Advanced NumPy Techniques

NumPy becomes powerful when you move beyond basics (array creation, slicing) and start using vectorization, broadcasting, views, strides, structured arrays, and performance tricks.

1. Vectorization (Fast Computation without Loops)

Vectorization allows replacing slow Python loops with fast NumPy operations.

❌ Slow (Python loop)

result = [x**2 for x in range(10_000)]

✔ Fast (NumPy vectorized)

import numpy as np
arr = np.arange(10_000)
result = arr ** 2

2. Broadcasting (Different Shapes Operations)

NumPy automatically “broadcasts” arrays of different sizes.

Example:

a = np.array([1, 2, 3])
b = 5
print(a + b)  # [6, 7, 8]

Broadcasting Rules:

1. Dimensions must be equal or

2. One of them must be 1.

3. Advanced Indexing

✔ Boolean indexing

arr = np.array([10,20,30,40,50])
arr[arr > 25]

✔ Fancy indexing (index array)

arr[[0, 3, 4]]  # pick elements

✔ 2D fancy indexing

mat = np.arange(16).reshape(4,4)
mat[[0,2], [1,3]]  # picks (0,1) and (2,3)

4. Views vs Copies (Memory Efficiency)

✔ View → shares memory

✔ Copy → separate memory

a = np.arange(10)
b = a[2:6]     # view
b[0] = 999
print(a)       # original changes

To force a copy:

c = a[2:6].copy()

5. Strides (Internal Memory Tricks)

Strides define how NumPy moves in memory.

Example: Create a sliding window view:

from numpy.lib.stride_tricks import sliding_window_view

arr = np.arange(10)
windows = sliding_window_view(arr, window_shape=3)

6. Vectorized String Operations

np.char.upper(['india', 'usa', 'uK'])
np.char.replace(['patna', 'delhi'], 'a', '@')

7. Matrix Operations for ML

✔ Dot Product

np.dot(a, b)

✔ Matrix Multiplication

A @ B

✔ Transpose

A.T

✔ Determinant & Inverse

np.linalg.det(A)
np.linalg.inv(A)

8. NumPy Broadcasting Tricks

Add vector to each row

A + b   # shape (3,3) + (3,)

Add vector to each column

A + b[:, None]

9. Performance Tips

✔ Use vectorized operations

✔ Avoid Python loops

✔ Use astype() wisely

✔ Use in-place operations

arr += 5  # faster

✔ Use efficient data types

arr = arr.astype(np.float32)

✔ Preallocate arrays instead of appending

10. Structured Arrays (Like SQL Tables)

data = np.array([
    (1, 'Ram', 25),
    (2, 'Shyam', 30)
], dtype=[('id','i4'), ('name','U10'), ('age','i4')])

data['name']

11. Random Module (ML Essential)

np.random.rand(3,3)
np.random.randn(3,3)
np.random.randint(1, 10, size=5)
np.random.seed(42)  # reproducibility

12. Concatenation & Splitting

np.concatenate([a, b])
np.vstack([A, B])
np.hstack([A, B])
np.split(arr, 3)

Summary of Advanced Concepts