Skip to main content
Donate
All projects
Machine learning project

Handwritten-digit recogniser

scikit-learn ships with about 1,800 tiny scans of handwritten digits, each 8 by 8 pixels. You'll draw a few of them as text so you can see what the model sees, train a k-nearest neighbours model, and look at the digits it gets wrong.

Skills you'll practise

  • Images as numbers
  • k-nearest neighbours
  • Accuracy
  • Reading mistakes
Open starter in playgroundSee the steps

Steps

  1. Step 1: Look at a digit

    Load the digits and print one 8×8 image as text, using a character for each shade of grey.

    Show a hint

    Pixel values run from 0 (blank) to 16 (darkest). Picking a character by pixel // 5 gives four shades.

  2. Step 2: Split the scans

    Use digits.data, where each image is flattened into a row of 64 numbers, and hold back a quarter for testing.

  3. Step 3: Train kNN

    Fit a KNeighborsClassifier with three neighbours on the training scans.

    Show a hint

    kNN labels a new digit by finding the training digits whose 64 pixel values are closest to it.

  4. Step 4: Measure it

    Print the accuracy on the test scans.

  5. Step 5: Look at the mistakes

    Find the test digits the model got wrong and draw one, with what it guessed and what it really was.

    Show a hint

    wrong = np.where(predictions != y_test)[0] gives the positions of the mistakes.

Starter code

It already runs. The TODO comments mark where to start.

Open in playground
main.py
from sklearn.datasets import load_digits

digits = load_digits()

SHADES = " .:#"  # blank, light, medium, dark


def draw(image):
    for row in image:
        print("".join(SHADES[int(pixel) // 5] for pixel in row))


print("This is a", digits.target[0])
draw(digits.images[0])

# TODO: split digits.data and digits.target, train a
# KNeighborsClassifier(n_neighbors=3), and print its accuracy.

Example solution

One way to finish it. Have a go first; yours doesn't need to match.

Reveal the solution
main.py
import numpy as np
from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
from sklearn.neighbors import KNeighborsClassifier
from sklearn.metrics import accuracy_score

digits = load_digits()

SHADES = " .:#"  # blank, light, medium, dark


def draw(image):
    for row in image:
        print("".join(SHADES[int(pixel) // 5] for pixel in row))


X_train, X_test, y_train, y_test = train_test_split(
    digits.data, digits.target, test_size=0.25, random_state=0, stratify=digits.target
)

model = KNeighborsClassifier(n_neighbors=3)
model.fit(X_train, y_train)

predictions = model.predict(X_test)
print(f"Accuracy on unseen digits: {accuracy_score(y_test, predictions):.1%}")

wrong = np.where(predictions != y_test)[0]
print(f"It got {len(wrong)} of {len(y_test)} wrong.")

if len(wrong) > 0:
    first = wrong[0]
    print()
    print(f"It guessed {predictions[first]}, but this is a {y_test[first]}:")
    draw(X_test[first].reshape(8, 8))
Open solution in playground

Stretch goal

Use matplotlib to show a grid of the digits it got wrong with plt.imshow(image, cmap="gray_r"), titled with the guess and the true label.

Chapters that help