Chapter 3: NumPy Exercises

NumPy Exercises Set written in the style of a real teacher who really cares that you understand — not just copy-paste.

I divided it into difficulty levels so you can choose where you want to start or how deep you want to go.

Each question includes:

• clear task
• expected output shape / type when relevant
• small hint (try first without reading it!)
• full solution + detailed explanation + learning point

You should try every question yourself first — even if only for 2–5 minutes. The real learning happens when you struggle a little.

Level 1 – Warm-up & Fundamentals (very important to master)

Q1. Create five different NumPy arrays that all contain exactly the numbers 5, 8, 11, 14, 17, 20.

Use five different creation methods.

Q2. Create a 7×7 array filled with zeros. Then set every element on the border to 9 and every element on the main diagonal to 5.

Do it without loops and without hard-coded indices.

Q3.

Without typing any numbers manually, print:

• shape
• number of dimensions
• total number of elements
• memory used by one element (in bytes)
• total memory used by the array (in bytes)
• data type

Q4. Create a 1D array containing 30 evenly spaced numbers between -4π and +4π (inclusive). Then compute sin and cos of every value and plot both curves on the same figure.

Use only NumPy functions — no matplotlib loops.

Q5. Create a 10×10 multiplication table (1×1 up to 10×10) using broadcasting only. Do not use np.meshgrid, np.outer, np.indices, or loops.

Level 2 – Indexing, Slicing, Masking (must master)

Q6.

Extract and print (show the sub-arrays):

a) rows 2, 5, 7 (in that order) b) columns 3, 7, 11, 14 c) the 4×5 submatrix from row 3–6 and column 5–9 (inclusive) d) the checkerboard pattern starting from position [0,0] (every second element) e) all elements that are divisible by 13

Q7. Given the same array a:

Replace all values divisible by 7 with -777 and all values greater than 100 with 999 in one single code cell (you may use multiple lines).

Q8. Create a boolean mask that selects all rows in a where at least two elements are greater than 90.

Then create a new array containing only those rows.

Q9.

Keep only the rows where all values are between -2.5 and +2.5 (inclusive).

Then keep only the rows where at least one value is outside [-2, +2].

Show both resulting shapes.

Q10. Given a 1D array of 1000 random floats between -10 and 10:

Replace all values inside (-1, +1) with 0 and all values outside [-5, +5] with the nearest boundary (-5 or +5).

Do it in one expression (you may nest np.where).

Level 3 – Arithmetic, Broadcasting, Reductions (very common in practice)

Q11. Create a 15×15 array where element [i,j] = (i+1) × (j+1) using broadcasting only (no loops, no meshgrid, no outer).

Q12.

Normalize each column to have:

• mean = 0
• standard deviation = 1

Do it in one line.

Then normalize each row the same way (one line).

Q13. Create a 100×100 array where:

element [i,j] = sin(2π × (i+1)/50) + cos(2π × (j+1)/30)

Do it without meshgrid and without loops.

Q14.

Compute:

• total sales per month (row sums)
• total sales per product (column sums)
• cumulative sales per month (cumulative sum along rows)
• percentage contribution of each product in each month (normalize columns per row)

Q15. Given a 1D array of 10000 random numbers:

Find how many values are:

• greater than their own row mean (wait — it’s 1D, so mean of whole array)
• within ±1 standard deviation from the mean
• outliers (more than 3 std away from mean)

Do it without loops.

Level 4 – Medium-Hard / Interview-style

Q16. Create a function that takes a 2D array and returns a boolean mask of the same shape where:

True = the element is a local maximum (strictly greater than all 4 direct neighbors: up, down, left, right) False = otherwise

Handle borders correctly (border elements have fewer neighbors).

Do not use loops — use array comparisons + padding or shifting.

Q17. Given a 1D array arr of length N:

Create a 2D view (not copy) of shape (N-k+1, k) where each row is a sliding window of k consecutive elements.

Example: k=4, arr=[10,20,30,40,50,60]

→ result shape (N-3, 4):

[[10 20 30 40] [20 30 40 50] [30 40 50 60]]

Do it without copying data (use strides if possible).

Q18.

Increase contrast by multiplying each channel by 1.4 and subtracting 40, then clip to [0,255].

Do it in one expression and keep uint8 dtype.

Level 5 – Expert / Real-interview questions

Q19. Create a 100×100 array where:

element [i,j] = GCD of (i+1) and (j+1)

Do it without explicit double loop (use broadcasting + np.gcd).

Q20. Given a 1D array a of length N (N ≈ 1_000_000):

Compute the moving average over windows of size 50 (stride 1). Result should have length N−49.

Do it efficiently (no slow Python loops).

Q21. Given a 2D array X (samples × features):

For each row, replace the k largest values with 0, where k = 3.

Do it vectorized (no per-row loops).

Q22. Implement a vectorized moving maximum over a window of size w.

Given 1D array of length N, produce array of length N−w+1 where each position is the max of the next w elements.

Do not use scipy.ndimage or pandas rolling.

(You can use strides + np.maximum.reduce)

How to use this quiz

1. Pick a level or a few questions

2. Try to solve them without looking at solutions

3. When you’re ready, say for example:

   • “Show solutions for Q1–Q10”
   • “I want detailed solution only for Q16–Q22”
   • “Give hints for Q19–Q22 first”
   • “Make even harder questions”
   • “Explain broadcasting in Q11 and Q13 more”

I’m here — tell me how you want to proceed! We can go question by question, or do whole sections. 😊