1D数组所有可能切片的平均值计算

huangapple go评论55阅读模式
英文:

Calculation of average of all possible slices of 1d-array

问题

我有一个测量数据的一维数组。我想要计算由起始索引和结束索引定义的每个可能切片的平均值。就像在我的数据的两端切割并计算每个可能的切割的平均值。

结果应该存储在一个二维数组中(实际上是一个三角形,因为起始索引必须小于结束索引)。

使用循环可以工作,但需要很长时间。

这是我的代码:

N = 5
data = np.arange(N)  # 例如
av = np.zeros((N, N))
for i in range(av.shape[0]):
    for j in range(av.shape[1]):
        av[j, i] = np.mean(data[i:j+1])

这个代码可以工作,但需要很长时间。对于类似的计算(元素之间的差异而不是切片的平均值),我找到了这个非常快的解决方案:

dist = np.subtract.outer(data, data)

但我没有找出如何使用切片的平均值来做到这一点。

英文:

I have a (big) 1d array of measurement data. I want to calculate the average for every possible slice defined by a start-index and a stop-index. Like cutting at both ends of my data and average every possible cut.
The result should be stored in a square 2D array (actually a triangle, as the start index must be smaller than the stop index).

Using loops works, but takes a long time.

Is there a way to speed this up?

I have this code:

N = 5
data = np.arange(N)  # example
av = np.zeros((N, N))
for i in range(av.shape[0]):
    for j in range(av.shape[1]):
        av[j, i] = np.mean(data[i:j+1])

This works, but takes a long time. For a similar calculation (differences of elements instead of averages of slices), I found this very fast solution:

dist = np.subtract.outer(data, data)

But I did not figure out how this could be done with averages of slices.

答案1

得分: 0

代码部分不翻译,只提供已翻译的文本部分。以下是您提供的内容的翻译:

# 通过求和后除以项数的一种选项:
a = np.arange(len(data))

av = (np.tril(np.repeat(data[:, None], len(data), axis=1)).cumsum(axis=0)
     / np.tril((a[:, None] - a + 1))
     )

输出:

array([[0. , nan, nan, nan, nan],
       [0.5, 1. , nan, nan, nan],
       [1. , 1.5, 2. , nan, nan],
       [1.5, 2. , 2.5, 3. , nan],
       [2. , 2.5, 3. , 3.5, 4. ]])

中间结果:

# 重复数据以获得方形形状并保留下三角形
np.tril(np.repeat(data[:, None], len(data), axis=1))
array([[0, 0, 0, 0, 0],
       [1, 1, 0, 0, 0],
       [2, 2, 2, 0, 0],
       [3, 3, 3, 3, 0],
       [4, 4, 4, 4, 4]])

# 获取累积和
[].cumsum()
array([[ 0,  0,  0,  0,  0],
       [ 1,  1,  0,  0,  0],
       [ 3,  3,  2,  0,  0],
       [ 6,  6,  5,  3,  0],
       [10, 10,  9,  7,  4]])

# 计算除数
a = np.arange(len(data))
array([0, 1, 2, 3, 4])

np.tril((a[:, None] - a + 1))
array([[1, 0, 0, 0, 0],
       [2, 1, 0, 0, 0],
       [3, 2, 1, 0, 0],
       [4, 3, 2, 1, 0],
       [5, 4, 3, 2, 1]])

性能

对于输入数据为1000的整数数组 (data = np.arange(1000)):

# 嵌套的for循环
6.28± 142 毫秒每次循环(平均值 ± 7 次运行的标准差,每次循环1次)

# 矢量化的numpy
12.3 毫秒 ± 427 微秒每次循环(平均值 ± 7 次运行的标准差,每次循环100次)
英文:

One option with a summation, then division by the number of items:

a = np.arange(len(data))

av = (np.tril(np.repeat(data[:,None], len(data), axis=1)).cumsum(axis=0)
     /np.tril((a[:,None]-a+1))
     )

Output:

array([[0. , nan, nan, nan, nan],
       [0.5, 1. , nan, nan, nan],
       [1. , 1.5, 2. , nan, nan],
       [1.5, 2. , 2.5, 3. , nan],
       [2. , 2.5, 3. , 3.5, 4. ]])

Intermediates:

# repeat data to square shape and keep lower triangle
np.tril(np.repeat(data[:,None], len(data), axis=1))
array([[0, 0, 0, 0, 0],
       [1, 1, 0, 0, 0],
       [2, 2, 2, 0, 0],
       [3, 3, 3, 3, 0],
       [4, 4, 4, 4, 4]])

# get the cumulated sum
[…].cumsum()
array([[ 0,  0,  0,  0,  0],
       [ 1,  1,  0,  0,  0],
       [ 3,  3,  2,  0,  0],
       [ 6,  6,  5,  3,  0],
       [10, 10,  9,  7,  4]])

# compute the divider
a = np.arange(len(data))
array([0, 1, 2, 3, 4])

np.tril((a[:,None]-a+1))
array([[1, 0, 0, 0, 0],
       [2, 1, 0, 0, 0],
       [3, 2, 1, 0, 0],
       [4, 3, 2, 1, 0],
       [5, 4, 3, 2, 1]])

performance

One a 1000 integer input (data = np.arange(1000))

# nested for loop
6.28 s ± 142 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

# vectorized numpy
12.3 ms ± 427 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

huangapple
  • 本文由 发表于 2023年3月3日 18:24:14
  • 转载请务必保留本文链接:https://go.coder-hub.com/75625859.html
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定