I would like to run an operation (e.g. subtracting the median) on rows of a numpy array.
One way to do that is using comprehension lists:
import numpy as np
from statistics import median
x = np.array([[1, 2, 3, 4], [5, 6, 7 ,8], [9, 10, 11, 12]])
xm = np.vstack(([x[i,:] - median(x[i,:]) for i in range(x.shape[0])]))
Each row is processed, then stacked vertically as numpy array.
Is there a more efficient/elegant way to do that?
x - np.median(x, axis=1)[:, np.newaxis]
given np.median has a keepdims parameter you can also avoid the manual slicing to make it broadcasting-friendly
x - np.median(x, axis=1, keepdims=True)
if you want to apply an arbitrary function row by row, like median from statistics, you can use np.apply_along_axis, just beware it's basically a for loop so you don't really get any vectorization speedup:
x - np.apply_along_axis(median, axis=1, x)[:,np.newaxis]
statistics library, I guess you could use np.apply_along_axis(median, 1, x) but it's not as efficent as numpy native one
numpy's median and statistics' median - they should return the same results in all situations except they behave slightly differently when input is an empty list. I would just use numpy version.