时间:2023-02-12 10:53:52 | 栏目:Python代码 | 点击:次
前言
自动化测试中我们存放数据无非是使用文件或者数据库,那么文件可以是csv,xlsx,xml,甚至是txt文件,通常excel文件往往是我们的首选,无论是编写测试用例还是存放测试数据,excel都是很方便的。那么今天我们就把不同模块处理excel文件的方法做个总结,直接做封装,方便我们以后直接使用,增加工作效率。
openpyxl
openpyxl是个第三方库,首先我们使用命令 pip install openpyxl 直接安装
注:openpyxl操作excel时,行号和列号都是从1开始计算的
封装代码
""" ------------------------------------ @Time : 2019/5/13 18:00 @Auth : linux超 @File : ParseExcel.py @IDE : PyCharm @Motto: Real warriors,dare to face the bleak warning,dare to face the incisive error! ------------------------------------ """ from openpyxl import load_workbook from openpyxl.styles import Font from openpyxl.styles.colors import BLACK from collections import namedtuple class ParseExcel(object): """解析excel文件""" def __init__(self, filename, sheet_name=None): try: self.filename = filename self.sheet_name = sheet_name self.wb = load_workbook(self.filename) if self.sheet_name is None: self.work_sheet = self.wb.active else: self.work_sheet = self.wb[self.sheet_name] except FileNotFoundError as e: raise e def get_max_row_num(self): """获取最大行号""" max_row_num = self.work_sheet.max_row return max_row_num def get_max_column_num(self): """获取最大列号""" max_column = self.work_sheet.max_column return max_column def get_cell_value(self, coordinate=None, row=None, column=None): """获取指定单元格的数据""" if coordinate is not None: try: return self.work_sheet[coordinate].value except Exception as e: raise e elif coordinate is None and row is not None and column is not None: if isinstance(row, int) and isinstance(column, int): return self.work_sheet.cell(row=row, column=column).value else: raise TypeError('row and column must be type int') else: raise Exception("Insufficient Coordinate of cell!") def get_row_value(self, row): """获取某一行的数据""" column_num = self.get_max_column_num() row_value = [] if isinstance(row, int): for column in range(1, column_num + 1): values_row = self.work_sheet.cell(row, column).value row_value.append(values_row) return row_value else: raise TypeError('row must be type int') def get_column_value(self, column): """获取某一列数据""" row_num = self.get_max_column_num() column_value = [] if isinstance(column, int): for row in range(1, row_num + 1): values_column = self.work_sheet.cell(row, column).value column_value.append(values_column) return column_value else: raise TypeError('column must be type int') def get_all_value_1(self): """获取指定表单的所有数据(除去表头)""" max_row_num = self.get_max_row_num() max_column = self.get_max_column_num() values = [] for row in range(2, max_row_num + 1): value_list = [] for column in range(1, max_column + 1): value = self.work_sheet.cell(row, column).value value_list.append(value) values.append(value_list) return values def get_all_value_2(self): """获取指定表单的所有数据(除去表头)""" rows_obj = self.work_sheet.iter_rows(min_row=2, max_row=self.work_sheet.max_row, values_only=True) # 指定values_only 会直接提取数据不需要再使用cell().value values = [] for row_tuple in rows_obj: value_list = [] for value in row_tuple: value_list.append(value) values.append(value_list) return values def get_excel_title(self): """获取sheet表头""" title_key = tuple(self.work_sheet.iter_rows(max_row=1, values_only=True))[0] return title_key def get_listdict_all_value(self): """获取所有数据,返回嵌套字典的列表""" sheet_title = self.get_excel_title() all_values = self.get_all_value_2() value_list = [] for value in all_values: value_list.append(dict(zip(sheet_title, value))) return value_list def get_list_nametuple_all_value(self): """获取所有数据,返回嵌套命名元组的列表""" sheet_title = self.get_excel_title() values = self.get_all_value_2() excel = namedtuple('excel', sheet_title) value_list = [] for value in values: e = excel(*value) value_list.append(e) return value_list def write_cell(self, row, column, value=None, bold=True, color=BLACK): """ 指定单元格写入数据 :param work_sheet: :param row: 行号 :param column: 列号 :param value: 待写入数据 :param bold: 加粗, 默认加粗 :param color: 字体颜色,默认黑色 :return: """ try: if isinstance(row, int) and isinstance(column, int): cell_obj = self.work_sheet.cell(row, column) cell_obj.font = Font(color=color, bold=bold) cell_obj.value = value self.wb.save(self.filename) else: raise TypeError('row and column must be type int') except Exception as e: raise e if __name__ == '__main__': pe = ParseExcel('testdata.xlsx') # sheet = pe.get_sheet_object('testcase') column_row = pe.get_max_column_num() print('最大列号:', column_row) max_row = pe.get_max_row_num() print('最大行号:', max_row) # cell_value_1 = pe.get_cell_value(row=2, column=3) print('第%d行, 第%d列的数据为: %s' % (2, 3, cell_value_1)) cell_value_2 = pe.get_cell_value(coordinate='A5') print('A5单元格的数据为: {}'.format(cell_value_2)) value_row = pe.get_row_value(3) print('第{}行的数据为:{}'.format(3, value_row)) value_column = pe.get_column_value(2) print('第{}列的数据为:{}'.format(2, value_column)) # values_1 = pe.get_all_value_1() print('第一种方式获取所有数据\n', values_1) values_2 = pe.get_all_value_2() print('第二种方式获取所有数据\n', values_2) title = pe.get_excel_title() print('表头为\n{}'.format(title)) dict_value = pe.get_listdict_all_value() print('所有数据组成的嵌套字典的列表:\n', dict_value) # namedtuple_value = pe.get_list_nametuple_all_value() print('所有数据组成的嵌套命名元组的列表:\n', namedtuple_value) pe.write_cell(1, 2, 'Tc_title')
# add by linux超 at 2019/05/22 15:58
上面这个封装如如果用来同时操作同一个excel文件的两个sheet写入数据时,会有点小bug(写完后你会发现两个表单有一个是没有数据的)
其实原因很简单:不同对象拥有自己独立的属性, 当你写操作的时候其实每个对象只针对自己的表单做了保存,所以最后一个对象写完数据后,只保存了自己的表单,其他的对象的表单实际是没有保存的。针对这个问题,对上面封装的代码进行了轻微改动
""" ------------------------------------ @Time : 2019/5/22 9:11 @Auth : linux超 @File : ParseExcel.py @IDE : PyCharm @Motto: Real warriors,dare to face the bleak warning,dare to face the incisive error! ------------------------------------ """ from openpyxl import load_workbook from openpyxl.styles import Font from openpyxl.styles.colors import BLACK from collections import namedtuple class ParseExcel(object): """解析excel文件""" def __init__(self, filename): try: self.filename = filename self.__wb = load_workbook(self.filename) except FileNotFoundError as e: raise e def get_max_row_num(self, sheet_name): """获取最大行号""" max_row_num = self.__wb[sheet_name].max_row return max_row_num def get_max_column_num(self, sheet_name): """获取最大列号""" max_column = self.__wb[sheet_name].max_column return max_column def get_cell_value(self, sheet_name, coordinate=None, row=None, column=None): """获取指定单元格的数据""" if coordinate is not None: try: return self.__wb[sheet_name][coordinate].value except Exception as e: raise e elif coordinate is None and row is not None and column is not None: if isinstance(row, int) and isinstance(column, int): return self.__wb[sheet_name].cell(row=row, column=column).value else: raise TypeError('row and column must be type int') else: raise Exception("Insufficient Coordinate of cell!") def get_row_value(self, sheet_name, row): """获取某一行的数据""" column_num = self.get_max_column_num(sheet_name) row_value = [] if isinstance(row, int): for column in range(1, column_num + 1): values_row = self.__wb[sheet_name].cell(row, column).value row_value.append(values_row) return row_value else: raise TypeError('row must be type int') def get_column_value(self, sheet_name, column): """获取某一列数据""" row_num = self.get_max_column_num(sheet_name) column_value = [] if isinstance(column, int): for row in range(1, row_num + 1): values_column = self.__wb[sheet_name].cell(row, column).value column_value.append(values_column) return column_value else: raise TypeError('column must be type int') def get_all_value_1(self, sheet_name): """获取指定表单的所有数据(除去表头)""" max_row_num = self.get_max_row_num(sheet_name) max_column = self.get_max_column_num(sheet_name) values = [] for row in range(2, max_row_num + 1): value_list = [] for column in range(1, max_column + 1): value = self.__wb[sheet_name].cell(row, column).value value_list.append(value) values.append(value_list) return values def get_all_value_2(self, sheet_name): """获取指定表单的所有数据(除去表头)""" rows_obj = self.__wb[sheet_name].iter_rows(min_row=2, max_row=self.__wb[sheet_name].max_row, values_only=True) values = [] for row_tuple in rows_obj: value_list = [] for value in row_tuple: value_list.append(value) values.append(value_list) return values def get_excel_title(self, sheet_name): """获取sheet表头""" title_key = tuple(self.__wb[sheet_name].iter_rows(max_row=1, values_only=True))[0] return title_key def get_listdict_all_value(self, sheet_name): """获取所有数据,返回嵌套字典的列表""" sheet_title = self.get_excel_title(sheet_name) all_values = self.get_all_value_2(sheet_name) value_list = [] for value in all_values: value_list.append(dict(zip(sheet_title, value))) return value_list def get_list_nametuple_all_value(self, sheet_name): """获取所有数据,返回嵌套命名元组的列表""" sheet_title = self.get_excel_title(sheet_name) values = self.get_all_value_2(sheet_name) excel = namedtuple('excel', sheet_title) value_list = [] for value in values: e = excel(*value) value_list.append(e) return value_list def write_cell(self, sheet_name, row, column, value=None, bold=True, color=BLACK): if isinstance(row, int) and isinstance(column, int): try: cell_obj = self.__wb[sheet_name].cell(row, column) cell_obj.font = Font(color=color, bold=bold) cell_obj.value = value self.__wb.save(self.filename) except Exception as e: raise e else: raise TypeError('row and column must be type int') if __name__ == '__main__': pe = ParseExcel('testdata.xlsx') print(pe.get_all_value_2('division')) print(pe.get_list_nametuple_all_value('division')) column_row = pe.get_max_column_num('division') print('最大列号:', column_row) max_row = pe.get_max_row_num('division') print('最大行号:', max_row) cell_value_1 = pe.get_cell_value('division', row=2, column=3) print('第%d行, 第%d列的数据为: %s' % (2, 3, cell_value_1)) cell_value_2 = pe.get_cell_value('division', coordinate='A5') print('A5单元格的数据为: {}'.format(cell_value_2)) value_row = pe.get_row_value('division', 3) print('第{}行的数据为:{}'.format(3, value_row)) value_column = pe.get_column_value('division', 2) print('第{}列的数据为:{}'.format(2, value_column)) values_1 = pe.get_all_value_1('division') print('第一种方式获取所有数据\n', values_1) values_2 = pe.get_all_value_2('division') print('第二种方式获取所有数据\n', values_2) title = pe.get_excel_title('division') print('表头为\n{}'.format(title)) dict_value = pe.get_listdict_all_value('division') print('所有数据组成的嵌套字典的列表:\n', dict_value) namedtuple_value = pe.get_list_nametuple_all_value('division') print('所有数据组成的嵌套命名元组的列表:\n', namedtuple_value) pe.write_cell('division', 1, 2, 'Tc_title')
xlrd
安装xlrd,此模块只支持读操作, 如果要写需要使用xlwt或者使用xlutils配合xlrd, 但是使用xlwt只能对新的excel文件进行写操作,无法对原有文件进行写, 所以这里选择是用xlutils
但是还有一个问题就是,如果使用xlutils, 那么我们的excel文件需要以.xls 为后缀。因为以xlsx为后缀无法实现写,会报错(亲测,因为formatting_info参数还没有对新版本的xlsx的格式完成兼容)
注:xlrd操作excel时,行号和列号都是从0开始计算的
封装代码
""" ------------------------------------ @Time : 2019/5/13 21:22 @Auth : linux超 @File : ParseExcel_xlrd.py @IDE : PyCharm @Motto: Real warriors,dare to face the bleak warning,dare to face the incisive error! ------------------------------------ """ import xlrd from xlutils import copy from collections import namedtuple class ParseExcel(object): # xlrd 解析excel, 行号和列号都是从0开始的 def __init__(self, filename, sheet): try: self.filename = filename self.sheet = sheet self.wb = xlrd.open_workbook(self.filename, formatting_info=True) if isinstance(sheet, str): self.sheet = self.wb.sheet_by_name(sheet) elif isinstance(sheet, int): self.sheet = self.wb.sheet_by_index(sheet) else: raise TypeError('sheet must be int or str') except Exception as e: raise e def get_max_row(self): """获取表单的最大行号""" max_row_num = self.sheet.nrows return max_row_num def get_max_column(self): """获取表单的最大列号""" min_row_num = self.sheet.ncols return min_row_num def get_cell_value(self, row, column): """获取某个单元格的数据""" if isinstance(row, int) and isinstance(column, int): values = self.sheet.cell(row-1, column-1).value return values else: raise TypeError('row and column must be type int') def get_row_values(self, row): """获取某一行的数据""" if isinstance(row, int): values = self.sheet.row_values(row-1) return values else: raise TypeError('row must be type int') def get_column_values(self, column): """获取某一列的数据""" if isinstance(column, int): values = self.sheet.col_values(column-1) return values else: raise TypeError('column must be type int') def get_table_title(self): """获取表头""" table_title = self.get_row_values(1) return table_title def get_all_values_dict(self): """获取所有的数据,不包括表头,返回一个嵌套字典的列表""" max_row = self.get_max_row() table_title = self.get_table_title() value_list = [] for row in range(2, max_row): values = self.get_row_values(row) value_list.append(dict(zip(table_title, values))) return value_list def get_all_values_nametuple(self): """获取所有的数据,不包括表头,返回一个嵌套命名元组的列表""" table_title = self.get_table_title() max_row = self.get_max_row() excel = namedtuple('excel', table_title) value_list = [] for row in range(2, max_row): values = self.get_row_values(row) e = excel(*values) value_list.append(e) return value_list def write_value(self, sheet_index, row, column, value): """写入某个单元格数据""" if isinstance(row, int) and isinstance(column, int): if isinstance(sheet_index, int): wb = copy.copy(self.wb) worksheet = wb.get_sheet(sheet_index) worksheet.write(row-1, column-1, value) wb.save(self.filename) else: raise TypeError('{} must be int'.format(sheet_index)) else: raise TypeError('{} and {} must be int'.format(row, column)) if __name__ == '__main__': pe = ParseExcel('testdata.xls', 'testcase') print('最大行号:', pe.get_max_row()) print('最大列号:', pe.get_max_column()) print('第2行第3列数据:', pe.get_cell_value(2, 3)) print('第2行数据', pe.get_row_values(2)) print('第3列数据', pe.get_column_values(3)) print('表头:', pe.get_table_title()) print('所有的数据返回嵌套字典的列表:', pe.get_all_values_dict()) print('所有的数据返回嵌套命名元组的列表:', pe.get_all_values_nametuple()) pe.write_value(0, 1, 3, 'test')
pandas
pandas是一个做数据分析的库, 总是感觉在自动化测试中使用pandas解析excel文件读取数据有点大材小用,不论怎样吧,还是把pandas解析excel文件写一下把
我这里只封装了读,写的话我这有点小问题,后面改好再追加代码吧。
请先pip install pandas安装pandas
封装代码
""" ------------------------------------ @Time : 2019/5/13 14:00 @Auth : linux超 @File : ParseExcel_pandas.py @IDE : PyCharm @Motto: Real warriors,dare to face the bleak warning,dare to face the incisive error! ------------------------------------ """ import pandas as pd class ParseExcel(object): def __init__(self, filename, sheet_name=None): try: self.filename = filename self.sheet_name = sheet_name self.df = pd.read_excel(self.filename, self.sheet_name) except Exception as e: raise e def get_row_num(self): """获取行号组成的列表, 从0开始的""" row_num_list = self.df.index.values return row_num_list def get_cell_value(self, row, column): """获取某一个单元格的数据""" try: if isinstance(row, int) and isinstance(column, int): cell_value = self.df.ix[row-2, column-1] # ix的行参数是按照有效数据行,且从0开始 return cell_value else: raise TypeError('row and column must be type int') except Exception as e: raise e def get_table_title(self): """获取表头, 返回列表""" table_title = self.df.columns.values return table_title def get_row_value(self, row): """获取某一行的数据, 行号从1开始""" try: if isinstance(row, int): row_data = self.df.ix[row-2].values return row_data else: raise TypeError('row must be type int') except Exception as e: raise e def get_column_value(self, col_name): """获取某一列数据""" try: if isinstance(col_name, str): col_data = self.df[col_name].values return col_data else: raise TypeError('col_name must be type str') except Exception as e: raise e def get_all_value(self): """获取所有的数据,不包括表头, 返回嵌套字典的列表""" rows_num = self.get_row_num() table_title = self.get_table_title() values_list = [] for i in rows_num: row_data = self.df.ix[i, table_title].to_dict() values_list.append(row_data) return values_list if __name__ == '__main__': pe = ParseExcel('testdata.xlsx', 'testcase') print(pe.get_row_num()) print(pe.get_table_title()) print(pe.get_all_value()) print(pe.get_row_value(2)) print(pe.get_cell_value(2, 3)) print(pe.get_column_value('Tc_title'))
总结
使用了3种方法,4个库 xlrd,openpyxl,xlwt,pandas 操作excel文件,个人感觉还是使用openpyxl比较适合在自动化中使用,当然不同人有不同选择,用哪个区别也不是很大。
以上3种方法,都可以拿来直接使用,不需要再做封装了 !