似乎没有函数可以简单地计算numpy/scipy的移动平均值，这导致了复杂的解决方案。
我的问题有两个方面:
用numpy(正确地)实现移动平均的最简单方法是什么? 既然这似乎不是小事，而且容易出错，有没有一个很好的理由不包括电池在这种情况下?
当前回答
这里有许多实现这一点的方法，以及一些基准测试。最好的方法是使用来自其他库的优化代码。瓶颈。Move_mean方法可能是最好的方法。scipy。卷积方法也非常快，可扩展，并且语法和概念简单，但是对于非常大的窗口值不能很好地扩展。numpy。如果你需要一个纯numpy方法，Cumsum方法是很好的。
注意:其中一些(例如:瓶颈。move_mean)不是居中的，并且会转移你的数据。
import numpy as np
import scipy as sci
import scipy.signal as sig
import pandas as pd
import bottleneck as bn
import time as time
def rollavg_direct(a,n):
'Direct "for" loop'
assert n%2==1
b = a*0.0
for i in range(len(a)) :
b[i]=a[max(i-n//2,0):min(i+n//2+1,len(a))].mean()
return b
def rollavg_comprehension(a,n):
'List comprehension'
assert n%2==1
r,N = int(n/2),len(a)
return np.array([a[max(i-r,0):min(i+r+1,N)].mean() for i in range(N)])
def rollavg_convolve(a,n):
'scipy.convolve'
assert n%2==1
return sci.convolve(a,np.ones(n,dtype='float')/n, 'same')[n//2:-n//2+1]
def rollavg_convolve_edges(a,n):
'scipy.convolve, edge handling'
assert n%2==1
return sci.convolve(a,np.ones(n,dtype='float'), 'same')/sci.convolve(np.ones(len(a)),np.ones(n), 'same')
def rollavg_cumsum(a,n):
'numpy.cumsum'
assert n%2==1
cumsum_vec = np.cumsum(np.insert(a, 0, 0))
return (cumsum_vec[n:] - cumsum_vec[:-n]) / n
def rollavg_cumsum_edges(a,n):
'numpy.cumsum, edge handling'
assert n%2==1
N = len(a)
cumsum_vec = np.cumsum(np.insert(np.pad(a,(n-1,n-1),'constant'), 0, 0))
d = np.hstack((np.arange(n//2+1,n),np.ones(N-n)*n,np.arange(n,n//2,-1)))
return (cumsum_vec[n+n//2:-n//2+1] - cumsum_vec[n//2:-n-n//2]) / d
def rollavg_roll(a,n):
'Numpy array rolling'
assert n%2==1
N = len(a)
rolling_idx = np.mod((N-1)*np.arange(n)[:,None] + np.arange(N), N)
return a[rolling_idx].mean(axis=0)[n-1:]
def rollavg_roll_edges(a,n):
# see https://stackoverflow.com/questions/42101082/fast-numpy-roll
'Numpy array rolling, edge handling'
assert n%2==1
a = np.pad(a,(0,n-1-n//2), 'constant')*np.ones(n)[:,None]
m = a.shape[1]
idx = np.mod((m-1)*np.arange(n)[:,None] + np.arange(m), m) # Rolling index
out = a[np.arange(-n//2,n//2)[:,None], idx]
d = np.hstack((np.arange(1,n),np.ones(m-2*n+1+n//2)*n,np.arange(n,n//2,-1)))
return (out.sum(axis=0)/d)[n//2:]
def rollavg_pandas(a,n):
'Pandas rolling average'
return pd.DataFrame(a).rolling(n, center=True, min_periods=1).mean().to_numpy()
def rollavg_bottlneck(a,n):
'bottleneck.move_mean'
return bn.move_mean(a, window=n, min_count=1)
N = 10**6
a = np.random.rand(N)
functions = [rollavg_direct, rollavg_comprehension, rollavg_convolve,
rollavg_convolve_edges, rollavg_cumsum, rollavg_cumsum_edges,
rollavg_pandas, rollavg_bottlneck, rollavg_roll, rollavg_roll_edges]
print('Small window (n=3)')
%load_ext memory_profiler
for f in functions :
print('\n'+f.__doc__+ ' : ')
%timeit b=f(a,3)
print('\nLarge window (n=1001)')
for f in functions[0:-2] :
print('\n'+f.__doc__+ ' : ')
%timeit b=f(a,1001)
print('\nMemory\n')
print('Small window (n=3)')
N = 10**7
a = np.random.rand(N)
%load_ext memory_profiler
for f in functions[2:] :
print('\n'+f.__doc__+ ' : ')
%memit b=f(a,3)
print('\nLarge window (n=1001)')
for f in functions[2:-2] :
print('\n'+f.__doc__+ ' : ')
%memit b=f(a,1001)
定时，小窗口(n=3)
Direct "for" loop :
4.14 s ± 23.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
List comprehension :
3.96 s ± 27.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
scipy.convolve :
1.07 ms ± 26.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
scipy.convolve, edge handling :
4.68 ms ± 9.69 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
numpy.cumsum :
5.31 ms ± 5.11 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
numpy.cumsum, edge handling :
8.52 ms ± 11.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
Pandas rolling average :
9.85 ms ± 9.63 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
bottleneck.move_mean :
1.3 ms ± 12.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
Numpy array rolling :
31.3 ms ± 91.9 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
Numpy array rolling, edge handling :
61.1 ms ± 55.9 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
定时，大窗口(n=1001)
Direct "for" loop :
4.67 s ± 34 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
List comprehension :
4.46 s ± 14.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
scipy.convolve :
103 ms ± 165 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
scipy.convolve, edge handling :
272 ms ± 1.23 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
numpy.cumsum :
5.19 ms ± 12.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
numpy.cumsum, edge handling :
8.7 ms ± 11.5 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
Pandas rolling average :
9.67 ms ± 199 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
bottleneck.move_mean :
1.31 ms ± 15.7 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
内存，小窗口(n=3)
The memory_profiler extension is already loaded. To reload it, use:
%reload_ext memory_profiler
scipy.convolve :
peak memory: 362.66 MiB, increment: 73.61 MiB
scipy.convolve, edge handling :
peak memory: 510.24 MiB, increment: 221.19 MiB
numpy.cumsum :
peak memory: 441.81 MiB, increment: 152.76 MiB
numpy.cumsum, edge handling :
peak memory: 518.14 MiB, increment: 228.84 MiB
Pandas rolling average :
peak memory: 449.34 MiB, increment: 160.02 MiB
bottleneck.move_mean :
peak memory: 374.17 MiB, increment: 75.54 MiB
Numpy array rolling :
peak memory: 661.29 MiB, increment: 362.65 MiB
Numpy array rolling, edge handling :
peak memory: 1111.25 MiB, increment: 812.61 MiB
内存，大窗口(n=1001)
scipy.convolve :
peak memory: 370.62 MiB, increment: 71.83 MiB
scipy.convolve, edge handling :
peak memory: 521.98 MiB, increment: 223.18 MiB
numpy.cumsum :
peak memory: 451.32 MiB, increment: 152.52 MiB
numpy.cumsum, edge handling :
peak memory: 527.51 MiB, increment: 228.71 MiB
Pandas rolling average :
peak memory: 451.25 MiB, increment: 152.50 MiB
bottleneck.move_mean :
peak memory: 374.64 MiB, increment: 75.85 MiB
其他回答
下面是一个使用numba的快速实现(注意类型)。注意它确实包含移位的nan。
import numpy as np
import numba as nb
@nb.jit(nb.float64[:](nb.float64[:],nb.int64),
fastmath=True,nopython=True)
def moving_average( array, window ):
ret = np.cumsum(array)
ret[window:] = ret[window:] - ret[:-window]
ma = ret[window - 1:] / window
n = np.empty(window-1); n.fill(np.nan)
return np.concatenate((n.ravel(), ma.ravel()))
这里有许多实现这一点的方法，以及一些基准测试。最好的方法是使用来自其他库的优化代码。瓶颈。Move_mean方法可能是最好的方法。scipy。卷积方法也非常快，可扩展，并且语法和概念简单，但是对于非常大的窗口值不能很好地扩展。numpy。如果你需要一个纯numpy方法，Cumsum方法是很好的。
注意:其中一些(例如:瓶颈。move_mean)不是居中的，并且会转移你的数据。
import numpy as np
import scipy as sci
import scipy.signal as sig
import pandas as pd
import bottleneck as bn
import time as time
def rollavg_direct(a,n):
'Direct "for" loop'
assert n%2==1
b = a*0.0
for i in range(len(a)) :
b[i]=a[max(i-n//2,0):min(i+n//2+1,len(a))].mean()
return b
def rollavg_comprehension(a,n):
'List comprehension'
assert n%2==1
r,N = int(n/2),len(a)
return np.array([a[max(i-r,0):min(i+r+1,N)].mean() for i in range(N)])
def rollavg_convolve(a,n):
'scipy.convolve'
assert n%2==1
return sci.convolve(a,np.ones(n,dtype='float')/n, 'same')[n//2:-n//2+1]
def rollavg_convolve_edges(a,n):
'scipy.convolve, edge handling'
assert n%2==1
return sci.convolve(a,np.ones(n,dtype='float'), 'same')/sci.convolve(np.ones(len(a)),np.ones(n), 'same')
def rollavg_cumsum(a,n):
'numpy.cumsum'
assert n%2==1
cumsum_vec = np.cumsum(np.insert(a, 0, 0))
return (cumsum_vec[n:] - cumsum_vec[:-n]) / n
def rollavg_cumsum_edges(a,n):
'numpy.cumsum, edge handling'
assert n%2==1
N = len(a)
cumsum_vec = np.cumsum(np.insert(np.pad(a,(n-1,n-1),'constant'), 0, 0))
d = np.hstack((np.arange(n//2+1,n),np.ones(N-n)*n,np.arange(n,n//2,-1)))
return (cumsum_vec[n+n//2:-n//2+1] - cumsum_vec[n//2:-n-n//2]) / d
def rollavg_roll(a,n):
'Numpy array rolling'
assert n%2==1
N = len(a)
rolling_idx = np.mod((N-1)*np.arange(n)[:,None] + np.arange(N), N)
return a[rolling_idx].mean(axis=0)[n-1:]
def rollavg_roll_edges(a,n):
# see https://stackoverflow.com/questions/42101082/fast-numpy-roll
'Numpy array rolling, edge handling'
assert n%2==1
a = np.pad(a,(0,n-1-n//2), 'constant')*np.ones(n)[:,None]
m = a.shape[1]
idx = np.mod((m-1)*np.arange(n)[:,None] + np.arange(m), m) # Rolling index
out = a[np.arange(-n//2,n//2)[:,None], idx]
d = np.hstack((np.arange(1,n),np.ones(m-2*n+1+n//2)*n,np.arange(n,n//2,-1)))
return (out.sum(axis=0)/d)[n//2:]
def rollavg_pandas(a,n):
'Pandas rolling average'
return pd.DataFrame(a).rolling(n, center=True, min_periods=1).mean().to_numpy()
def rollavg_bottlneck(a,n):
'bottleneck.move_mean'
return bn.move_mean(a, window=n, min_count=1)
N = 10**6
a = np.random.rand(N)
functions = [rollavg_direct, rollavg_comprehension, rollavg_convolve,
rollavg_convolve_edges, rollavg_cumsum, rollavg_cumsum_edges,
rollavg_pandas, rollavg_bottlneck, rollavg_roll, rollavg_roll_edges]
print('Small window (n=3)')
%load_ext memory_profiler
for f in functions :
print('\n'+f.__doc__+ ' : ')
%timeit b=f(a,3)
print('\nLarge window (n=1001)')
for f in functions[0:-2] :
print('\n'+f.__doc__+ ' : ')
%timeit b=f(a,1001)
print('\nMemory\n')
print('Small window (n=3)')
N = 10**7
a = np.random.rand(N)
%load_ext memory_profiler
for f in functions[2:] :
print('\n'+f.__doc__+ ' : ')
%memit b=f(a,3)
print('\nLarge window (n=1001)')
for f in functions[2:-2] :
print('\n'+f.__doc__+ ' : ')
%memit b=f(a,1001)
定时，小窗口(n=3)
Direct "for" loop :
4.14 s ± 23.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
List comprehension :
3.96 s ± 27.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
scipy.convolve :
1.07 ms ± 26.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
scipy.convolve, edge handling :
4.68 ms ± 9.69 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
numpy.cumsum :
5.31 ms ± 5.11 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
numpy.cumsum, edge handling :
8.52 ms ± 11.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
Pandas rolling average :
9.85 ms ± 9.63 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
bottleneck.move_mean :
1.3 ms ± 12.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
Numpy array rolling :
31.3 ms ± 91.9 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
Numpy array rolling, edge handling :
61.1 ms ± 55.9 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
定时，大窗口(n=1001)
Direct "for" loop :
4.67 s ± 34 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
List comprehension :
4.46 s ± 14.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
scipy.convolve :
103 ms ± 165 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
scipy.convolve, edge handling :
272 ms ± 1.23 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
numpy.cumsum :
5.19 ms ± 12.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
numpy.cumsum, edge handling :
8.7 ms ± 11.5 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
Pandas rolling average :
9.67 ms ± 199 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
bottleneck.move_mean :
1.31 ms ± 15.7 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
内存，小窗口(n=3)
The memory_profiler extension is already loaded. To reload it, use:
%reload_ext memory_profiler
scipy.convolve :
peak memory: 362.66 MiB, increment: 73.61 MiB
scipy.convolve, edge handling :
peak memory: 510.24 MiB, increment: 221.19 MiB
numpy.cumsum :
peak memory: 441.81 MiB, increment: 152.76 MiB
numpy.cumsum, edge handling :
peak memory: 518.14 MiB, increment: 228.84 MiB
Pandas rolling average :
peak memory: 449.34 MiB, increment: 160.02 MiB
bottleneck.move_mean :
peak memory: 374.17 MiB, increment: 75.54 MiB
Numpy array rolling :
peak memory: 661.29 MiB, increment: 362.65 MiB
Numpy array rolling, edge handling :
peak memory: 1111.25 MiB, increment: 812.61 MiB
内存，大窗口(n=1001)
scipy.convolve :
peak memory: 370.62 MiB, increment: 71.83 MiB
scipy.convolve, edge handling :
peak memory: 521.98 MiB, increment: 223.18 MiB
numpy.cumsum :
peak memory: 451.32 MiB, increment: 152.52 MiB
numpy.cumsum, edge handling :
peak memory: 527.51 MiB, increment: 228.71 MiB
Pandas rolling average :
peak memory: 451.25 MiB, increment: 152.50 MiB
bottleneck.move_mean :
peak memory: 374.64 MiB, increment: 75.85 MiB
从Numpy 1.20开始，sliding_window_view提供了一种在元素窗口中滑动/滚动的方法。然后你可以分别取平均值。
例如，对于一个4元素的窗口:
from numpy.lib.stride_tricks import sliding_window_view
# values = np.array([5, 3, 8, 10, 2, 1, 5, 1, 0, 2])
np.average(sliding_window_view(values, window_shape = 4), axis=1)
# array([6.5, 5.75, 5.25, 4.5, 2.25, 1.75, 2])
注意sliding_window_view的中间结果:
# values = np.array([5, 3, 8, 10, 2, 1, 5, 1, 0, 2])
sliding_window_view(values, window_shape = 4)
# array([[ 5, 3, 8, 10],
# [ 3, 8, 10, 2],
# [ 8, 10, 2, 1],
# [10, 2, 1, 5],
# [ 2, 1, 5, 1],
# [ 1, 5, 1, 0],
# [ 5, 1, 0, 2]])
移动平均线 迭代器方法 在i处反转数组，简单地求i到n的均值。 使用列表推导式在运行中生成迷你数组。
x = np.random.randint(10, size=20)
def moving_average(arr, n):
return [ (arr[:i+1][::-1][:n]).mean() for i, ele in enumerate(arr) ]
d = 5
moving_average(x, d)
张量卷积
moving_average = np.convolve(x, np.ones(d)/d, mode='valid')
所有的答案似乎都集中在预先计算的列表的情况下。对于实际运行的用例，数字一个接一个地进来，这里有一个简单的类，它提供了对最后N个值求平均的服务:
import numpy as np
class RunningAverage():
def __init__(self, stack_size):
self.stack = [0 for _ in range(stack_size)]
self.ptr = 0
self.full_cycle = False
def add(self,value):
self.stack[self.ptr] = value
self.ptr += 1
if self.ptr == len(self.stack):
self.full_cycle = True
self.ptr = 0
def get_avg(self):
if self.full_cycle:
return np.mean(self.stack)
else:
return np.mean(self.stack[:self.ptr])
用法:
N = 50 # size of the averaging window
run_avg = RunningAverage(N)
for i in range(1000):
value = <my computation>
run_avg.add(value)
if i % 20 ==0: # print once in 20 iters:
print(f'the average value is {run_avg.get_avg()}')
推荐文章
- 如何读取文件的前N行?
- 如何删除matplotlib中的顶部和右侧轴?
- 解析.py文件，读取AST，修改它，然后写回修改后的源代码
- Visual Studio Code:如何调试Python脚本的参数
- 使用元组/列表等等。从输入vs直接引用类型如list/tuple/etc
- 结合conda环境。Yml和PIP requirements.txt
- 将命名元组转换为字典
- 如何使x轴和y轴的刻度相等呢?
- Numpy在这里函数多个条件
- 在Python中，使用argparse只允许正整数
- 如何排序mongodb与pymongo
- 不可变与可变类型
- 列表是线程安全的吗?
- 操作系统。makdirs在我的路径上不理解“~”
- 如何在Django模板中获得我的网站的域名?
