NumPy Learning Roadmap

Why NumPy

NumPy is the foundation of Python’s scientific computing stack. pandas, scikit-learn, PyTorch, and TensorFlow all use NumPy arrays at their core. Without NumPy fluency, you’ll fight your tools instead of solving problems.

Stage 1: Array Fundamentals (Week 1)

Goal: Create, index, and manipulate arrays without looking at docs.

Topics:

• Introduction to NumPy — what arrays are, why not lists
• Arrays and Dtypes — shapes, dtypes, memory layout
• Indexing and Slicing — fancy indexing, boolean masks, views vs copies

Milestone: Write a function that takes a 2D array of student scores and returns per-student grades, class averages per subject, and a pass/fail mask — using only array operations, no loops.

Stage 2: Computation Patterns (Week 2)

Goal: Write vectorized code instead of Python loops.

Topics:

• Broadcasting — shape rules, outer products, adding dimensions
• Universal Functions — ufunc methods, at(), reduceat()
• Vectorization — benchmark loops vs vectorized, when to profile

Milestone: Implement pairwise cosine similarity for N vectors using broadcasting only (no loops, no sklearn).

Stage 3: Math and Algorithms (Week 3)

Goal: Use NumPy for linear algebra, statistics, and matrix operations.

Topics:

• Matrix Operations — dot, matmul, einsum
• Linear Algebra — SVD, eigenvalues, solving linear systems
• Random and Simulation — Generator API, Monte Carlo, bootstrap

Milestone: Implement PCA from scratch: center data, compute covariance, eigen-decompose, project to k components. Verify against sklearn’s PCA.

Stage 4: Performance and Scale (Week 4)

Goal: Write NumPy code that performs well on large datasets.

Topics:

• Performance Optimization — memory layout, copy vs view, profiling

Milestone: Benchmark 5 implementations of a sliding window mean: pure Python loop, numpy loop, stride tricks, and vectorized diff. Understand where each beats the others.

Stage 5: Projects

Beginner: Statistics calculator, grade book analyzer, dice simulator
Intermediate: K-Means from scratch, image compression via SVD, Monte Carlo options pricing
Advanced: Backpropagation engine, numerical ODE solver, FFT-based convolution

See NumPy Projects for full descriptions.

What Comes After NumPy

• pandas — tabular data built on NumPy arrays
• scikit-learn — ML algorithms that consume NumPy arrays
• PyTorch / TensorFlow — deep learning frameworks with NumPy-like APIs
• SciPy — scientific algorithms (optimization, signal processing, statistics)

Resources

• Official NumPy documentation and User Guide
• “Python for Data Analysis” by Wes McKinney (chapters 4-5)
• NumPy’s own tutorial: numpy.org/doc/stable/user/quickstart.html