<p>所以，我知道这个问题已经问了很长时间了，因为我碰到了同样的问题，在处理长时间序列时，你真的希望避免对你不感兴趣的值进行不必要的计算。由于该方法不实现<code>step</code>参数，所以我使用numpy编写了一个变通方法</p>
<p>它基本上是<a href="https://stackoverflow.com/questions/40084931/taking-subarrays-from-numpy-array-with-given-stride-stepsize">this link</a>中的解决方案和BENY提出的索引的组合</p>
<pre class="lang-py prettyprint-override"><code>def apply_rolling_data(data, col, function, window, step=1, labels=None):
"""Perform a rolling window analysis at the column `col` from `data`
Given a dataframe `data` with time series, call `function` at
sections of length `window` at the data of column `col`. Append
the results to `data` at a new columns with name `label`.
Parameters
----------
data : DataFrame
Data to be analyzed, the dataframe must stores time series
columnwise, i.e., each column represent a time series and each
row a time index
col : str
Name of the column from `data` to be analyzed
function : callable
Function to be called to calculate the rolling window
analysis, the function must receive as input an array or
pandas series. Its output must be either a number or a pandas
series
window : int
length of the window to perform the analysis
step : int
step to take between two consecutive windows
labels : str
Name of the column for the output, if None it defaults to
'MEASURE'. It is only used if `function` outputs a number, if
it outputs a Series then each index of the series is going to
be used as the names of their respective columns in the output
Returns
-------
data : DataFrame
Input dataframe with added columns with the result of the
analysis performed
"""
x = _strided_app(data[col].to_numpy(), window, step)
rolled = np.apply_along_axis(function, 1, x)
if labels is None:
labels = [f"metric_{i}" for i in range(rolled.shape[1])]
for col in labels:
data[col] = np.nan
data.loc[
data.index[
[False]*(window-1)
+ list(np.arange(len(data) - (window-1)) % step == 0)],
labels] = rolled
return data
def _strided_app(a, L, S): # Window len = L, Stride len/stepsize = S
"""returns an array that is strided
"""
nrows = ((a.size-L)//S)+1
n = a.strides[0]
return np.lib.stride_tricks.as_strided(
a, shape=(nrows, L), strides=(S*n, n))
</code></pre>
