之前跟一些小伙伴有个讨论:
大概就是很多跟数据打交道的朋友都面对过很复杂的excel公式,有时嵌套层数特别多,肉眼观看很容易蒙圈。
有了这样的需求,我就有了解决问题的想法,说干就干,于是一个比较牛逼的excel公式格式化的工具就出现了。
效果体验
先看看效果吧:
=IF(C11>100%*C4,IF(C11<=200%*C4,C11*50%-C4*15%,C11*60%-C4*35%),IF(C11<=C4*50%,C11*30%,C11*40%-C4*5%))
的格式化结果是:
=IF( C11>100%*C4, IF( C11<=200%*C4, C11*50%-C4*15%, C11*60%-C4*35% ), IF( C11<=C4*50%, C11*30%, C11*40%-C4*5% ) )
(SMA(MAX(CLOSE-DELAY(CLOSE,1),0),12,1)/SMA(ABS(CLOSE-DELAY(CLOSE,1)),12,1)*100-MIN(SMA(MAX(CLOSE-DELAY( CLOSE,1),0),12,1)/SMA(ABS(CLOSE-DELAY(CLOSE,1)),12,1)*100,12))/(MAX(SMA(MAX(CLOSE-DELAY(CLOSE,1),0),12, 1)/SMA(ABS(CLOSE-DELAY(CLOSE,1)),12,1)*100,12)-MIN(SMA(MAX(CLOSE-DELAY(CLOSE,1),0),12,1)/SMA(ABS( CLOSE-DELAY(CLOSE,1)),12,1)*100,12))
的格式化结果为:
( SMA(MAX(CLOSE-DELAY(CLOSE,1),0),12,1) / SMA(ABS(CLOSE-DELAY(CLOSE,1)),12,1) * 100-MIN( SMA(MAX(CLOSE-DELAY(CLOSE,1),0),12,1) / SMA(ABS(CLOSE-DELAY(CLOSE,1)),12,1)*100, 12 ) ) / ( MAX( SMA(MAX(CLOSE-DELAY(CLOSE,1),0),12,1) / SMA(ABS(CLOSE-DELAY(CLOSE,1)),12,1)*100, 12 ) - MIN( SMA(MAX(CLOSE-DELAY(CLOSE,1),0),12,1) / SMA(ABS(CLOSE-DELAY(CLOSE,1)),12,1)*100, 12 ) )
=IF(ROW()>COLUMN(),"",IF(ROW()=COLUMN(),$B15,ROUNDDOWN($B15*INDIRECT(SUBSTITUTE(ADDRESS(1,3+COLUMN()-ROW(), 4),1,"")&56),0)))
的格式化结果为:
=IF( ROW()>COLUMN(), "", IF( ROW()=COLUMN(), $B15, ROUNDDOWN( $B15*INDIRECT( SUBSTITUTE(ADDRESS(1,3+COLUMN()-ROW(), 4),1,"") & 56 ), 0 ) ) )
(文末有体验网址)
不过接下来,将公布这套格式化程序的完整代码和开发思想,有技术能力的小伙伴可以考虑改进该代码。
完整代码
__author__ = 'xiaoxiaoming' from collections import deque import re class Node: def __init__(self, parent=None, tab_size=0): self.parent = parent self.tab_size = tab_size self.data = [] def is_single_node(self): for e in self.data: if not isinstance(e, str): return False return True def get_single_text(self): return "".join(self.data) def split_text_blocks(excel_func_text): """ 将excel公式字符串,按照一定的规则切割成数组 :param excel_func_text: 被切割的excel公式字符串 :return: 切割后的结果 """ excel_func_text = excel_func_text.replace('\n', '').replace('\r', '') excel_func_text = re.sub(" +", " ", excel_func_text) lines = [] i, j = 0, 0 while j < len(excel_func_text): c = excel_func_text[j] if (c == '(' and excel_func_text[j + 1] != ')') or c == ',': lines.append(excel_func_text[i:j + 1]) i = j = j + 1 elif c == ')' and excel_func_text[j - 1] != '(': if i < j: lines.append(excel_func_text[i:j]) i = j # 起始文件块置于)处 # 以下代码查找,如果中间不包含(或),则将)和,之间的文本块加入到划分结果 k = excel_func_text.find(",", j + 1) l = excel_func_text.find("(", j + 1, k) m = excel_func_text.find(")", j + 1, k) if k != -1 and l == -1 and m == -1: lines.append(excel_func_text[i:k + 1]) i = j = k + 1 elif j + 1 < len(excel_func_text) and excel_func_text[j + 1] != ')': lines.append(")") lines.append(excel_func_text[j + 1]) i = j = j + 2 else: lines.append(")") i = j = j + 1 elif c == '"': j = excel_func_text.find('"', j + 1) + 1 else: j += 1 return lines blank_char_count = 2 def combine_node(root, text_max_length=60, max_combine_layer=3): """ 合并最内层的只有纯文本子节点的节点为单个文本节点 :param root: 被合并的节点 :param text_max_length: 合并后的文本长度不超过该参数,则应用该合并替换原节点 :param max_combine_layer: 最大合并层数 :return: """ for _ in range(max_combine_layer): no_change = True stack = deque([root]) while stack: node = stack.pop() tmp = {} for i, e in enumerate(node.data): if isinstance(e, Node): if e.is_single_node(): single_text = e.get_single_text() if len(single_text) < text_max_length: tmp[i] = single_text else: stack.append(e) for i, e in tmp.items(): node.data[i] = e if len(tmp) != 0: no_change = False if no_change: break def node_next_line(node): for i, e in enumerate(node.data): if isinstance(e, str): if i == 0 or i == len(node.data) - 1: tab = node.tab_size - 1 else: tab = node.tab_size yield f"{' ' * blank_char_count * tab}{e}" else: yield from node_next_line(e) def excel_func_format(excel_func_text, blank_count=2, combine_single_node=True, text_max_length=60, max_combine_layer=3): """ 将excel公式格式化成比较容易阅读的格式 :param excel_func_text: 被格式化的excel公式字符串 :param blank_count: 最终显示的格式化字符串的1个tab用几个空格表示 :param combine_single_node: 是否合并纯文本节点,该参数设置为True后面的参数才生效 :param text_max_length: 合并后的文本长度不超过该参数,则应用该合并替换原节点 :param max_combine_layer: 最大合并层数 :return: 格式化后的字符串 """ global blank_char_count blank_char_count = blank_count blocks = split_text_blocks(excel_func_text) # print("\n".join(blocks)) # print('-----------拆分结果-----------') tab_size = 0 node = root = Node() for block in blocks: if block.endswith("("): tab_size += 1 child_node = Node(node, tab_size) node.data.append(child_node) node = child_node node.data.append(block) elif block.startswith(")"): tab_size -= 1 node.data.append(block) node = node.parent else: node.data.append(block) if combine_single_node: combine_node(root, text_max_length, max_combine_layer) result = [line for line in node_next_line(root)] return "\n".join(result)
处理流程浅析
下面都以如下公式作为示例:
=IF(ROW()>COLUMN(),"",IF(ROW()=COLUMN(),$B15,ROUNDDOWN($B15*INDIRECT(SUBSTITUTE(ADDRESS(1,3+COLUMN()-ROW(), 4),1,"")&56),0)))
文本分块切分
def split_text_blocks(excel_func_text): """ 将excel公式字符串,按照一定的规则切割成数组 :param excel_func_text: 被切割的excel公式字符串 :return: 切割后的结果 """ excel_func_text = excel_func_text.replace('\n', '').replace('\r', '') excel_func_text = re.sub(" +", " ", excel_func_text) lines = [] i, j = 0, 0 while j < len(excel_func_text): c = excel_func_text[j] if (c == '(' and excel_func_text[j + 1] != ')') or c == ',': lines.append(excel_func_text[i:j + 1]) i = j = j + 1 elif c == ')' and excel_func_text[j - 1] != '(': if i < j: lines.append(excel_func_text[i:j]) i = j # 起始文件块置于)处 # 以下代码查找,如果中间不包含(或),则将)和,之间的文本块加入到划分结果 k = excel_func_text.find(",", j + 1) l = excel_func_text.find("(", j + 1, k) m = excel_func_text.find(")", j + 1, k) if k != -1 and l == -1 and m == -1: lines.append(excel_func_text[i:k + 1]) i = j = k + 1 elif j + 1 < len(excel_func_text) and excel_func_text[j + 1] != ')': lines.append(")") lines.append(excel_func_text[j + 1]) i = j = j + 2 else: lines.append(")") i = j = j + 1 elif c == '"': j = excel_func_text.find('"', j + 1) + 1 else: j += 1 return lines s = """=IF(ROW()>COLUMN(),"",IF(ROW()=COLUMN(),$B15,ROUNDDOWN($B15*INDIRECT(SUBSTITUTE(ADDRESS(1,3+COLUMN()-ROW(), 4),1,"")&56),0))) """ blocks = split_text_blocks(s) for block in blocks: print(block)
的运行结果为:
=IF( ROW()>COLUMN(), "", IF( ROW()=COLUMN(), $B15, ROUNDDOWN( $B15*INDIRECT( SUBSTITUTE( ADDRESS( 1, 3+COLUMN()-ROW(), 4 ), 1, "" ) & 56 ), 0 ) ) )
这端代码首先替换掉所有的换行符,将多个空格替换为单个空格,然后将左右括号和逗号作为切分点进行切分。
但存在一些特殊情况,例如ROW()和COLUMN()括号内部没有任何内容,所有这种括号应该作为普通字符处理,另外被""包含的字符串可能包含括号,也应该作为普通字符。
构建多叉树层次结构
设计数据结构:
class Node: def __init__(self, parent=None, tab_size=0): self.parent = parent self.tab_size = tab_size self.data = []
parent存储父节点的指针,tab_size存储当前节点的层级,data存储当前节点的所有数据。
构建代码:
tab_size = 0 node = root = Node() for block in blocks: if block.endswith("("): tab_size += 1 child_node = Node(node, tab_size) node.data.append(child_node) node = child_node node.data.append(block) elif block.startswith(")"): tab_size -= 1 node.data.append(block) node = node.parent else: node.data.append(block)
构建完毕后,这段数据在内存中的结构(仅展示data)如下:
遍历打印这颗多叉树
def node_next_line(node): for i, e in enumerate(node.data): if isinstance(e, str): if i == 0 or i == len(node.data) - 1: tab = node.tab_size - 1 else: tab = node.tab_size yield f"{' ' * 2 * tab}{e}" else: yield from node_next_line(e) result = [line for line in node_next_line(root)] print("\n".join(result))
结果:
=IF( ROW()>COLUMN(), "", IF( ROW()=COLUMN(), $B15, ROUNDDOWN( $B15*INDIRECT( SUBSTITUTE( ADDRESS( 1, 3+COLUMN()-ROW(), 4 ), 1, "" ) & 56 ), 0 ) ) )
合并最内层的节点
显然将最内层的node5节点合并一下阅读性更好:
首先给数据结构增加判断是否为纯文本节点的方法:
class Node: def __init__(self, parent=None, tab_size=0): self.parent = parent self.tab_size = tab_size self.data = [] def is_single_node(self): for e in self.data: if not isinstance(e, str): return False return True def get_single_text(self): return "".join(self.data)
下面是合并纯文本节点的实现,max_combine_layer决定了合并的最大次数,如果合并后长度超过text_max_length参数,则不应用这次合并:
from collections import deque def combine_node(root, text_max_length=60, max_combine_layer=3): """ 合并最内层的只有纯文本子节点的节点为单个文本节点 :param root: 被合并的节点 :param text_max_length: 合并后的文本长度不超过该参数,则应用该合并替换原节点 :param max_combine_layer: 最大合并层数 :return: """ for _ in range(max_combine_layer): no_change = True stack = deque([root]) while stack: node = stack.pop() tmp = {} for i, e in enumerate(node.data): if isinstance(e, Node): if e.is_single_node(): single_text = e.get_single_text() if len(single_text) < text_max_length: tmp[i] = single_text else: stack.append(e) for i, e in tmp.items(): node.data[i] = e if len(tmp) != 0: no_change = False if no_change: break
合并一次:
combine_node(root, 100, 1) result = [line for line in node_next_line(root)] print("\n".join(result))
结果:
=IF( ROW()>COLUMN(), "", IF( ROW()=COLUMN(), $B15, ROUNDDOWN( $B15*INDIRECT( SUBSTITUTE( ADDRESS(1,3+COLUMN()-ROW(), 4), 1, "" ) & 56 ), 0 ) ) )
合并二次:
combine_node(root, 100, 2) result = [line for line in node_next_line(root)] print("\n".join(result))
结果:
=IF( ROW()>COLUMN(), "", IF( ROW()=COLUMN(), $B15, ROUNDDOWN( $B15*INDIRECT( SUBSTITUTE(ADDRESS(1,3+COLUMN()-ROW(), 4),1,"") & 56 ), 0 ) ) )
合并三次:
combine_node(root, 100, 3) result = [line for line in node_next_line(root)] print("\n".join(result))
结果:
=IF( ROW()>COLUMN(), "", IF( ROW()=COLUMN(), $B15, ROUNDDOWN( $B15*INDIRECT(SUBSTITUTE(ADDRESS(1,3+COLUMN()-ROW(), 4),1,"")&56), 0 ) ) )
合并三次后的内存情况:
体验网址
http://xiaoxiaoming.xyz:8088/excel
不保证永久有效。
免责声明:本站文章均来自网站采集或用户投稿,网站不提供任何软件下载或自行开发的软件!
如有用户或公司发现本站内容信息存在侵权行为,请邮件告知! 858582#qq.com
内蒙古资源网 Copyright www.nmgbbs.com
暂无“python实现excel公式格式化的示例代码”评论...