python爬虫实例100条|用python爬取去哪儿游记攻略为十月假期做准备。。。爬虫之路,永无止境!
用python爬取去哪儿游记攻略为十月假期做准备。。。爬虫之路,永无止境! 热热闹闹的开学季又来了,小伙伴们又可以在一起玩耍了,不对是在一起学习了,哈哈。
文章图片
再过几周就是国庆假期,想想还是很激动的,我决定给大家做个游记爬虫,大家早做准备。。嘿嘿
文章图片
代码操作展示:
今天目标地址:https://travel.qunar.com/place/
开发环境:
windows10
python3.6
开发工具:
pycharm
库:
tkinter、re、os、lxml、threading、xlwt、xlrd
1.首先先将全国所有的城市名称和id拿到
文章图片
2.右击检查,进行抓包,找到数据所在的包
文章图片
3.发送请求,获取响应,解析响应
# 发送请求,获取响应,解析响应
response = session.get(self.start_url, headers=self.headers).html
# 提取所有目的地(城市)的url
city_url_list = response.xpath(
'//*[@id="js_destination_recommend"]/div[2]/div[1]/div[2]/dl/dd/div/ul/li/a/@href')
city_id_list = [''.join(re.findall(r'-cs(.*?)-', i)) for i in city_url_list]
# 提取所有的城市名称
city_name_list = response.xpath(
'//*[@id="js_destination_recommend"]/div[2]/div[1]/div[2]/dl/dd/div/ul/li/a/text()')
4.随机点一个城市,进入该城市,查看游记攻略,本文选的是上海
文章图片
5.进行抓包,查找需要的信息
文章图片
# 提取游记作者
author_list = html.xpath('//span[@class="user_name"]/a/text()')
# 出发时间
date_list = html.xpath('//span[@class="date"]/text()')
# 游玩时间
days_list = html.xpath('//span[@class="days"]/text()')
# 阅读量
read_list = html.xpath('//span[@class="icon_view"]/span/text()')
# 点赞量
like_count_list = html.xpath('//span[@class="icon_love"]/span/text()')
# 评论量
icon_list = html.xpath('//span[@class="icon_comment"]/span/text()')
# 游记地址
text_url_list = html.xpath('//h3[@class="tit"]/a/@href')
6.进行翻页抓包,第二页为异步加载
文章图片
文章图片
7.在上一张图中点击Headers,获取游记攻略的请求地址
文章图片
这样思路是不是清晰一点,来观察一下请求地址https://travel.qunar.com/place/api/html/books/dist/299878?sortField=0&pageSize=10&page=2
其中299878是上海城市的id page是页数
源码展示:
# !/usr/bin/nev python
# -*-coding:utf8-*-
import tkinter as tk
import re
from lxml import etree
import tkinter.messagebox as msgbox
from requests_html import HTMLSession
from threading import Thread
import os, xlwt, xlrd
from xlutils.copy import copy
session = HTMLSession()class TKSpider(object):def __init__(self):
# 定义循环条件
self.is_running = True
# 定义起始的页码
self.start_page = 1
"""定义可视化窗口,并设置窗口和主题大小布局"""
self.window = tk.Tk()
self.window.title('爬虫数据采集')
self.window.geometry('1000x800')"""创建label_user按钮,与说明书"""
self.label_user = tk.Label(self.window, text='请选择城市输入序号:', font=('Arial', 12), width=30, height=2)
self.label_user.pack()
"""创建label_user关联输入"""
self.entry_user = tk.Entry(self.window, show=None, font=('Arial', 14))
self.entry_user.pack(after=self.label_user)"""创建Text富文本框,用于按钮操作结果的展示"""
# 定义富文本框滑动条
scroll = tk.Scrollbar()
# 放到窗口的右侧, 填充Y竖直方向
scroll.pack(side=tk.RIGHT, fill=tk.Y)
self.text1 = tk.Text(self.window, font=('Arial', 12), width=75, height=20)
self.text1.pack()
# 两个控件关联
scroll.config(command=self.text1.yview)
self.text1.config(yscrollcommand=scroll.set)"""定义按钮1,绑定触发事件方法"""
self.button_1 = tk.Button(self.window, text='下载游记', font=('Arial', 12), width=10, height=1, command=self.parse_hit_click_1)
self.button_1.pack(before=self.text1)"""定义按钮2,绑定触发事件方法"""
self.button_2 = tk.Button(self.window, text='清除', font=('Arial', 12), width=10, height=1, command=self.parse_hit_click_2)
self.button_2.pack(anchor="e")"""定义富文本文字的显示"""
self.start_url = 'https://travel.qunar.com/place/'
self.headers = {'user-agent': 'Mozilla/5.0 (Windows NT 10.0;
Win64;
x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/92.0.4515.159 Safari/537.36'
}
# 发送请求,获取响应,解析响应
response = session.get(self.start_url, headers=self.headers).html
# 提取所有目的地(城市)的url
city_url_list = response.xpath(
'//*[@id="js_destination_recommend"]/div[2]/div[1]/div[2]/dl/dd/div/ul/li/a/@href')
city_id_list = [''.join(re.findall(r'-cs(.*?)-', i)) for i in city_url_list]
# 提取所有的城市名称
city_name_list = response.xpath(
'//*[@id="js_destination_recommend"]/div[2]/div[1]/div[2]/dl/dd/div/ul/li/a/text()')
self.dict_data = https://www.it610.com/article/dict(zip(city_id_list, city_name_list))
text_info =''
for city_id, city_name in zip(city_id_list, city_name_list):
text_info += city_id + ":" + city_name + '\t\t'
self.text1.insert("insert", text_info)def parse_hit_click_1(self):
"""线程关联:定义触发事件1, 将执行结果显示在文本框中"""
Thread(target=self.parse_start_url_job).start()def parse_city_id_name_info(self):
"""
富文本内容展示:
:return:
"""
# 发送请求,获取响应,解析响应
response = session.get(self.start_url, headers=self.headers).html
# 提取所有目的地(城市)的url
city_url_list = response.xpath(
'//*[@id="js_destination_recommend"]/div[2]/div[1]/div[2]/dl/dd/div/ul/li/a/@href')
city_id_list = [''.join(re.findall(r'-cs(.*?)-', i)) for i in city_url_list]
# 提取所有的城市名称
city_name_list = response.xpath(
'//*[@id="js_destination_recommend"]/div[2]/div[1]/div[2]/dl/dd/div/ul/li/a/text()')
self.dict_data = https://www.it610.com/article/dict(zip(city_id_list, city_name_list))
text_info =''
for city_id, city_name in zip(city_id_list, city_name_list):
text_info += city_id + ":" + city_name + '\t\t****'
self.text1.insert("insert", text_info)def parse_start_url_job(self):
# 从输入窗口获取输入
city_id = self.entry_user.get()
self.text1.delete("1.0", "end")
# 异常捕捉,是否存在key值
try:
city_name = self.dict_data[city_id]
self.parse_request_yj(city_id, city_name)
except:
# 报错提示
msgbox.showerror(title='错误', message='检测到瞎搞,请重新输入!')def parse_request_yj(self, city_id, city_name):
"""
拼接游记地址,进行翻页
:param city_id: 城市的id
:param city_name: 城市的名称
:return:
"""
while self.is_running:
yj_url = f'https://travel.qunar.com/place/api/html/books/dist/{
city_id}?sortField=0&pageSize=10&page={
self.start_page}'
response = session.get(yj_url, headers=self.headers).json()['data'].replace('\\', '')
html = etree.HTML(response)
# 提取游记的标题:
title_list = html.xpath('//h3[@class="tit"]/a/text()')
# 死循环的终止条件
if not title_list:
print("该城市已经采集到最后一页----------城市切换中---------logging!!!")
break
else:
# 提取游记作者
author_list = html.xpath('//span[@class="user_name"]/a/text()')
# 出发时间
date_list = html.xpath('//span[@class="date"]/text()')
# 游玩时间
days_list = html.xpath('//span[@class="days"]/text()')
# 阅读量
read_list = html.xpath('//span[@class="icon_view"]/span/text()')
# 点赞量
like_count_list = html.xpath('//span[@class="icon_love"]/span/text()')
# 评论量
icon_list = html.xpath('//span[@class="icon_comment"]/span/text()')
# 游记地址
text_url_list = html.xpath('//h3[@class="tit"]/a/@href')
"""构造保存数据格式字典"""
for a, b, c, d, e, f, g, h in zip(title_list, author_list, date_list, days_list, read_list, like_count_list, icon_list, text_url_list):
dict_dd = {city_name: [a, b, c, d, e, f, g, h]
}
self.text1.insert("insert", f"{
a}-----------采集完成!!!" + '\n')
self.SaveExcels(dict_dd)
self.start_page += 1def SaveExcels(self, data):
# 获取表的名称
sheet_name = [i for i in data.keys()][0]
# 创建保存excel表格的文件夹
# os.getcwd() 获取当前文件路径
os_mkdir_path = os.getcwd() + '/数据/'
# 判断这个路径是否存在,不存在就创建
if not os.path.exists(os_mkdir_path):
os.mkdir(os_mkdir_path)
# 判断excel表格是否存在工作簿文件名称
os_excel_path = os_mkdir_path + '数据.xls'
if not os.path.exists(os_excel_path):
# 不存在,创建工作簿(也就是创建excel表格)
workbook = xlwt.Workbook(encoding='utf-8')
"""工作簿中创建新的sheet表"""# 设置表名
worksheet1 = workbook.add_sheet(sheet_name, cell_overwrite_ok=True)
"""设置sheet表的表头"""
sheet1_headers = ('游记标题', '发布游记的作者名称', '出发时间', '游玩时间', '阅读量', '点赞量', '评论量', '游记地址')
# 将表头写入工作簿
for header_num in range(0, len(sheet1_headers)):
# 设置表格长度
worksheet1.col(header_num).width = 2560 * 3
# 写入表头行,列,内容
worksheet1.write(0, header_num, sheet1_headers[header_num])
# 循环结束,代表表头写入完成,保存工作簿
workbook.save(os_excel_path)
"""=============================已有工作簿添加新表==============================================="""
# 打开工作薄
workbook = xlrd.open_workbook(os_excel_path)
# 获取工作薄中所有表的名称
sheets_list = workbook.sheet_names()
# 如果表名称:字典的key值不在工作簿的表名列表中
if sheet_name not in sheets_list:
# 复制先有工作簿对象
work = copy(workbook)
# 通过复制过来的工作簿对象,创建新表-- 保留原有表结构
sh = work.add_sheet(sheet_name)
# 给新表设置表头
excel_headers_tuple = ('游记标题', '发布游记的作者名称', '出发时间', '游玩时间', '阅读量', '点赞量', '评论量', '游记地址')
for head_num in range(0, len(excel_headers_tuple)):
sh.col(head_num).width = 2560 * 3
#行,列,内容,样式
sh.write(0, head_num, excel_headers_tuple[head_num])
work.save(os_excel_path)
"""========================================================================================="""
# 判断工作簿是否存在
if os.path.exists(os_excel_path):
# 打开工作簿
workbook = xlrd.open_workbook(os_excel_path)
# 获取工作薄中所有表的个数
sheets = workbook.sheet_names()
for i in range(len(sheets)):
for name in data.keys():
worksheet = workbook.sheet_by_name(sheets[i])
# 获取工作薄中所有表中的表名与数据名对比
if worksheet.name == name:
# 获取表中已存在的行数
rows_old = worksheet.nrows
# 将xlrd对象拷贝转化为xlwt对象
new_workbook = copy(workbook)
# 获取转化后的工作薄中的第i张表
new_worksheet = new_workbook.get_sheet(i)
for num in range(0, len(data[name])):
new_worksheet.write(rows_old, num, data[name][num])
new_workbook.save(os_excel_path)def parse_hit_click_2(self):
"""定义触发事件2,删除文本框中内容"""
self.entry_user.delete(0, "end")
# self.entry_passwd.delete(0, "end")
self.text1.delete("1.0", "end")
self.parse_city_id_name_info()def center(self):
"""创建窗口居中函数方法"""
ws = self.window.winfo_screenwidth()
hs = self.window.winfo_screenheight()
x = int((ws / 2) - (800 / 2))
y = int((hs / 2) - (600 / 2))
self.window.geometry('{}x{}+{}+{}'.format(800, 600, x, y))def run_loop(self):
"""禁止修改窗体大小规格"""
self.window.resizable(False, False)
"""窗口居中"""
self.center()
"""窗口维持--持久化"""
self.window.mainloop()if __name__ == '__main__':
t = TKSpider()
t.run_loop()
代码仅供学习!! 代码仅供学习!! 代码仅供学习!! 祝大家学习python顺利! 【python爬虫实例100条|用python爬取去哪儿游记攻略为十月假期做准备。。。爬虫之路,永无止境!】
文章图片
推荐阅读
- python学习之|python学习之 实现QQ自动发送消息
- 逻辑回归的理解与python示例
- python自定义封装带颜色的logging模块
- 【Leetcode/Python】001-Two|【Leetcode/Python】001-Two Sum
- Python基础|Python基础 - 练习1
- Python爬虫|Python爬虫 --- 1.4 正则表达式(re库)
- Python(pathlib模块)
- python青少年编程比赛_第十一届蓝桥杯大赛青少年创意编程组比赛细则
- Python数据分析(一)(Matplotlib使用)
- 爬虫数据处理HTML转义字符