十大经典排序算法(python实现)(原创)

Posted 万雨

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了十大经典排序算法(python实现)(原创)相关的知识,希望对你有一定的参考价值。

使用场景:

1,空间复杂度 越低越好、n值较大:

  堆排序  O(nlog2n)  O(1)

2,无空间复杂度要求、n值较大:

  桶排序  O(n+k)    O(n+k)

 

 

 

经典排序算法图解:

经典排序算法的复杂度:

 

大类一(比较排序法):

1、冒泡排序(Bubble Sort)【前后比较-交换】

 

python代码实现:

 1 d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]
 2 d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
 3 
 4 while 1:
 5     state = 0  # 假设本次循环没有改变
 6     for i in range(len(d0) - 1):
 7         if d0[i] > d0[i + 1]:
 8             d0[i], d0[i + 1] = d0[i + 1], d0[i]
 9             state = 1  # 有数值交换,那么状态值置1
10     if not state:  # 如果没有数值交换,那么就跳出
11         break
12 
13 print(d0)
14 print(d0_out)

 

2、选择排序(Selection Sort)【选择最小的数据放在前面】

python代码实现:

 1 def select_sort(data):
 2     d1 = []
 3     while len(data):
 4         min = [0, data[0]]
 5         for i in range(len(data)):
 6             if min[1] > data[i]:
 7                 min = [i, data[i]]
 8         del data[min[0]]  # 找到剩余部分的最小值,并且从原数组中删除
 9         d1.append(min[1])  # 在新数组中添加
10     return d1
11 
12 if __name__ == "__main__":
13     d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]  # 原始乱序
14     d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
15     d1 = select_sort(d0)
16     print(d1)
17     print(d0_out)

3、插入排序(Insertion Sort)【逐个插入到前面的有序数中】

 

直接插入排序-python实现:

 1 def direct_insertion_sort(d):   # 直接插入排序,因为要用到后面的希尔排序,所以转成function
 2     d1 = [d[0]]
 3     for i in d[1:]:
 4         state = 1
 5         for j in range(len(d1) - 1, -1, -1):
 6             if i >= d1[j]:
 7                 d1.insert(j + 1, i)  # 将元素插入数组
 8                 state = 0
 9                 break
10         if state:
11             d1.insert(0, i)
12     return d1
13 
14 
15 if __name__ == "__main__":
16     d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]  # 原始乱序
17     d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
18     d1 = direct_insertion_sort(d0)
19     print(d1)
20     print(d0_out)

 

折半插入排序-python实现:

 1 d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]  # 原始乱序
 2 d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
 3 
 4 d1 = [d0[0]]
 5 del d0[0]
 6 
 7 for i in d0:
 8     index_now = [0, len(d1)]
 9     while 1:
10         index = index_now[0] + int((index_now[1] - index_now[0]) / 2)
11         if i == d1[index]:
12             d1.insert(index+1,i)
13             break
14         elif index in index_now:  # 如果更新的index值在index_now中存在(也有可能是边界),那么就表明无法继续更新
15             d1.insert(index+1,i)
16             break
17         elif i > d1[index]:
18             index_now[0] = index
19         elif i < d1[index]:
20             index_now[1] = index
21 
22 print(d1)
23 print(d0_out)

 

4、希尔排序(Shell Sort)【从大范围到小范围进行比较-交换】类似冒泡和插入的联合

 

python代码实现:

 1 def direct_insertion_sort(d):  # 直接插入排序,因为要用到后面的希尔排序,所以转成function
 2     d1 = [d[0]]
 3     for i in d[1:]:
 4         state = 1
 5         for j in range(len(d1) - 1, -1, -1):
 6             if i >= d1[j]:
 7                 d1.insert(j + 1, i)  # 将元素插入数组
 8                 state = 0
 9                 break
10         if state:
11             d1.insert(0, i)
12     return d1
13 
14 
15 def shell_sort(d):  # d 为乱序数组,l为初始增量,其中l<len(d),取为len(d)/2比较好操作。最后还是直接省略length输入
16     length = int(len(d) / 2)  # 10
17     num = int(len(d) / length)  # 2
18     while 1:
19 
20         for i in range(length):
21             d_mid = []
22             for j in range(num):
23                 d_mid.append(d[i + j * length])
24             d_mid = direct_insertion_sort(d_mid)
25             for j in range(num):
26                 d[i + j * length] = d_mid[j]
27         # print(d)
28         length = int(length / 2)
29         if length == 0:
30             return d
31             break
32         # print(\'length:\',length)
33         num = int(len(d) / length)
34 
35 
36 if __name__ == "__main__":
37     d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]  # 原始乱序
38     d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
39     d1 = shell_sort(d0)
40     print(d1)
41     print(d0_out)

 

5、归并排序(Merge Sort)【分治法-2-4-8插入排序】

python代码实现(这个地方是由大往小进行递归):

 1 # 归并排序,还有些问题。其中有些细节需要重新理解
 2 # 也是递归问题
 3 def merge_sort(data):  # 分治发的典型应用,大问题拆分成小问题,逐个击破,之后将结果合并
 4     half_index = int(len(data) / 2)  # 将数组拆分
 5 
 6     d0 = data[:half_index]
 7     d1 = data[half_index:]
 8 
 9     if len(d0) > 1:
10         d0 = merge_sort(d0)
11 
12     if len(d1) > 1:
13         d1 = merge_sort(d1)
14 
15     index = 0
16     for i in range(len(d1)):
17         state = 1
18         for j in range(index, len(d0)):
19             if d1[i] < d0[j]:
20                 state = 0
21                 index = j + 1
22                 d0.insert(j, d1[i])
23                 break
24         if state == 1:  # 如果大于d0这个队列的所有值,那么直接extend所有数据
25             d0.extend(d1[i:])
26             break
27     return d0
28 
29 
30 if __name__ == "__main__":
31     d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]  # 原始乱序
32     d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
33     d1 = merge_sort(d0)
34     print(d1)
35     print(d0_out)

 

python代码实现(由小扩展到大的子序列):
 1 
 2 def list_sort(d0, d1):  # 基元组往大扩展
 3     index = 0
 4     for i in range(len(d1)):  # 遍历d1数组
 5         state = 1
 6         for j in range(index, len(d0)):  # 遍历d0数组
 7             if d0[j] > d1[i]:
 8                 state = 0
 9                 index = j + 1
10                 d0.insert(j, d1[i])
11                 break
12         if state == 1:  # 如果大于d0这个队列的所有值,那么直接extend所有数据
13             d0.extend(d1[i:])
14             break
15     return d0
16 
17 
18 def merge_sort(data):
19     d0 = [[x] for x in data]
20     while len(d0) != 1:    # 循环条件
21         length = len(d0)
22         half = int(length/2)    # 除2的整数部分
23         quo = length%2          # 除2的商
24         d0_mid = []
25         for i in range(half):
26             d0_mid.append(list_sort(d0[i*2], d0[i*2+1]))
27         if quo:
28             d0_mid.append(d0[-1])
29         d0 = d0_mid
30 
31     return d0[0]
32 
33 
34 
35 if __name__ == "__main__":
36     d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]  # 原始乱序
37     d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
38     d1 = merge_sort(d0)
39     print(d1)
40     print(d0_out)

 

 

6、快速排序(Quick Sort)【选取一个基准值,小数在左大数在在右】

python代码实现:

 1 # import sys
 2 # sys.setrecursionlimit(1000000)
 3 
 4 def quick_sort(data):
 5     d = [[], [], []]
 6     d_pivot = data[-1]  # 因为是乱序数组,所以第几个都是可以的,理论上是一样的
 7     for i in data:
 8         if i < d_pivot:  # 小于基准值的放在前
 9             d[0].append(i)
10         elif i > d_pivot:  # 大于基准值的放在后
11             d[2].append(i)
12         else:  # 等于基准值的放在中间
13             d[1].append(i)
14 
15     # print(d[0], d[1], d[2])
16     if len(d[0]) > 1:  # 大于基准值的子数组,递归
17         d[0] = quick_sort(d[0])
18 
19     if len(d[2]) > 1:  # 小于基准值的子数组,递归
20         d[2] = quick_sort(d[2])
21 
22     d[0].extend(d[1])
23     d[0].extend(d[2])
24     return d[0]
25 
26 
27 if __name__ == "__main__":
28     d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]  # 原始乱序
29     d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]  # 正确排序
30     d1 = quick_sort(d0)
31     print(d1)
32     print(d0_out)

 

 

7、堆排序(Heap Sort)【利用最大堆和最小堆的特性】

python代码实现:

 1 d0 = [99, 5, 36, 7, 22, 17, 46, 12, 2, 19, 25, 28, 1, 92]
 2 
 3 
 4 def sort_max(data):  # 直接冒泡一下吧,小到大
 5     for i in range(len(data) - 1):
 6         for j in range(len(data) - 1):
 7             if data[j] > data[j + 1]:
 8                 data[j], data[j + 1] = data[j + 1], data[j]
 9     return data
10 
11 def heap_min(data,type):
12     index = 0
13     if not type:
14         for i in range(len(data[1:])):
15             if data[index] > data[i+1]:
16                 index = i+1
17         data[0],data[index] = data[index],data[0]
18         return data
19     else:
20         for i in range(len(data[1:])):
21             if data[index] < data[i+1]:
22                 index = i+1
23         data[0],data[index] = data[index],data[0]
24         return data
25 
26 
27 # d0 = [3,2,1,10,3]
28 # print(heap_min(d0,1))
29 # print(heap_min(d0,0))
30 
31 import numpy as np
32 
33 
34 def heap_adj(data, type):  # data 原始堆,type=1最大堆,type=0最小堆
35     length = len(data)
36     floor = int(np.log2(length))
37     for i in range(floor, 0, -1):  # 3(7 6 5 4)-2(3 2)-1(1)
38         for j in range(2 ** floor - 1, 2 ** (floor - i) - 1, -1):
39             # print(i,j)    # j-1 为当前父节点
40             d_mid = [data[j - 1]]  # j = 7,j-1 =6 index
41             if j * 2 <= length:  # 14
42                 d_mid.append(data[j * 2 - 1])
43             if j * 2 + 1 <= length:
44                 d_mid.append(data[j * 2])
45 
46             d_mid = heap_min(d_mid, type)
47 
48             if len(d_mid) == 2:
49                 data[j - 1], data[j * 2 - 1] = d_mid[0], d_mid[1]
50             elif len(d_mid) == 3:
51                 data[j - 1], data[j * 2 - 1], data[j * 2] = d_mid[0], d_mid[1], d_mid[2]
52     return data
53 
54 d1 = []
55 for i in range(len(d0)):
56     data = heap_adj(d0, 0)
57     d1.append(d0[0])
58     del d0[0]
59 
60 
61 print(d1)

 

大类二(非比较排序法):

8、计数排序(Counting Sort)【字典计数-还原】

python代码实现:

 1 d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]
 2 d0_out = [2, 2, 2, 3, 4, 4, 4, 4, 5, 5, 5, 6, 6, 7, 9, 12, 15, 44, 45, 54, 64]
 3 
 4 d_max = 0
 5 d_min = 0
 6 for i in d0:
 7     if d_max<i:
 8         d_max = i
 9     if d_min>i:
10         d_min = i
11 
12 d1 = {}
13 for i in d0:
14     if i in d1.keys():
15         d1[i] += 1
16     else:
17         d1[i] = 1
18 
19 d2 = []
20 for i in range(d_min,d_max+1):
21     if i in d1.keys():
22         for j in range(d1[i]):
23             d2.append(i)
24 
25 print(d2)

 

9、桶排序(Bucket Sort)【链表】

python代码实现:

 1 d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]
 2 
 3 d1 = [[] for x in range(10)]
 4 for i in d0:
 5     d1[int(i/10)].append(i)
 6 
 7 # print(d1)
 8 
 9 
10 for i in range(len(d1)):
11     if d1[i] != []:
12         d2 = [[] for x in range(10)]
13         for j in d1[i]:
14             d2[j%10].append(j)
15         d1[i] = d2
16 
17 # print(d1)
18 
19 d3 = []
20 for i in d1:
21     if i:
22         for j in i:
23             if j:
24                 for k in j:
25                     if k:
26                         d3.append(k)
27 print(d3)

 

10、基数排序(Radix Sort)

 

python代码实现:

 1 d0 = [2, 15, 5, 9, 7, 6, 4, 12, 5, 4, 2, 64, 5, 6, 4, 2, 3, 54, 45, 4, 44]
 2 
 3 d1 = [[] for x in range(10)]
 4 
 5 # 第一次 最小位次排序
 6 for i in d0:
 7     d1[i % 10].append(i)
 8 
 9 print(d1)
10 
11 d0_1 = []
12 for i in d1:
13     if i:
14         for j in i:
15             d0_1.append(j)
16 print(d0_1)
17 
18 # 第二次 次低位排序
19 d2 = [[] for x in range(10)]
20 for i in d0_1:
21     d2[int(i/10)].append(i)
22 print(d2)
23 
24 d0_2 = []
25 for i in d2:
26     if i:
27         for j in i:
28             d0_2.append(j)
29 print(d0_2)

 

文末再次感谢:博主 一像素 https://www.cnblogs.com/onepixel/

参考文档出处:

十大经典排序算法(动图演示) https://www.cnblogs.com/onepixel/articles/7674659.html

分治算法 https://www.cnblogs.com/chentingk/p/5671808.html

堆 https://www.cnblogs.com/chenweichu/articles/5710567.html

 

以上是关于十大经典排序算法(python实现)(原创)的主要内容,如果未能解决你的问题,请参考以下文章

python实现十大经典排序算法

(附代码)动图图解 | 十大经典排序算法Python版实现

面试官常问十大经典算法排序(用Python实现)

动画详解十大经典排序算法Java版实现(上)

动画详解十大经典排序算法Java版实现(上)

十大经典排序(下)——Python3实现