-
Notifications
You must be signed in to change notification settings - Fork 0
/
bert.py
134 lines (115 loc) · 4.22 KB
/
bert.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
import pandas as pd
import numpy as np
from sklearn.preprocessing import MinMaxScaler
import time
import copy
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.autograd as autograd
import torch.nn.functional
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
import warnings
import torch
import time
import argparse
import get_data
import os
#main.py
from transformers import BertTokenizer
from transformers import BertForSequenceClassification
from transformers import BertConfig
#from transformers import BertPreTrainedModel
embedding_dim=400
hidden_dim=256
vocab_size=51158
target=1
Batchsize=128
stringlen=25
Epoch=20
lr=0.01
from transformers import BertModel
USE_CUDA = torch.cuda.is_available()
texta,textb,labels,evala,evalb,evallabels=get_data.train_data(stringlen)
resulta,resultb=get_data.result_data(stringlen)
if USE_CUDA:
texta = texta.cuda()
textb= textb.cuda()
labels= labels.cuda()
evala= evala.cuda()
evalb= evalb.cuda()
evallabels= evallabels.cuda()
resulta=resulta.cuda()
resultb=resulta.cuda()
tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
tokenizer.add_tokens([str(i) for i in range(0,51157)])
len_token=len(tokenizer)
print(len_token)
def get_train_args():
parser=argparse.ArgumentParser()
parser.add_argument('--batch_size',type=int,default=2,help = '每批数据的数量')
parser.add_argument('--nepoch',type=int,default=30,help = '训练的轮次')
parser.add_argument('--lr',type=float,default=0.001,help = '学习率')
parser.add_argument('--gpu',type=bool,default=True,help = '是否使用gpu')
parser.add_argument('--num_workers',type=int,default=2,help='dataloader使用的线程数量')
parser.add_argument('--num_labels',type=int,default=2,help='分类类数')
parser.add_argument('--data_path',type=str,default='./data',help='数据路径')
opt=parser.parse_args()
print(opt)
return opt
def get_model(opt,len_token):
model = BertForSequenceClassification.from_pretrained('bert-base-cased',num_labels=opt.num_labels)
#model = bert_.BertForSequenceClassification.from_pretrained('bert-base-cased', num_labels=opt.num_labels)
#model = bert_LSTM.Net()
model.resize_token_embeddings(len_token)
return model
def train(net, texta,textb,labels,evala,evalb,evallabels,resulta,resultb,num_epochs, learning_rate, batch_size):
net.train()
#loss_fct = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=learning_rate)
#optimizer = torch.optim.AdamW(net.parameters(), lr=learning_rate)
dataset = torch.utils.data.TensorDataset(texta,textb,labels)
train_iter = torch.utils.data.DataLoader(dataset, batch_size, shuffle=True)
pre=0.5
acc_list=[]
index_list=[]
pre_list=[]
for epoch in range(num_epochs):
correct = 0
total=0
iter = 0
net.train()
for XA, XB , y in train_iter:
iter += 1
XA = XA.long()
XB = XB.long()
y=y.long()
y=y.view(-1)
#print(y)
if XA.size(0)!= batch_size:
break
optimizer.zero_grad()
outputs= net(XA,XB, labels=y)
loss, logits = outputs[0], outputs[1]
_, predicted = torch.max(logits.data, 1)
loss.backward()
optimizer.step()
total += XA.size(0)
#print("predict",predicted)
#print("label",y)
correct += predicted.data.eq(y.data).cpu().sum()
s = ("Acc:%.3f" % ((1.0 * correct.numpy()) / total))
if iter % 200 ==0:
print(s)
print(s)
#print(net.state_dict()["word_embeddings.weight"])
s = ((1.0 * correct) / total)
print("epoch: ",epoch, " ",correct,"/" , total, "TrainAcc:", s)
acc_list.append(s)
index_list.append(epoch)
return acc_list,index_list,pre_list
opt = get_train_args()
model=get_model(opt,len_token)
model=model.cuda()
a,b,c=train(model,texta,textb,labels,evala,evalb,evallabels,resulta,resultb,Epoch,lr,2)