NumPy Best Practices for ML
You'll learn: How to write efficient, maintainable NumPy code that integrates smoothly with machine learning workflows.
Why Best Practices Matter
After learning individual NumPy operations, it's time to combine them into professional ML code. Poor NumPy habits waste memory, slow down training, and create subtle bugs that corrupt your models.
Avoid Explicit Loops
Python loops are slow. You've already learned vectorization and broadcasting—always use them instead of for loops:
# BAD: Looping through elements
result = np.zeros(10000)
for i in range(10000):
result[i] = data[i] * 2 + 1
# GOOD: Vectorized operation
result = data * 2 + 1 # 100x faster!
Manage Data Types Carefully
The dtype controls both precision and memory usage. Use float32 instead of float64 when precision allows—it halves memory and speeds up computation:
# Default is float64 (8 bytes per number)
X = np.random.randn(10000, 100)
# Use float32 for most ML tasks (4 bytes per number)
X = np.random.randn(10000, 100).astype(np.float32)
Integrate with scikit-learn
scikit-learn expects NumPy arrays in specific shapes:
- Features: 2D array
(n_samples, n_features) - Labels: 1D array
(n_samples,)or 2D(n_samples, 1)
Always verify shapes before passing data to models using .shape and reshape when needed.
Memory and Performance
Reuse arrays when possible instead of creating new ones. Use in-place operations (like +=) and preallocate arrays with np.zeros() for sequential filling.
Key Takeaway: Professional NumPy code avoids loops through vectorization, uses appropriate dtypes to balance precision and memory, and maintains array shapes compatible with ML libraries—these practices prevent bugs and dramatically improve performance.