Fashion-MNIST

Import Fashion-MNIST via Keras.

Insert the new custom layer into the model.

Run 5-10 epochs of training.

Verify if loss converges faster than the baseline.

Quantize a pre-trained Fashion-MNIST model to INT8.

Deploy the model to the target edge device.

Run the 10,000 test images through the device.

Measure latency per image and classification accuracy drop.

Provide a Jupyter notebook with the data pre-loaded.

Have students visualize the 10 classes of clothing.

Guide students to build a simple MLP (Multi-Layer Perceptron).

Observe the performance gap compared to MNIST.

Define a Generator and Discriminator network.

Train using the Fashion-MNIST images.

Iterate on the architecture to prevent mode collapse.

Evaluate generated 'fake' boots or shirts visually.

Set up a GitHub Action triggered on push.

Script a 'smoke test' that trains a model on a subset of Fashion-MNIST.

Assert that accuracy exceeds a 70% threshold.

Fail the build if the threshold is not met.

Train an autoencoder to compress images to a latent vector.

Reconstruct images from the vector.

Calculate the Mean Squared Error (MSE).

Compare results against standard JPEG or WebP compression.

Start with a very small labeled subset of Fashion-MNIST.

Use an uncertainty sampling algorithm to pick the next batch.

Label the picked batch.

Measure the accuracy gain per labeled image.