Image Classification with a CNN

A TensorFlow/Keras study of ten-class, 28 × 28 grayscale image classification, from IDX data loading to convolutional training and evaluation.

Explore the notebook · Read the report

Inside the experiment

The notebook builds a compact CNN with two convolutional stages (32 and 64 filters), batch normalization, max pooling, a 128-unit dense layer, and a ten-class softmax output. Inputs are normalized to [0, 1], and 20% of the training data is reserved for validation with a fixed split seed.

Training is configured for 100 epochs with batches of 128, categorical cross-entropy, and momentum SGD. The notebook also plots training and validation accuracy and loss, evaluates a held-out test set, and saves the model as fashion_mnist_model.keras.

Recorded result

Original saved training plot from the committed notebook; not a new run.

Measure	Saved notebook output
Test accuracy	90.84%
Test loss	0.6774

These are historical outputs stored in the notebook, not a freshly reproduced benchmark. The saved training log stops partway through epoch 5 even though the code requests 100 epochs; it does not establish the exact training history of the evaluated model.

Open and run

Use Python 3.11 and an isolated environment:

python -m pip install -r requirements.txt
jupyter lab ImageClassificationWithCNN.ipynb

Run all cells from a fresh kernel. The notebook explicitly loads Fashion-MNIST using Keras; the first run downloads the dataset. Training is configured by EPOCHS = 100, which you can lower for an exploratory run. All random generators are seeded through Keras and deterministic TensorFlow operations are enabled. Reproducibility still depends on the software/hardware environment.

For existing local IDX data, data_utils.load_mnist(path, kind='train') and kind='t10k' validate headers, dimensions, counts and class ranges. IDX format alone cannot establish dataset identity; use Fashion-MNIST files for this experiment. Preprocessing supports variable sample counts while requiring 28 × 28 grayscale images.

Maintenance validation

The model now uses the supported SGD optimizer and an explicit input layer. No augmentation is applied, matching the original executed training path; the unused augmentation generator was removed. Corrected notebook cells have cleared outputs so old metrics cannot be mistaken for a new run.

python -m pip install -r requirements-dev.txt
python -m pytest --rootdir=. tests

The validation checks corrupted IDX headers/counts, small valid datasets, preprocessing, and a real model build plus one training batch on synthetic images. The model smoke test uses TensorFlow 2.16.2/Keras 3.15.1 on Python 3.11. It does not reproduce the historical 90.84% accuracy, download the full dataset, or establish generalization performance. The saved figure and report retain the earlier results with that distinction explicit.

Automated checks

GitHub Actions runs the validation suite with Python 3.11 on pushes and pull requests. It checks IDX handling and performs one synthetic training batch; it does not download Fashion-MNIST, run the notebook, train for 100 epochs, or reproduce the saved accuracy.