TopolearnSummaries per row, per column or over everything are the most common analysis step, and axis mistakes are the most common bug.
Aggregations collapse many values into fewer: sum, mean, median, min, max, std, var, prod, argmin/argmax (position of the extreme), any/all, and the cumulative versions cumsum/cumprod. Most exist as both functions (np.sum(a)) and methods (a.sum()).
Without arguments they reduce over every element. The axis argument chooses which axis to collapse:
- For a 2-D table of shape
(rows, cols),axis=0collapses the rows and gives one value per column, shape(cols,). axis=1collapses the columns and gives one value per row, shape(rows,).
A reliable way to remember: the axis you name is the one that disappears from the shape. With a shape (10, 24, 7) array, a.mean(axis=1) has shape (10, 7). You can pass a tuple, axis=(1, 2), to collapse several at once.
keepdims=True keeps the collapsed axis with length 1, so the result broadcasts back against the original. That is how you normalise each row: a / a.sum(axis=1, keepdims=True).
Two details trip people up. np.std uses the population formula by default (ddof=0); pass ddof=1 for the sample standard deviation that pandas and most statistics texts use. And argmax without an axis returns an index into the flattened array; np.unravel_index converts it back to a (row, col) pair.
If the data contains nan, regular aggregations return nan; the np.nansum, np.nanmean family skips missing values.
Example
import numpy as np
sales = np.array([[3, 7, 2], # rows: stores, cols: days
[5, 1, 8]])
print(sales.sum()) # 26
print(sales.sum(axis=0)) # [ 8 8 10] per day (column)
print(sales.sum(axis=1)) # [12 14] per store (row)
print(sales.max(axis=1)) # [7 8]
i = sales.argmax() # index in the flattened array
print(i, np.unravel_index(i, sales.shape)) # 5 (np.int64(1), np.int64(2))
share = sales / sales.sum(axis=1, keepdims=True)
print(share.round(2)) # each row sums to 1
x = np.array([2.0, 4.0, 4.0, 6.0])
print(x.std(), x.std(ddof=1)) # 1.414... 1.632...
print(np.nanmean([1.0, np.nan, 3.0])) # 2.0Remember
axis=kcollapses axis k: it vanishes from the result's shape- On a 2-D table, axis=0 gives per-column results, axis=1 per-row
keepdims=Truekeeps a length-1 axis so results broadcast backstddefaults to ddof=0; usenan*functions to skip missing values
Watch out for
- Mixing up axis=0 and axis=1: check the result's shape against what you expect.