1-3,文本数据建模流程案例

!pip install torchkeras 
!pip install prettytable 
!pip install torchtext

一，准备数据

imdb数据集的目标是根据电影评论的文本内容预测评论的情感标签。
训练集有20000条电影评论文本，测试集有5000条电影评论文本，其中正面评论和负面评论都各占一半。

文本数据预处理较为繁琐，包括中文切词（本示例不涉及），构建词典，编码转换，序列填充，构建数据管道等等。
在torch中预处理文本数据一般使用torchtext或者自定义Dataset，torchtext功能非常强大，可以构建文本分类，序列标注，问答模型，机器翻译等NLP任务的数据集。
下面仅演示使用它来构建文本分类数据集的方法。
较完整的教程可以参考以下知乎文章：《pytorch学习笔记—Torchtext》
https://zhuanlan.zhihu.com/p/65833208

torchtext常见API一览

• torchtext.data.Example : 用来表示一个样本，数据和标签
• torchtext.vocab.Vocab: 词汇表，可以导入一些预训练词向量
• torchtext.data.Datasets: 数据集类，__getitem__返回 Example实例, torchtext.data.TabularDataset是其子类。
• torchtext.data.Field : 用来定义字段的处理方法（文本字段，标签字段）创建 Example时的 预处理，batch 时的一些处理操作。
• torchtext.data.Iterator: 迭代器，用来生成 batch
• torchtext.datasets: 包含了常见的数据集.

import numpy as np 
import pandas as pd 
from collections import OrderedDict
import re,string
MAX_WORDS = 10000  # 仅考虑最高频的10000个词
MAX_LEN = 200  # 每个样本保留200个词的长度
BATCH_SIZE = 20 
train_data_path = '/home/kesci/input/data6936/data/imdb/train.tsv'
test_data_path = '/home/kesci/input/data6936/data/imdb/test.tsv'
train_token_path = '/home/kesci/input/data6936/data/imdb/train_token.tsv'
test_token_path =  '/home/kesci/input/data6936/data/imdb/test_token.tsv'
train_samples_path = '/home/kesci/input/data6936/data/imdb/train_samples/'
test_samples_path =  '/home/kesci/input/data6936/data/imdb/test_samples/'

首先我们构建词典，并保留最高频的MAX_WORDS个词。

##构建词典

word_count_dict = {}

#清洗文本
def clean_text(text):
    lowercase = text.lower().replace("\n"," ")
    stripped_html = re.sub('<br />', ' ',lowercase)
    cleaned_punctuation = re.sub('[%s]'%re.escape(string.punctuation),'',stripped_html)
    return cleaned_punctuation

with open(train_data_path,"r",encoding = 'utf-8') as f:
    for line in f:
        label,text = line.split("\t")
        cleaned_text = clean_text(text)
        for word in cleaned_text.split(" "):
            word_count_dict[word] = word_count_dict.get(word,0)+1 

df_word_dict = pd.DataFrame(pd.Series(word_count_dict,name = "count"))
df_word_dict = df_word_dict.sort_values(by = "count",ascending =False)

df_word_dict = df_word_dict[0:MAX_WORDS-2] #  
df_word_dict["word_id"] = range(2,MAX_WORDS) #编号0和1分别留给未知词<unkown>和填充<padding>

word_id_dict = df_word_dict["word_id"].to_dict()

df_word_dict.head(10)

然后我们利用构建好的词典，将文本转换成token序号。

#转换token

# 填充文本
def pad(data_list,pad_length):
    padded_list = data_list.copy()
    if len(data_list)> pad_length:
         padded_list = data_list[-pad_length:]
    if len(data_list)< pad_length:
         padded_list = [1]*(pad_length-len(data_list))+data_list
    return padded_list

def text_to_token(text_file,token_file):
    with open(text_file,"r",encoding = 'utf-8') as fin,\
      open(token_file,"w",encoding = 'utf-8') as fout:
        for line in fin:
            label,text = line.split("\t")
            cleaned_text = clean_text(text)
            word_token_list = [word_id_dict.get(word, 0) for word in cleaned_text.split(" ")]
            pad_list = pad(word_token_list,MAX_LEN)
            out_line = label+"\t"+" ".join([str(x) for x in pad_list])
            fout.write(out_line+"\n")
        
text_to_token(train_data_path,train_token_path)
text_to_token(test_data_path,test_token_path)

接着将token文本按照样本分割，每个文件存放一个样本的数据。

# 分割样本
import os

if not os.path.exists(train_samples_path):
    os.mkdir(train_samples_path)
    
if not os.path.exists(test_samples_path):
    os.mkdir(test_samples_path)
    
def split_samples(token_path,samples_dir):
    with open(token_path,"r",encoding = 'utf-8') as fin:
        i = 0
        for line in fin:
            with open(samples_dir+"%d.txt"%i,"w",encoding = "utf-8") as fout:
                fout.write(line)
            i = i+1

split_samples(train_token_path,train_samples_path)
split_samples(test_token_path,test_samples_path)

print(os.listdir(train_samples_path)[0:100])

一切准备就绪，我们可以创建数据集Dataset, 从文件名称列表中读取文件内容了。

import os
from torch.utils.data import Dataset,DataLoader 
class imdbDataset(Dataset):
    def __init__(self,samples_dir):
        self.samples_dir = samples_dir
        self.samples_paths = os.listdir(samples_dir)
    
    def __len__(self):
        return len(self.samples_paths)
    
    def __getitem__(self,index):
        path = self.samples_dir + self.samples_paths[index]
        with open(path,"r",encoding = "utf-8") as f:
            line = f.readline()
            label,tokens = line.split("\t")
            label = torch.tensor([float(label)],dtype = torch.float)
            feature = torch.tensor([int(x) for x in tokens.split(" ")],dtype = torch.long)
            return  (feature,label)
ds_train = imdbDataset(train_samples_path)
ds_test = imdbDataset(test_samples_path)

print(len(ds_train))
print(len(ds_test))

dl_train = DataLoader(ds_train,batch_size = BATCH_SIZE,shuffle = True,num_workers=4)
dl_test = DataLoader(ds_test,batch_size = BATCH_SIZE,num_workers=4)

for features,labels in dl_train:
    print(features)
    print(labels)
    break

二，定义模型

使用Pytorch通常有三种方式构建模型：使用nn.Sequential按层顺序构建模型，继承nn.Module基类构建自定义模型，继承nn.Module基类构建模型并辅助应用模型容器(nn.Sequential,nn.ModuleList,nn.ModuleDict)进行封装。
此处选择使用第三种方式进行构建。
由于接下来使用类形式的训练循环，我们将模型封装成torchkeras.Model类来获得类似Keras中高阶模型接口的功能。
Model类实际上继承自nn.Module类。

import torch
from torch import nn 
import torchkeras
torch.random.seed()
import torch
from torch import nn 

class Net(torchkeras.Model):
    
    def __init__(self):
        super(Net, self).__init__()
        
        #设置padding_idx参数后将在训练过程中将填充的token始终赋值为0向量
        self.embedding = nn.Embedding(num_embeddings = MAX_WORDS,embedding_dim = 3,padding_idx = 1)
        self.conv = nn.Sequential()
        self.conv.add_module("conv_1",nn.Conv1d(in_channels = 3,out_channels = 16,kernel_size = 5))
        self.conv.add_module("pool_1",nn.MaxPool1d(kernel_size = 2))
        self.conv.add_module("relu_1",nn.ReLU())
        self.conv.add_module("conv_2",nn.Conv1d(in_channels = 16,out_channels = 128,kernel_size = 2))
        self.conv.add_module("pool_2",nn.MaxPool1d(kernel_size = 2))
        self.conv.add_module("relu_2",nn.ReLU())
        
        self.dense = nn.Sequential()
        self.dense.add_module("flatten",nn.Flatten())
        self.dense.add_module("linear",nn.Linear(6144,1))
        self.dense.add_module("sigmoid",nn.Sigmoid())
        
    def forward(self,x):
        x = self.embedding(x).transpose(1,2)
        x = self.conv(x)
        y = self.dense(x)
        return y
model = Net()
print(model)
model.summary(input_shape = (200,),input_dtype = torch.LongTensor)

三，训练模型

训练Pytorch通常需要用户编写自定义训练循环，训练循环的代码风格因人而异。

有3类典型的训练循环代码风格：脚本形式训练循环，函数形式训练循环，类形式训练循环。

此处介绍一种类形式的训练循环。

我们仿照Keras定义了一个高阶的模型接口Model,实现 fit, validate，predict, summary 方法，相当于用户自定义高阶API。

# 准确率
def accuracy(y_pred,y_true):
    y_pred = torch.where(y_pred>0.5,torch.ones_like(y_pred,dtype = torch.float32),
                      torch.zeros_like(y_pred,dtype = torch.float32))
    acc = torch.mean(1-torch.abs(y_true-y_pred))
    return acc

model.compile(loss_func = nn.BCELoss(),optimizer= torch.optim.Adagrad(model.parameters(),lr = 0.02),
             metrics_dict={"accuracy":accuracy})

# 有时候模型训练过程中不收敛，需要多试几次
dfhistory = model.fit(20,dl_train,dl_val=dl_test,log_step_freq= 200)

四，评估模型

%matplotlib inline
%config InlineBackend.figure_format = 'svg'

import matplotlib.pyplot as plt

def plot_metric(dfhistory, metric):
    train_metrics = dfhistory[metric]
    val_metrics = dfhistory['val_'+metric]
    epochs = range(1, len(train_metrics) + 1)
    plt.plot(epochs, train_metrics, 'bo--')
    plt.plot(epochs, val_metrics, 'ro-')
    plt.title('Training and validation '+ metric)
    plt.xlabel("Epochs")
    plt.ylabel(metric)
    plt.legend(["train_"+metric, 'val_'+metric])
    plt.show()

plot_metric(dfhistory,"loss")

plot_metric(dfhistory,"accuracy")

# 评估
model.evaluate(dl_test)

五，使用模型

model.predict(dl_test)

六，保存模型

推荐使用保存参数方式保存Pytorch模型。

print(model.state_dict().keys())

# 保存模型参数
torch.save(model.state_dict(), "./data/model_parameter.pkl")
model_clone = Net()
model_clone.load_state_dict(torch.load("./data/model_parameter.pkl"))
model_clone.compile(loss_func = nn.BCELoss(),optimizer= torch.optim.Adagrad(model.parameters(),lr = 0.02),
             metrics_dict={"accuracy":accuracy})
# 评估模型
model_clone.evaluate(dl_test)

项目github地址：https://github.com/lyhue1991/eat_pytorch_in_20_days
项目和鲸专栏地址：https://www.kesci.com/home/column/5f2ac5d8af3980002cb1bc08