- 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
- 🍖 原作者：K同学啊 | 接辅导、项目定制
- 🚀 文章来源：K同学的学习圈子
目录
一、准备工作
1.任务说明
文本分类流程图：
2.加载数据
编辑 二、数据的预处理
1.构建词典
2.生成数据批次和迭代器
三、模型构建
四、训练模型
五、小结
一、准备工作
1.任务说明
本次将使用PyTorch实现中文文本分类。主要代码与N1周基本一致，不同的是本次任务中使用了本地的中文数据，数据示例如下:
本周任务:
1.学习如何进行中文本文预处理
2.根据文本内容(第1列)预测文本标签(第2列)
进阶任务:
1.尝试根据第一周的内容独立实现，尽可能的不看本文的代码
2.构建更复杂的网络模型，将准确率提升至91%
文本分类流程图：
2.加载数据
import torch
import torch.nn as nn
import torchvision
from torchvision import transforms,datasets
import os,PIL,pathlib,warningswarnings.filterwarnings("ignore") #忽略警告信息device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)import pandas as pd#加载自定义中文数据
train_data = pd.read_csv('./train.csv',sep='\t',header = None)
#构造数据集迭代器
def coustom_data_iter(texts,labels):for x,y in zip(texts,labels):yield x,ytrain_iter =coustom_data_iter(train_data[0].values[:],train_data[1].values[:])
输出：
二、数据的预处理
1.构建词典
#构建词典
from torchtext.data.utils import get_tokenizer
from torchtext.vocab import build_vocab_from_iterator
import jieba#中文分词方法
tokenizer = jieba.lcut
def yield_tokens(data_iter):for text,_ in data_iter:yield tokenizer(text)vocab = build_vocab_from_iterator(yield_tokens(train_iter),specials=["<unk>"])
vocab.set_default_index(vocab["<unk>"]) #设置默认索引，如果找不到单词，则会选择默认索引
vocab(['我','想','看','和平','精英','上','战神','必备','技巧','的','游戏','视频'])
label_name = list(set(train_data[1].values[:]))
print(label_name)
text_pipeline = lambda x : vocab(tokenizer(x))
label_pipeline = lambda x : label_name.index(x)print(text_pipeline('我想看和平精英上战神必备技巧的游戏视频'))
print(label_pipeline('Video-Play'))
输出：
2.生成数据批次和迭代器
#生成数据批次和迭代器
from torch.utils.data import DataLoaderdef collate_batch(batch):label_list, text_list, offsets = [],[],[0] for(_text, _label) in batch:#标签列表label_list.append(label_pipeline(_label))#文本列表processed_text = torch.tensor(text_pipeline(_text), dtype=torch.int64)text_list.append(processed_text)#偏移量offsets.append(processed_text.size(0))label_list = torch.tensor(label_list,dtype=torch.int64)text_list = torch.cat(text_list)offsets = torch.tensor(offsets[:-1]).cumsum(dim=0) #返回维度dim中输入元素的累计和return text_list.to(device), label_list.to(device), offsets.to(device)#数据加载器
dataloader = DataLoader(train_iter,batch_size = 8,shuffle = False,collate_fn = collate_batch
)三、模型构建
#搭建模型
from torch import nnclass TextClassificationModel(nn.Module):def __init__(self, vocab_size, embed_dim, num_class):super(TextClassificationModel,self).__init__()self.embedding = nn.EmbeddingBag(vocab_size, #词典大小embed_dim, # 嵌入的维度sparse=False) #self.fc = nn.Linear(embed_dim, num_class)self.init_weights()def init_weights(self):initrange = 0.5self.embedding.weight.data.uniform_(-initrange, initrange)self.fc.weight.data.uniform_(-initrange, initrange)self.fc.bias.data.zero_()def forward(self, text, offsets):embedded = self.embedding(text, offsets)return self.fc(embedded)
#初始化模型
#定义实例
num_class = len(label_name)
vocab_size = len(vocab)
em_size = 64
model = TextClassificationModel(vocab_size, em_size, num_class).to(device)
#定义训练与评估函数
import timedef train(dataloader):model.train() #切换为训练模式total_acc, train_loss, total_count = 0,0,0log_interval = 50start_time = time.time()for idx, (text,label, offsets) in enumerate(dataloader):predicted_label = model(text, offsets)optimizer.zero_grad() #grad属性归零loss = criterion(predicted_label, label) #计算网络输出和真实值之间的差距，label为真loss.backward() #反向传播torch.nn.utils.clip_grad_norm_(model.parameters(),0.1) #梯度裁剪optimizer.step() #每一步自动更新#记录acc与losstotal_acc += (predicted_label.argmax(1) == label).sum().item()train_loss += loss.item()total_count += label.size(0)if idx % log_interval == 0 and idx > 0:elapsed = time.time() - start_timeprint('|epoch{:d}|{:4d}/{:4d} batches|train_acc{:4.3f} train_loss{:4.5f}'.format(epoch,idx,len(dataloader),total_acc/total_count,train_loss/total_count))total_acc,train_loss,total_count = 0,0,0staet_time = time.time()def evaluate(dataloader):model.eval() #切换为测试模式total_acc,train_loss,total_count = 0,0,0with torch.no_grad():for idx,(text,label,offsets) in enumerate(dataloader):predicted_label = model(text, offsets)loss = criterion(predicted_label,label) #计算loss值#记录测试数据total_acc += (predicted_label.argmax(1) == label).sum().item()train_loss += loss.item()total_count += label.size(0)return total_acc/total_count, train_loss/total_count四、训练模型
#拆分数据集并运行模型
from torch.utils.data.dataset import random_split
from torchtext.data.functional import to_map_style_dataset# 超参数设定
EPOCHS = 10 #epoch
LR = 5 #learningRate
BATCH_SIZE = 64 #batch size for training#设置损失函数、选择优化器、设置学习率调整函数
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr = LR)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 1.0, gamma = 0.1)
total_accu = None# 构建数据集
train_iter = custom_data_iter(train_data[0].values[:],train_data[1].values[:])
train_dataset = to_map_style_dataset(train_iter)
split_train_, split_valid_ = random_split(train_dataset,[int(len(train_dataset)*0.8),int(len(train_dataset)*0.2)])train_dataloader = DataLoader(split_train_, batch_size = BATCH_SIZE, shuffle = True, collate_fn = collate_batch)
valid_dataloader = DataLoader(split_valid_, batch_size = BATCH_SIZE, shuffle = True, collate_fn = collate_batch)for epoch in range(1, EPOCHS + 1):epoch_start_time = time.time()train(train_dataloader)val_acc, val_loss = evaluate(valid_dataloader)#获取当前的学习率lr = optimizer.state_dict()['param_groups'][0]['lr']if total_accu is not None and total_accu > val_acc:scheduler.step()else:total_accu = val_accprint('-' * 69)print('| epoch {:d} | time:{:4.2f}s | valid_acc {:4.3f} valid_loss {:4.3f}'.format(epoch,time.time() - epoch_start_time,val_acc,val_loss))print('-' * 69)
test_acc,test_loss = evaluate(valid_dataloader)
print('模型准确率为：{:5.4f}'.format(test_acc))
#测试指定的数据
def predict(text, text_pipeline):with torch.no_grad():text = torch.tensor(text_pipeline(text))output = model(text, torch.tensor([0]))return output.argmax(1).item()ex_text_str = "还有双鸭山到淮阴的汽车票吗13号的"
model = model.to("cpu")print("该文本的类别是: %s" %label_name[predict(ex_text_str,text_pipeline)])输出：
|epoch1| 50/ 152 batches|train_acc0.431 train_loss0.03045
|epoch1| 100/ 152 batches|train_acc0.700 train_loss0.01936
|epoch1| 150/ 152 batches|train_acc0.768 train_loss0.01370
---------------------------------------------------------------------
| epoch 1 | time:1.58s | valid_acc 0.789 valid_loss 0.012
---------------------------------------------------------------------
|epoch2| 50/ 152 batches|train_acc0.818 train_loss0.01030
|epoch2| 100/ 152 batches|train_acc0.831 train_loss0.00932
|epoch2| 150/ 152 batches|train_acc0.850 train_loss0.00811
---------------------------------------------------------------------
| epoch 2 | time:1.47s | valid_acc 0.837 valid_loss 0.008
---------------------------------------------------------------------
|epoch3| 50/ 152 batches|train_acc0.870 train_loss0.00688
|epoch3| 100/ 152 batches|train_acc0.887 train_loss0.00658
|epoch3| 150/ 152 batches|train_acc0.893 train_loss0.00575
---------------------------------------------------------------------
| epoch 3 | time:1.46s | valid_acc 0.866 valid_loss 0.007
---------------------------------------------------------------------
|epoch4| 50/ 152 batches|train_acc0.906 train_loss0.00507
|epoch4| 100/ 152 batches|train_acc0.918 train_loss0.00468
|epoch4| 150/ 152 batches|train_acc0.915 train_loss0.00478
---------------------------------------------------------------------
| epoch 4 | time:1.47s | valid_acc 0.886 valid_loss 0.006
---------------------------------------------------------------------
|epoch5| 50/ 152 batches|train_acc0.938 train_loss0.00378
|epoch5| 100/ 152 batches|train_acc0.935 train_loss0.00379
|epoch5| 150/ 152 batches|train_acc0.932 train_loss0.00376
---------------------------------------------------------------------
| epoch 5 | time:1.51s | valid_acc 0.890 valid_loss 0.006
---------------------------------------------------------------------
|epoch6| 50/ 152 batches|train_acc0.951 train_loss0.00310
|epoch6| 100/ 152 batches|train_acc0.952 train_loss0.00287
|epoch6| 150/ 152 batches|train_acc0.950 train_loss0.00289
---------------------------------------------------------------------
| epoch 6 | time:1.50s | valid_acc 0.894 valid_loss 0.006
---------------------------------------------------------------------
|epoch7| 50/ 152 batches|train_acc0.963 train_loss0.00233
|epoch7| 100/ 152 batches|train_acc0.963 train_loss0.00244
|epoch7| 150/ 152 batches|train_acc0.965 train_loss0.00222
---------------------------------------------------------------------
| epoch 7 | time:1.49s | valid_acc 0.898 valid_loss 0.005
---------------------------------------------------------------------
|epoch8| 50/ 152 batches|train_acc0.975 train_loss0.00183
|epoch8| 100/ 152 batches|train_acc0.976 train_loss0.00176
|epoch8| 150/ 152 batches|train_acc0.971 train_loss0.00188
---------------------------------------------------------------------
| epoch 8 | time:1.67s | valid_acc 0.900 valid_loss 0.005
---------------------------------------------------------------------
|epoch9| 50/ 152 batches|train_acc0.982 train_loss0.00145
|epoch9| 100/ 152 batches|train_acc0.982 train_loss0.00139
|epoch9| 150/ 152 batches|train_acc0.980 train_loss0.00141
---------------------------------------------------------------------
| epoch 9 | time:2.05s | valid_acc 0.901 valid_loss 0.006
---------------------------------------------------------------------
|epoch10| 50/ 152 batches|train_acc0.990 train_loss0.00108
|epoch10| 100/ 152 batches|train_acc0.984 train_loss0.00119
|epoch10| 150/ 152 batches|train_acc0.986 train_loss0.00105
---------------------------------------------------------------------
| epoch 10 | time:1.98s | valid_acc 0.900 valid_loss 0.005
---------------------------------------------------------------------
模型准确率为：0.8996
该文本的类别是: Travel-Query
五、小结
- 数据加载：
- 定义一个生成器函数，将文本和标签成对迭代。这是为了后续的数据处理和加载做准备。
- 分词与词汇表：
- 使用
jieba进行中文分词，jieba.lcut可以将中文文本切割成单个词语列表。 - 使用
torchtext的build_vocab_from_iterator从分词后的文本中构建词汇表，并设置默认索引为<unk>，表示未知词汇。这对处理未见过的词汇非常重要。
- 使用
- 数据管道：创建文本和标签处理管道。
-
创建两个处理管道：
text_pipeline：将文本转换为词汇表中的索引。label_pipeline：将标签转换为索引。
-
- 模型构建：定义带嵌入层和全连接层的文本分类模型。
- 定义一个文本分类模型
TextClassificationModel，包括一个嵌入层nn.EmbeddingBag和一个全连接层nn.Linear。nn.EmbeddingBag在处理变长序列时性能较好，因为它不需要明确的填充操作。
- 定义一个文本分类模型
