Skip to main content

Stage 9: Use It on New Images

Course progressStage 9 of 10
~90 min
Your workspace

Keep your Colab notebook tab open all session. Open in a new tab — don’t use the buttons in this page to leave the course.

Train your AIGoogle ColabOpen in a new tab ↗No-code warm-upMachine Learning for KidsOpen in a new tab ↗
Build

an inference tool that preprocesses uploaded images and prints top guesses

Learn

why prediction uses the same preparation rules as training

Ship

a tested upload-and-predict workflow with confidence interpretation

Teacher demo

Upload one valid CIFAR-10 class image and one off-menu image, like pizza. Show that the model always chooses from its 10 known classes.

The big idea​

Using a trained model on a new image is called inference. The model does not learn during inference. It only guesses.

The new image must be translated into the same format as training data: 32 by 32 pixels, RGB, values from 0 to 1, and a batch dimension.

How the Python ML workflow connects
  1. 1
    Photos / CIFAR-10labeled image examplesStage 1
  2. 2
    Notebook variablesx_train, y_train, class_namesSetup-2
  3. 3
    Prepared datanormalized pixels and fair pilesStage 3
  4. 4
    Keras modelCNN layers and summaryStage 4
  5. 5
    Training historyepochs, loss, accuracyStage 5
  6. 6
    Test evidencesealed score and mistakesStages 6-7
  7. 7
    Improved modelaugmentation comparisonStage 8
  8. 8
    Inferenceuploaded image to top-3 guessesStage 9
  9. 9
    Demo evidencetable, confidence, limitationStage 10

Stage 9 uses the trained model without changing it. Uploaded images must pass through the same preparation rules before the model can make top-3 guesses.

New words
inference
using a trained model to make a prediction
preprocess
prepare new data in the same format as training data
confidence
the model's score for a predicted class
top-3
the three most likely classes according to the model
Before you start

Use augmented_model from Stage 8 if it performed at least as well as baseline. Otherwise use model. Set final_model in Step 1.

Build it​

Step 1 — Choose the model to demo​

final_model = augmented_model

If your augmented model performed worse and you want to demo the baseline, use:

# final_model = model

Step 2 — Write the preprocessing function​

The image helper loads a file and converts it to an array. np.expand_dims adds the batch dimension, turning one image into a batch of one image. The model was trained on batches, so even a single uploaded image must keep that shape.

def load_and_preprocess_image(image_path):
    from tensorflow.keras.preprocessing import image
    import numpy as np

    img = image.load_img(image_path, target_size=(32, 32))
    img_array = image.img_to_array(img)
    img_array = img_array / 255.0
    img_array = np.expand_dims(img_array, axis=0)
    return img_array

This matches the training format exactly.

Step 3 — Predict top guesses​

Upload an image in Colab, copy its path, and paste it below.

new_image_path = 'my_image.jpg'
new_image = load_and_preprocess_image(new_image_path)

predictions = final_model.predict(new_image)[0]
top_indexes = predictions.argsort()[-3:][::-1]

plt.imshow(plt.imread(new_image_path))
plt.axis('off')
plt.show()

for index in top_indexes:
    print(f"{class_names[index]}: {predictions[index]:.0%}")

The top class is the model's answer. The next two show what it almost chose.

Step 4 — Test an off-menu image​

Upload something not in CIFAR-10: pizza, phone, person, or shoe. Run the same cell.

The model still picks one of its 10 classes because it has no "none of these" output.

Step 5 — Write a prediction note​

Prediction note:
- My in-class image was __________.
- Top guess: __________ at _____%.
- Off-menu image: __________.
- The model called it __________ because it only knows 10 classes.

Understand it​

Preprocessing is not optional. A model trained on 32x32 normalized images expects every new image to look like that numerically.

Confidence is not the same as truth. A wrong model can be confident, especially when forced to choose from only 10 classes. Top-3 results give a more honest view of uncertainty.

Try this​

Learning beat

Try this

Three short experiments. Predict before you run, then test your guess.

Predict first

Predict whether a clear image or a blurry image will get higher confidence. Test both.

Compare

Compare top-3 guesses for an in-class image and an off-menu image. Which result looks more uncertain?

Connect

Which Stage 7 mistakes explain the new-image mistakes you saw today?

Test your stage​

  • You predicted top-3 classes for one in-class image.
  • You tested one off-menu image.
  • You wrote a prediction note.
  • Workflow check. Point to this stage on the workflow map and explain why uploaded images need the same preparation rules.
  • Evidence check. Your prediction note includes one top-3 result and one confidence value.
  • Design check. Explain why the model cannot say "none of these."

If it breaks​

  • FileNotFoundError. Re-copy the exact path from Colab Files.
  • final_model is undefined. Run Step 1.
  • The image is huge on screen. Add plt.figure(figsize=(3, 3)) before imshow.
Coach notes

The off-menu image is essential. It prevents students from thinking AI knows when it is outside its training world.