NumPy Best Practices

NumPy is the fundamental package for scientific computing with Python. It provides N-dimensional array objects, vectorized math operations, broadcasting, linear algebra, Fourier transforms, and random number generation. This skill covers best practices for writing correct, efficient, and maintainable NumPy code.

Import Convention

Always import NumPy with the standard alias:

import numpy as np

Never use from numpy import * — it pollutes the namespace and makes code harder to read.

Array Creation

Choose the right creation function

Use case	Function
Known values	np.array([1, 2, 3])
Zeros	np.zeros(shape)
Ones	np.ones(shape)
Uninitialized (fill later)	np.empty(shape)
Integer range	np.arange(start, stop, step)
Evenly spaced floats	np.linspace(start, stop, num)
Identity matrix	np.eye(n)
Like existing array	np.zeros_like(arr), np.ones_like(arr)

Specify dtype explicitly

Always specify dtype when the intended type differs from NumPy's default (float64 for floats, int64 for integers):

# Explicit dtype avoids silent precision issues
weights = np.ones(1000, dtype=np.float32)   # saves memory for ML models
indices = np.arange(100, dtype=np.int32)    # sufficient range, half memory
flags = np.zeros(50, dtype=np.bool_)        # boolean array

Do not rely on implicit upcasting — declare the dtype the data actually needs.

Use np.random.default_rng() for random numbers

The legacy np.random.* functions (e.g., np.random.rand) are deprecated in favour of the Generator API:

# Correct — reproducible, modern API
rng = np.random.default_rng(seed=42)
samples = rng.normal(loc=0.0, scale=1.0, size=(100, 3))
integers = rng.integers(0, 10, size=50)

# Avoid — legacy API, global state
np.random.seed(42)
samples = np.random.randn(100, 3)

Pass seed to default_rng for reproducibility in tests and experiments.

Vectorization Over Loops

Replace Python loops with vectorized NumPy operations wherever possible. NumPy operations execute in optimized C code, making them orders of magnitude faster.

# Avoid — Python loop
result = []
for x in data:
    result.append(x ** 2 + 2 * x + 1)
result = np.array(result)

# Correct — fully vectorized
result = data ** 2 + 2 * data + 1

Use np.vectorize only as a convenience wrapper for scalar functions — it does not improve performance since it still calls Python per element.

Broadcasting Rules

Broadcasting allows operations on arrays of different shapes without copying data. Apply broadcasting instead of explicit tile or repeat calls.

Broadcasting rules (trailing dimensions are compared):

1. Dimensions are equal — compatible.
2. One dimension is 1 — that dimension is stretched.
3. Otherwise — ValueError.

# Add a bias vector to each row of a matrix — broadcasting handles shape (3,) vs (4, 3)
matrix = np.zeros((4, 3))
bias = np.array([1.0, 2.0, 3.0])
result = matrix + bias   # shape (4, 3); no copy made

# Outer product via newaxis
a = np.array([0.0, 10.0, 20.0])    # shape (3,)
b = np.array([1.0, 2.0, 3.0])      # shape (3,)
outer = a[:, np.newaxis] + b        # shape (3, 3)

Avoid broadcasting that produces very large intermediate arrays — use an explicit loop for memory-constrained cases.

Views vs Copies

Basic indexing (slices) returns a view — modifying it modifies the original:

x = np.arange(10)
y = x[2:5]     # view — shares memory
y[0] = 99      # also changes x[2]

Advanced indexing (integer arrays, boolean masks) returns a copy:

x = np.arange(10)
idx = [1, 3, 5]
y = x[idx]     # copy — independent of x

Check ownership with arr.base:

y.base is None    # True → copy
y.base is x       # True → view of x

When to force a copy

Call .copy() explicitly when an independent array is needed:

backup = original.copy()

Use .ravel() (view when possible) over .flatten() (always copies) when write access to the parent is acceptable. Use reshape(-1) as the most reliable way to get a flat view.

Indexing and Selection

Boolean indexing for filtering

arr = np.array([1, -2, 3, -4, 5])
positive = arr[arr > 0]           # copy: [1, 3, 5]
arr[arr < 0] = 0                  # in-place modification via boolean mask

np.where for conditional selection

# Replace negatives with zero, keep positives
cleaned = np.where(arr > 0, arr, 0)

Avoid loops for aggregations

Use axis-aware aggregation functions instead of looping over rows or columns:

matrix = np.arange(12).reshape(3, 4)
row_sums = matrix.sum(axis=1)    # sum each row → shape (3,)
col_max  = matrix.max(axis=0)    # max each column → shape (4,)

Data Types and Precision

Choose the smallest sufficient dtype

Scenario	Recommended dtype
ML model weights	np.float32
High-precision scientific	np.float64
Small integer counts (<32 k)	np.int16
Large integer counts	np.int32 or np.int64
Boolean flags	np.bool_
Complex numbers	np.complex64 or np.complex128

Use arr.astype(np.float32, copy=False) to cast in-place when the data is already the right type — copy=False avoids an unnecessary allocation.

Watch for integer overflow

NumPy integer arithmetic wraps silently:

x = np.array([200], dtype=np.int8)   # max 127
x + 100    # array([ 44], dtype=int8) — silent overflow!

Cast to a wider type before operations that risk overflow.

Saving and Loading Arrays

Format	Function	Use case
Single array (binary)	np.save / np.load	Fast, preserves dtype and shape
Multiple arrays	np.savez / np.savez_compressed	Archive multiple arrays
Text (CSV etc.)	np.savetxt / np.loadtxt	Human-readable interchange

# Save and reload with full metadata preserved
np.save("data.npy", arr)
arr_loaded = np.load("data.npy")

# Save several arrays
np.savez("dataset.npz", X=X_train, y=y_train)
npz = np.load("dataset.npz")
X_train = npz["X"]

Prefer .npy/.npz over text formats for large arrays — binary I/O is faster and lossless.

Reshaping and Shape Manipulation

Use -1 as a wildcard dimension — NumPy infers the correct size:

flat = arr.reshape(-1)        # flatten to 1-D (view when possible)
col  = arr.reshape(-1, 1)     # column vector
row  = arr.reshape(1, -1)     # row vector

Use np.newaxis (equivalent to None) to insert a dimension for broadcasting:

a = np.array([1, 2, 3])       # shape (3,)
a_col = a[:, np.newaxis]      # shape (3, 1)
a_row = a[np.newaxis, :]      # shape (1, 3)

Linear Algebra

Use np.linalg for matrix operations:

A = np.array([[1, 2], [3, 4]], dtype=np.float64)

# Matrix multiplication — use @ operator (Python 3.5+)
C = A @ B             # preferred over np.dot(A, B) for 2-D

# Common operations
vals, vecs = np.linalg.eig(A)
inv_A      = np.linalg.inv(A)
rank       = np.linalg.matrix_rank(A)
det        = np.linalg.det(A)

# Solve Ax = b without computing inverse (faster, more stable)
x = np.linalg.solve(A, b)     # preferred over inv(A) @ b

Never use np.matrix — it is deprecated. Use 2-D ndarray with @ instead.

Quick Reference

Task	Idiomatic code
Import	import numpy as np
Array from list	np.array([1, 2, 3])
Shape / ndim / size	arr.shape, arr.ndim, arr.size
Reshape	arr.reshape(rows, -1)
Flatten (view)	arr.reshape(-1) or arr.ravel()
Flatten (copy)	arr.flatten()
Transpose	arr.T or arr.transpose()
Boolean mask	arr[arr > 0]
Axis aggregation	arr.sum(axis=0)
Matrix multiply	A @ B
Copy	arr.copy()
Check view	arr.base is not None
Cast dtype	arr.astype(np.float32, copy=False)
Random (modern)	np.random.default_rng(seed)
Save / load	np.save / np.load

Additional Resources

Reference Files

For deeper guidance, consult:

• references/performance-and-memory.md — Vectorization patterns, memory layout, dtype selection, profiling, and avoiding common performance traps
• references/array-operations.md — Broadcasting in depth, advanced indexing, ufuncs, structured arrays, and I/O patterns