Shulou(Shulou.com)06/02 Report--

This article mainly explains "how to use the pytorch framework", the content of the article is simple and clear, easy to learn and understand, now please follow the editor's ideas slowly in depth, together to study and learn "how to use the pytorch framework" bar!

Chinese News emotion Classification Bert-Pytorch-transformers

Using the pytorch framework and transformers package, as well as the Chinese pre-training model of Bert

File directory

Data

Train_DataSet.csv

Train_DataSet_Label.csv

Main.py

NewsData.py

# main.py

From transformers import BertTokenizer

From transformers import BertForSequenceClassification

From transformers import BertConfig

From transformers import BertPreTrainedModel

Import torch

Import torch.nn as nn

From transformers import BertModel

Import time

Import argparse

From NewsData import NewsData

Import os

Def get_train_args ():

Parser=argparse.ArgumentParser ()

Parser.add_argument ('--batch_size',type=int,default=10,help = 'quantity per batch')

Parser.add_argument ('--nepoch',type=int,default=3,help = 'training rounds')

Parser.add_argument ('--lr',type=float,default=0.001,help = 'learning rate')

Parser.add_argument ('--gpu',type=bool,default=True,help = 'whether to use gpu')

Parser.add_argument ('--number of threads used by num_workers',type=int,default=2,help='dataloader')

Parser.add_argument ('--number of num_labels',type=int,default=3,help=' categories')

Parser.add_argument ('--data_path',type=str,default='./data',help=' data path')

Opt=parser.parse_args ()

Print (opt)

Return opt

Def get_model (opt):

# Class method. From _ pretrained () gets the pre-training model, and num_labels is the number of classes classified.

Model = BertForSequenceClassification.from_pretrained ('bert-base-chinese',num_labels=opt.num_labels)

Return model

Def get_data (opt):

# NewsData inherits from pytorch's Dataset class

Trainset = NewsData (opt.data_path,is_train = 1)

Trainloader=torch.utils.data.DataLoader (trainset,batch_size=opt.batch_size,shuffle=True,num_workers=opt.num_workers)

Testset = NewsData (opt.data_path,is_train = 0)

Testloader=torch.utils.data.DataLoader (testset,batch_size=opt.batch_size,shuffle=False,num_workers=opt.num_workers)

Return trainloader,testloader

Def train (epoch,model,trainloader,testloader,optimizer,opt):

Print ('\ ntrain-Epoch:% d'% (epoch+1))

Model.train ()

Start_time = time.time ()

Print_step = int (len (trainloader) / 10)

For batch_idx, (sue,label,posi) in enumerate (trainloader):

If opt.gpu:

Sue = sue.cuda ()

Posi = posi.cuda ()

Label = label.unsqueeze (1) .cuda ()

Optimizer.zero_grad ()

# input parameters are word list, location list, label

Outputs = model (sue, position_ids=posi,labels = label)

Loss, logits = outputs [0], outputs [1]

Loss.backward ()

Optimizer.step ()

If batch_idx% print_step = = 0:

Print ("Epoch:%d [d |% d] loss:%f"% (epoch+1,batch_idx,len (trainloader), loss.mean ()

Print ("time:%.3f"% (time.time ()-start_time))

Def test (epoch,model,trainloader,testloader,opt):

Print ('\ ntest-Epoch:% d'% (epoch+1))

Model.eval ()

Total=0

Correct=0

With torch.no_grad ():

For batch_idx, (sue,label,posi) in enumerate (testloader):

If opt.gpu:

Sue = sue.cuda ()

Posi = posi.cuda ()

Labels = label.unsqueeze (1) .cuda ()

Label = label.cuda ()

Else:

Labels = label.unsqueeze (1)

Outputs = model (sue, labels=labels)

Loss, logits = outputs [: 2]

_, predicted=torch.max (logits.data,1)

Total+=sue.size (0)

Correct+=predicted.data.eq (label.data). Cpu (). Sum ()

S = ("Acc:%.3f"% ((1.0*correct.numpy ()) / total))

Print (s)

If _ _ name__=='__main__':

Opt = get_train_args ()

Model = get_model (opt)

Trainloader,testloader = get_data (opt)

If opt.gpu:

Model.cuda ()

Optimizer=torch.optim.SGD (model.parameters (), lr=opt.lr,momentum=0.9)

If not os.path.exists ('. / model.pth'):

For epoch in range (opt.nepoch):

Train (epoch,model,trainloader,testloader,optimizer,opt)

Test (epoch,model,trainloader,testloader,opt)

Torch.save (model.state_dict (),'. / model.pth')

Else: which hospital has good http://www.120kdfk.com/ for gynecology treatment in Zhengzhou?

Model.load_state_dict (torch.load ('model.pth'))

Print ('model exists, direct test')

Test (0meme model recorder loader dint testloader recorder opt)

# NewsData.py

From transformers import BertTokenizer

From transformers import BertForSequenceClassification

From transformers import BertConfig

From transformers import BertPreTrainedModel

Import torch

Import torch.nn as nn

From transformers import BertModel

Import time

Import argparse

Class NewsData (torch.utils.data.Dataset):

Def _ _ init__ (self,root,is_train = 1):

Self.tokenizer = BertTokenizer.from_pretrained ('bert-base-chinese')

Self.data_num = 7346

Self.x_list = []

Self.y_list = []

Self.posi = []

With open (root +'/ Train_DataSet.csv',encoding='UTF-8') as f:

For i in range (self.data_num+1):

Line = f.readline () [:-1] + 'this is a neutral data'

Data_one_str = line.split (',') [len (line.split (',')-2]

Data_two_str = line.split (',') [len (line.split (',')-1]

If len (data_one_str) < 6:

Z = len (data_one_str)

Data_one_str = data_one_str +','+ data_two_ str0: min (200Jing len (data_two_str))]

Else:

Data_one_str = data_one_str

If iTunes 0:

Continue

Word_l = self.tokenizer.encode (data_one_str, add_special_tokens=False)

If len (word_l)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.