怎么在python中利用openpyxl和xlsxwriter对Excel进行操作
本篇文章给大家分享的是有关怎么在python中利用openpyxl和xlsxwriter对Excel进行操作,小编觉得挺实用的,因此分享给大家学习,希望大家阅读完这篇文章后可以有所收获,话不多说,跟着小编一起来看看吧。
简介
Python中操作Excel的库非常多,为开发者提供了多种选择,如:xlrd
、 xlwt
、xlutils
、xlwings
、pandas
、 win32com
、openpyxl
、xlsxwriter
等等。
其中:
前三个一般混合使用,对Excel读写操作,适合旧版Excel,仅支持 xls 文件;
win32com
库功能丰富,性能强大,适用于Windows;xlwings
稍次于前者,但同样功能丰富;pandas
适合处理大量数据;xlsxwriter
适合大量数据的写操作,支持图片/表格/图表/筛选/格式/公式等;openpyxl
读写均可,简单易用,功能广泛,可插入图表等,类似前者。
以下主要描述一下后两种(
openpyxl
、xlsxwriter
)的简单使用
2.Excel库的使用
1.目标
2.openpyxl
的使用
2.1.安装
pip install openpyxl
2.2.写入Excel
import osfrom openpyxl import Workbookfrom openpyxl.styles import Alignment, Font, colors, PatternFillfrom openpyxl.utils import get_column_letterFILE_PATH = os.path.join(os.path.dirname(__file__), 'files/')def write_test(): wb = Workbook() filename = FILE_PATH + '/openpyxl_test.xlsx' # 活动sheet ws1 = wb.active ws1.title = "Test-1" # 列表追加 for row in range(1, 10): ws1.append(range(9)) # 创建sheet ws2 = wb.create_sheet(title="Test-2") # 合并单元格 ws2.merge_cells('F5:I5') # 拆分 # ws2.unmerge_cells('F5:I5') # 单元赋值 ws2['F5'] = 'hello world' # 居中 ws2['F5'].alignment = Alignment(horizontal='center', vertical='center') # sheet标签颜色 ws2.sheet_properties.tabColor = '1072BA' # 字体样式 bold_itatic_12_font = Font(name='仿宋', size=12, italic=True, color=BLUE, bold=True) ws2['F5'].font = bold_itatic_12_font # 背景颜色 bg_color = PatternFill('solid', fgColor='1874CD') ws2['F5'].fill = bg_color # 行高列宽 ws2.row_dimensions[5].height = 40 # 第 5 行 ws2.column_dimensions['F'].width = 30 # F 列 ws3 = wb.create_sheet(title="Test-3") for row in range(10, 20): for col in range(10, 20): ws3.cell(column=col, row=row, value="0}".format(get_column_letter(col))) print(ws3['S10'].value) # 保存 wb.save(filename)
2.3.读取Excel
from openpyxl import load_workbookdef read_test(filename): wb = load_workbook(filename) print('取得所有工作表的表名 :') print(wb.sheetnames, '\n') print('取得某张工作表 :') # sheet = wb['Sheet1'] # sheet = wb.worksheets[0] sheet = wb[wb.sheetnames[0]] print(type(sheet)) print('表名: ' + sheet.title, '\n') print('取得活动工作表 :') active_sheet = wb.active print('表名: ' + active_sheet.title, '\n') print('获取工作表的大小:') print('总行数: ' + str(active_sheet.max_row)) print('总列数: ' + str(active_sheet.max_column)) print('\n获取单元格数据:') for row in range(sheet.max_row): for col in range(sheet.max_column): print(f"第 {row + 1} 行 {col + 1} 列:", sheet.cell(row=row + 1, column=col + 1).value) print('\n获取行数据:') for i, cell_object in enumerate(list(sheet.rows)): cell_lst = [cell.value for cell in cell_object] print(f'第 {i + 1} 行:', cell_lst)
2.4.案例demo 数据源格式
# contents数据contents=[{ "uid": "1281948912", "group_name": "测试群-5", "domain": "ddos5.www.cn", "user_area": [ { "num": 1024, "region": "中国", "percent": 33.33 }, { "num": 1022, "region": "中国香港", "percent": 33.33 }, { "num": 1021, "region": "新加坡", "percent": 33.33 } ], "gf_area": [ { "num": 5680, "region": "中国香港", "percent": 97.8 }, { "num": 60, "region": "新加坡", "percent": 0.8 }, { "num": 55, "region": "美西", "percent": 0.8 } ], "sip_area": { "waf_ip":["aliyunwaf.com.cn"], "sip":["13.75.120.253","18.163.46.57"], "isp_region":[ { "country": "中国香港", "isp": "microsoft.com" }, { "country": "中国香港", "isp": "amazon.com" } ] } },]
写入Excel
import osimport timefrom openpyxl import Workbook, load_workbookfrom openpyxl.styles import Alignment, Font, colors, PatternFillFILE_PATH = os.path.join(os.path.dirname(__file__), 'files/')# 颜色BLACK = colors.COLOR_INDEX[0]WHITE = colors.COLOR_INDEX[1]RED = colors.COLOR_INDEX[2]DARKRED = colors.COLOR_INDEX[8]BLUE = colors.COLOR_INDEX[4]DARKBLUE = colors.COLOR_INDEX[12]GREEN = colors.COLOR_INDEX[3]DARKGREEN = colors.COLOR_INDEX[9]YELLOW = colors.COLOR_INDEX[5]DARKYELLOW = colors.COLOR_INDEX[19]def export_gf_excel_test(filename=None, sheetName=None, contents=None): filename = filename if filename else 'openpyxl_Test.xlsx' sheetName = sheetName if sheetName else '测试' contents = contents if contents else [] # 新建工作簿 wb = Workbook() ws = wb.worksheets[0] # 设置sheet名称 ws.title = sheetName # sheet标签颜色 ws.sheet_properties.tabColor = '1072BA' # 居中 pos_center = Alignment(horizontal='center', vertical='center') # 字体样式 bold_12_font = Font(name='仿宋', size=12, italic=False, color=BLACK, bold=True) # 背景颜色 bg_color = PatternFill('solid', fgColor='4DCFF6') # 设置标题 # 合并 merge_lst = [ 'A1:A3', 'B1:B3', 'C1:C3', 'D1:R1', 'S1:AA1', 'AB1:AE1', 'D2:F2', 'G2:I2', 'J2:L2', 'M2:O2', 'P2:R2', 'S2:U2', 'V2:X2', 'Y2:AA2', 'AB2:AB3', 'AC2:AC3', 'AD2:AD3', 'AE2:AE3' ] [ws.merge_cells(c) for c in merge_lst] # 填充字段 title_dic = { 'A1': 'UID', 'B1': '钉钉群', 'C1': '域名', 'D1': '用户区域', 'S1': '高防区域', 'AB1': '源站区域', 'D2': 'TOP1', 'G2': 'TOP2', 'J2': 'TOP3', 'M2': 'TOP4', 'P2': 'TOP5', 'S2': 'TOP1', 'V2': 'TOP2', 'Y2': 'TOP3', 'AB2': 'WAF IP', 'AC2': '源站IP', 'AD2': '源站IP区域', 'AE2': '运营商' } line3_v = ['物理区域', '请求量', '占比'] * 8 line3_k = [chr(i) + '3' for i in range(68, 91)] + ['AA3'] title_dic.update(dict(zip(line3_k, line3_v))) for k, v in title_dic.items(): ws[k].value = v ws[k].font = bold_12_font ws[k].alignment = pos_center ws[k].fill = bg_color # 列宽 width_dic = { 'A': 30, 'B': 30, 'C': 30, 'AB': 16, 'AC': 16, 'AD': 16, 'AE': 16 } for k, v in width_dic.items(): ws.column_dimensions[k].width = v # 内容 for i, dic in enumerate(contents): user_gf_mod = {'region': '', 'num': '', 'percent': ''} user_area = dic['user_area'] gf_area = dic['gf_area'] sip_area = dic['sip_area'] # UID+域名 data = [dic['uid'], dic['group_name'], dic['domain']] # 用户区域 if not user_area: user_area = [user_gf_mod] * 5 else: user_area = list( map(lambda item: { 'region': item['region'], 'num': item['num'], 'percent': item['percent']}, user_area) ) [user_area.append(user_gf_mod) for _ in range(5 - len(user_area))] [data.extend(user_area[u].values()) for u in range(len(user_area))] # 高防区域 if not gf_area: gf_area = [user_gf_mod] * 3 else: gf_area = list( map(lambda item: { 'region': item['region'], 'num': item['num'], 'percent': item['percent']}, gf_area) ) [gf_area.append(user_gf_mod) for _ in range(3 - len(gf_area))] [data.extend(gf_area[g].values()) for g in range(len(gf_area))] # 源站区域 waf_ip = sip_area['waf_ip'] sip = sip_area['sip'] isp_region = sip_area['isp_region'] data.append(','.join(waf_ip)) if waf_ip else data.append('') data.append(','.join(sip)) if sip else data.append('') if not isp_region: data.extend([''] * 2) else: try: country = ','.join(map(lambda item: item['country'], isp_region)) isp = ','.join(map(lambda item: item['isp'] if item['isp'] else '暂未查到', isp_region)) data.append(country) data.append(isp) except Exception as e: print(e) print(isp_region) # 写入Excel ws.append(data) # 保存文件 wb.save(filename=filename)if __name__ == "__main__":curTime = ''.join(map(lambda i: str(i) if len(str(i)) >= 2 else '%02d' % i, [i for i in time.localtime()[:-4]])) filename = os.path.join(FILE_PATH, 'openpyxl_Test_{}.xlsx'.format(curTime)) export_gf_excel_test(filename, contents=contents)
xlsxwriter
的使用
1.安装
pip install XlsxWriter
2.写入Excel
import osimport timeimport jsonimport xlsxwriterFILE_PATH = os.path.join(os.path.dirname(__file__), 'files/')def export_gf_excel_test(filename=None, sheetName=None, contents=None): filename = filename if filename else 'xlsxwriter_Test.xlsx' sheetName = sheetName if sheetName else '测试' contents = contents if contents else [] # 新建 wb = xlsxwriter.Workbook(filename) ws = wb.add_worksheet(name=sheetName) # 设置风格 style1 = wb.add_format({ "bold": True, 'font_name': '仿宋', 'font_size': 12, # 'font_color': '#217346', 'bg_color': '#4DCFF6', "align": 'center', "valign": 'vcenter', 'text_wrap': 1 }) style2 = wb.add_format({ # "bold": True, # 'font_name': '仿宋', 'font_size': 11, 'font_color': '#217346', 'bg_color': '#E6EDEC', "align": 'center', "valign": 'vcenter', # 'text_wrap': 1 }) # 标题 ws.set_column('A1:AE1', None, style1) # 合并单元格: first_row, first_col, last_row, last_col # 第 1 行 ws.merge_range(0, 0, 2, 0, 'UID') ws.merge_range(0, 1, 2, 1, '钉钉群') ws.merge_range(0, 2, 2, 2, '域名') ws.merge_range(0, 3, 0, 17, '用户区域') ws.merge_range(0, 18, 0, 26, '高防区域') ws.merge_range(0, 27, 0, 30, '源站区域') # 第 2 行 user_tl2 = ['TOP' + str(i) for i in range(1, 6)] gf_tl2 = user_tl2[:3] [ws.merge_range(1, 3 * (i + 1), 1, 3 * (i + 2) - 1, name) for i, name in enumerate(user_tl2 + gf_tl2)] # 第 3 行 user_gf_tl3 = ['物理区域', '请求量', '占比'] * 8 sip_tl3 = ['WAF IP', '源站IP', '源站IP区域', '运营商'] [ws.write(2, 3 + i, name) for i, name in enumerate(user_gf_tl3)] [ws.merge_range(1, 27 + i, 2, 27 + i, name) for i, name in enumerate(sip_tl3)] # ws.write(11, 2, '=SUM(1:10)') # 增加公式 # ws.set_default_row(35) # 设置默认行高 # 设置列宽 ws.set_column(0, 2, 30) ws.set_column(3, 26, 10) ws.set_column(27, 30, 16) # 内容 for i, dic in enumerate(contents): user_gf_mod = {'region': '', 'num': '', 'percent': ''} user_area = dic['user_area'] gf_area = dic['gf_area'] sip_area = dic['sip_area'] # UID+域名 data = [dic['uid'], dic['group_name'], dic['domain']] # 用户区域 if not user_area: user_area = [user_gf_mod] * 5 else: user_area = list( map(lambda item: { 'region': item['region'], 'num': item['num'], 'percent': item['percent']}, user_area) ) [user_area.append(user_gf_mod) for _ in range(5 - len(user_area))] [data.extend(user_area[u].values()) for u in range(len(user_area))] # 高防区域 if not gf_area: gf_area = [user_gf_mod] * 3 else: gf_area = list( map(lambda item: { 'region': item['region'], 'num': item['num'], 'percent': item['percent']}, gf_area) ) [gf_area.append(user_gf_mod) for _ in range(3 - len(gf_area))] [data.extend(gf_area[g].values()) for g in range(len(gf_area))] # 源站区域 waf_ip = sip_area['waf_ip'] sip = sip_area['sip'] isp_region = sip_area['isp_region'] data.append(','.join(waf_ip)) if waf_ip else data.append('') data.append(','.join(sip)) if sip else data.append('') if not isp_region: data.extend([''] * 2) else: try: country = ','.join(map(lambda item: item['country'], isp_region)) isp = ','.join(map(lambda item: item['isp'] if item['isp'] else '暂未查到', isp_region)) data.append(country) data.append(isp) except Exception as e: print(e) print(isp_region) # 写入Excel ws.write_row('A' + str(i + 4), data, style2) # 保存关闭文件 wb.close()if __name__ == '__main__':curTime = ''.join(map(lambda i: str(i) if len(str(i)) >= 2 else '%02d' % i, [i for i in time.localtime()[:-4]])) filename = os.path.join(FILE_PATH, 'xlsxwriter_Test_{}.xlsx'.format(curTime)) export_gf_excel_test(filename, contents=contents)
以上就是怎么在python中利用openpyxl和xlsxwriter对Excel进行操作,小编相信有部分知识点可能是我们日常工作会见到或用到的。希望你能通过这篇文章学到更多知识。更多详情敬请关注编程网行业资讯频道。
免责声明:
① 本站未注明“稿件来源”的信息均来自网络整理。其文字、图片和音视频稿件的所属权归原作者所有。本站收集整理出于非商业性的教育和科研之目的,并不意味着本站赞同其观点或证实其内容的真实性。仅作为临时的测试数据,供内部测试之用。本站并未授权任何人以任何方式主动获取本站任何信息。
② 本站未注明“稿件来源”的临时测试数据将在测试完成后最终做删除处理。有问题或投稿请发送至: 邮箱/279061341@qq.com QQ/279061341