Fashion Classifier
A neural network built from scratch. An exercise in understanding every step between a handful of pixels and a prediction.
- Python
- NumPy
- Fashion-MNIST
I implemented the layers, gradients, training loop, and image preprocessing in Python and NumPy.
Read projectThe process, the decisions, the lessons.
The question
What actually happens inside a neural network when it learns? I wanted to follow the whole path: from the pixels entering the model to the gradients updating its weights.
Fashion-MNIST gave me a focused problem: classify small grayscale images into ten clothing categories. I built a multilayer perceptron using NumPy so that the forward and backward passes stayed visible in the code.
From pixels to probabilities
Each 28 × 28 image becomes a vector of 784 values, normalized to the range [0, 1]. A fully connected layer maps those inputs to 128 hidden units, followed by ReLU and a final layer with 10 outputs.
The implementation includes He initialization, a numerically stable softmax, and cross-entropy loss. During backpropagation, each layer calculates its gradients and updates its weights using gradient descent.
model = Sequential([
Linear(784, 128),
ReLU(),
Linear(128, 10)
])What I built
- Reusable layers. Linear and ReLU each implement their own forward and backward passes.
- A training loop. The data is shuffled each epoch and processed in batches of 32. The current script runs for 25 epochs with a learning rate of 0.01.
- Weight persistence. Model parameters can be saved to a NumPy archive and loaded again for inference.
- Image preprocessing. The prediction script converts an image to grayscale, crops it to a square, resizes it, and normalizes its pixels before inference.
What this experiment taught me
Writing the backward pass makes the chain rule concrete: each layer receives a gradient, computes its own parameter gradients, and passes a gradient to the layer before it. Keeping track of matrix shapes is as important as getting the equations right.
It also makes preprocessing part of the model’s behaviour. A photograph prepared differently from the training images can produce an unreliable prediction, even when the code runs correctly.
Limits & next questions
This is a learning project. The training script reports training loss; the repository does not currently include a held-out evaluation script. There is no verified test-accuracy result to report here.
The next useful steps are a reproducible train/validation/test evaluation, a confusion matrix, and checks of the analytical gradients against numerical estimates. The prediction script’s softmax output should not be interpreted as calibrated confidence, especially for photographs outside the training distribution.
Explore the implementation
The repository contains the source and a detailed walkthrough of the maths and data pipeline. To run it, you need Python, NumPy, Pillow, idx2numpy, and the Fashion-MNIST data files. The training script creates the weights used by the prediction script.