Unlock the Power of NumPy: Visual Guide to Arrays and Operations

Introduction

NumPy is a fundamental library for multi‑dimensional array and matrix operations, essential for machine learning and many Python data‑processing packages such as pandas, PyTorch, TensorFlow, and Keras.

Core Concepts

NumPy’s core is the n‑dimensional array (ndarray). Operations look the same regardless of dimensionality, though 1‑D (vectors) and 2‑D (matrices) have special considerations.

Vectors – 1‑D arrays

Creating a vector can be done by converting a Python list, e.g. np.array([1, 2, 3]). All elements must share the same dtype; otherwise the array falls back to object. NumPy arrays cannot grow in place, so one usually creates a zero‑filled array with np.zeros or np.empty before filling it.

Common ways to generate vectors include np.arange, np.linspace, and random generators such as np.random.rand. np.arange is sensitive to floating‑point rounding; linspace is preferred when a specific number of points is required.

Indexing, slicing, boolean indexing, and functions like np.where and np.searchsorted are illustrated.

Matrices – 2‑D arrays

Matrix creation mirrors vector syntax but uses double brackets for the optional dtype, e.g. np.array([[1,2],[3,4]], dtype=int). Random matrices can be generated with np.random.rand or np.random.randn.

Two‑dimensional indexing is more convenient than nested Python lists. The axis argument determines whether operations act on rows ( axis=1) or columns ( axis=0).

Matrix arithmetic includes element‑wise operators (+, –, *, /, //) and the matrix‑multiplication operator @. Broadcasting allows mixing vectors and matrices.

Higher‑dimensional arrays

For 3‑D and higher arrays the index order is (z, y, x). When dealing with RGB images the conventional order is (y, x, z). Functions such as np.stack, np.concatenate, and np.transpose are used to reshape and reorder axes.

Common Operations

Statistical functions ( np.mean, np.std, np.min, np.max, etc.) accept an axis argument. Reduction functions like np.all and np.any also support axis.

Sorting NumPy arrays differs from Python list sorting; to sort by a column one can use a[a[:,0].argsort()] or np.lexsort. For stable multi‑column sorting, pandas’ DataFrame.sort_values is often more convenient.

Advanced utilities include np.einsum for Einstein summation, which can replace explicit loops and many tensor operations.

Practical Tips

When comparing floating‑point numbers use np.allclose with appropriate tolerances; be aware of its assumptions about the magnitude of the numbers. For more robust comparisons, Python’s math.isclose may be preferable.

Creating grids with np.meshgrid can be memory‑intensive; broadcasting a pair of 1‑D vectors is often sufficient.

Further Resources

For practice, see the “numpy‑100” repository on GitHub, which contains 100 challenging exercises.

Original article: “NumPy Illustrated – A Visual Guide to NumPy” (Medium).