Anatomy of an array

The attributes

import numpy as np

arr = np.array([[1, 2, 3], [4, 5, 6]])
print(arr.shape, arr.ndim, arr.size)
print(arr.dtype, arr.itemsize, arr.nbytes)
print(arr.T)
print(arr.strides)

(2, 3) 2 6
int64 8 48
[[1 4]
 [2 5]
 [3 6]]
(24, 8)

The strides is the internal detail that makes slicing so fast: to move one row, NumPy skips 24 bytes (3 elements of 8 bytes), and to move one column, it skips 8. You don't need to calculate it, but chapter 29 shows how it allows creating "views" without copying data.

Shape is not size

The most frequent mistake of beginners is to swap shape for size. They answer different questions: the shape tells how the elements are organized, and the size tells how many there are. An array of shape (4, 5) has size 20.

A one-dimensional array has shape of a element, and the comma is needed in the tuple. And there are zero-dimensional arrays (scalars), whose shape is the empty tuple:

print(len(arr), np.array([1, 2, 3]).shape, np.array(5).shape, np.array(5).ndim)

2 (3,) () 0

The len(arr) returns only the size of the first dimension (2 rows). For the total number of elements, use size.

Why dtype matters more than it seems

A million of float64 takes up double the memory of the same million in float32. In a toy dataset this is irrelevant, but when the dataset grows to hundreds of millions of values, choosing the right type decides whether it fits in memory. The next chapter deals with this.