糗事百科实例:

爬取糗事百科段子,假设页面的URL是 http://www.qiushibaike.com/8hr/page/1

要求:

  1. 使用requests获取页面信息,用XPath / re 做数据提取

  2. 获取每个帖子里的用户头像链接用户姓名段子内容点赞次数评论次数

  3. 保存到 json 文件内

参考代码

#coding=utf-8
import requests
from retrying import retry
from lxml import etree

class Qiubai_spider():
    def __init__(self):
        self.url = "http://www.qiushibaike.com/8hr/page/{}/"
        self.headers = {
            "User-Agent":"Mozilla/5.0 (compatible; MSIE 9.0; Windows NT 6.1 Trident/5.0;"
        }

    @retry(stop_max_attempt_number=5) #调用retry,当assert出错时候,重复请求5次
    def parse_url(self,url):
        response = requests.get(url,timeout=10,headers=self.headers) #请求url
        assert response.status_code==200  #当响应码不是200时候,做断言报错处理
        print(url)
        return etree.HTML(response.text) #返回etree之后的html

    def parse_content(self,html):
        item_temp = html.xpath("//div[@class='article block untagged mb15']")
        print(len(item_temp))
        for item in item_temp:
            #获取用户头像地址
            avatar = item.xpath("./div[1]/a[1]/img/@src")[0] if len(item.xpath("./div[1]/a[1]/img/@src"))>0 else None
            #为头像地址添加前缀
            if avatar is not None and not avatar.startswith("http:"):
                avatar = "http:"+avatar
            print(avatar)
            name = item.xpath("./div[1]/a[2]/h2/text()")[0] #获取用户名
            print(name)
            content = item.xpath("./a[@class='contentHerf']/div/span/text()")[0] #获取内容
            print(content)
            star_number = item.xpath("./div[@class='stats']/span[1]/i/text()")[0] #获取点赞数
            print(star_number)
            comment_number = item.xpath("./div[@class='stats']/span[2]/a/i/text()")[0] #获取评论数
            print(comment_number)
            print("*"*100)

    def run(self):
        '''函数的主要逻辑实现
        '''
        url = self.url.format(1) #获取到url
        html = self.parse_url(url) #请求url
        self.parse_content(html) #解析页面内容并把内容存入内容队列

if __name__ == "__main__":
    qiubai = Qiubai_spider()
    qiubai.run()

演示效果