Mahmoud Abas, Full-Stack Software Engineer & AI

Neural Networks - Neural Networks Assignment: Fashion MNIST with Keras

In the print dialog, choose "Save as PDF" as the destination.

Neural Networks - Lesson 7

Neural Networks Assignment: Fashion MNIST with Keras

Build, train, evaluate, and improve a neural network end to end

Objective and Grading

Build, train, evaluate, and improve a neural network on the Fashion MNIST dataset using Keras (TensorFlow)
Graded in 5 parts, one grade each: data, model, training, evaluation, improvement
Deliverable: a single, well-commented Jupyter Notebook

Step 1

Data Loading and Exploration

Load Fashion MNIST from tensorflow.keras.datasets
Print the shape of the training and test sets
Normalize pixel values to the range 0-1
Display 4 sample images from different classes with labels via matplotlib
Create a dictionary mapping class indices (0-9) to class names

Step 2

Build the Neural Network

Flatten

reshapes the input

Dense hidden layer

ReLU activation

Dense output layer

10 units, softmax

Build a Sequential model in Keras with at least one hidden layer
Print the model summary to verify the architecture

Step 3

Compile the Model

Setting	Value
Loss	`sparse_categorical_crossentropy`
Optimizer	`adam`
Metric	`accuracy`

Step 3, continued

Train and Visualize

Train for a few epochs (roughly 4-6)
Batch size guidance: 32 for a human-run notebook, 128 if run by an automated or robot process
Plot training accuracy and loss over epochs with matplotlib

Step 4

Evaluate and Predict

Evaluate the trained model on the test set; report test loss and accuracy
Predict the labels of 6 test images
Display each test image alongside its predicted label and its true label

Step 5

Improve the Model and Reflect

Modify the model to improve performance, for example adding more layers or changing activation functions or the number of units
Re-train the modified model and compare its accuracy to the original
Write a short reflection (3-5 sentences) on what changed and its effect on performance

Submission Guidelines

Submit a single, well-commented Jupyter Notebook
All cells must run top-to-bottom without errors
Include titles and axis labels on every plot
Use markdown cells to explain steps and results

Key Takeaways

Load

Explore

Build

Train

Evaluate

Improve

Data must be normalized and explored (sample images, shapes) before model building
A Flatten + Dense (ReLU) + Dense (softmax) Sequential model is the minimum viable image classifier
sparse_categorical_crossentropy + adam + accuracy is the standard compile configuration for integer-labeled multi-class problems
Visualizing accuracy and loss curves and predicted-vs-true labels is essential for diagnosing model behavior
Iteration (adding layers, tuning units and activations) and reflection on the result are graded as their own step

Open this lesson

Mahmoud AbasNeural Networks Assignment: Fashion MNIST with Keras

Setting

Value

Loss

sparse_categorical_crossentropy

Optimizer

adam

Metric

accuracy