=-1)[0].cpu()
for outputs_i, prob_i in zip(outputs, outputs_prob):
collected_outputs.append(outputs_i)
prob_outputs.append(prob_i)
steps += 1
epoch_loss = loss_sum / steps
return epoch_loss, collected_outputs, prob_outputs
def __save_ckpt(self, epoch):
save_path = self.config.save_path
if not os.path.exists(save_path):
os.mkdir(save_path)
path = os.path.join(save_path, self.config.tag + str(epoch) + ".pt")
torch.save(self.model.state_dict(), path)
def train(self):
no_improve = 0
most_count = 0
for epoch in range(1, self.config.num_epochs + 1):
self.model.train()
train_loss, _, _ = self.__forward_prop_all(self.train_dataloader, back_prop=True)
self.model.eval()
with torch.no_grad():
valid_loss, valid_output, valid_prob = self.__forward_prop(self.valid_dataloader, back_prop=False)
# print(f"train_loss: {train_loss}, valid_loss: {valid_loss}")
if not os.path.exists(self.config.save_path + '/tmp/'):
os.makedirs(self.config.save_path + '/tmp/')
save_decode_result_para(valid_output, valid_prob, self.valid_dataloader.dataset.data,
self.config.save_path + '/tmp/' + "valid_" + str(epoch) + ".txt")
count = save_decode_result_lbl(valid_output, self.valid_dataloader.dataset.data,
self.config.save_path + '/tmp/' + "valid_" + str(epoch) + ".lbl")
# try:
char_metrics, sent_metrics = csc_metrics(
pred=self.config.save_path + '/tmp/' + "valid_" + str(epoch) + ".lbl",
gold=self.config.lbl_path,
src='data/val_1.src',
)
get_best_score(self.best_score, self.best_epoch, epoch,
char_metrics["Correction"]["F1"], sent_metrics["Correction"]["F1"])
if max(self.best_epoch.values()) == epoch:
self.__save_ckpt(epoch)
print(f"curr epoch: {epoch} | curr best epoch {self.best_epoch}")
print(f"best socre:{self.best_score}")
print(f"no improve: {epoch - max(self.best_epoch.values())}")
if (epoch - max(self.best_epoch.values())) >= self.config.patience:
break
def main(config):
trainer = Trainer(config)
trainer.train()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--pretrained_model", required=True, type=str)
parser.add_argument("--train_path", required=True, type=str)
parser.add_argument("--dev_path", required=True, type=str)
parser.add_argument("--test_path", required=True, type=str)
parser.add_argument("--lbl_path", required=True, type=str)
parser.add_argument("--test_lbl_path", required=True, type=str)
parser.add_argument("--save_path", required=True, type=str)
parser.add_argument("--max_seq_len", default=128, type=int)
parser.add_argument("--batch_size", default=32, type=int)
parser.add_argument("--label_ignore_id", default=0, type=int)
parser.add_argument("--num_epochs", default=1, type=int)
parser.add_argument("--lr", default=2e-5, type=float)
parser.add_argument("--patience", default=10, type=int)
parser.add_argument("--freeze_bert", default=False, type=bool)
parser.add_argument("--tie_cls_weight", default=False, type=bool)
parser.add_argument("--tag", required=True, type=str)
parser.add_argument("--seed", default=2021, type=int)
args = parser.parse_args()
main(args)