Neural Network for Fashion MNIST

A neural network classifier built from scratch in NumPy (no PyTorch training loop) that recognizes 10 categories of clothing from 28×28 grayscale images using 5-fold cross-validation and a hand-coded Adam optimizer with learning rate decay.

---

Project Structure

NN for FASHION/
├── src/
│   ├── data/
│   │   └── fashion_numpy.py    # Download, flatten & normalise Fashion MNIST
│   ├── models/
│   │   └── numpy_net.py        # Layer_Dense, Dropout, ReLU, Softmax, Adam
│   └── training/
│       └── cross_val.py        # k-fold CV, train_model, evaluate_model
├── checkpoints/                # Saved weights (git-ignored)
├── logs/                       # Output logs (git-ignored)
├── train_numpy.py              # Entry point
├── conftest.py
├── pytest.ini
└── requirements.txt

---

Architecture

All layers are implemented manually in NumPy:

Component	Detail
Layer_Dense	Linear transform with L1/L2 regularization on weights & biases
Activation_ReLU	Element-wise rectified linear unit
Layer_Dropout	Inverted dropout (disabled during inference)
Activation_Softmax + Loss_CategoricalCrossentropy	Fused for numerically stable backprop
Optimizer_Adam	Bias-corrected Adam with inverse learning rate decay

Network: 784 → 512 → 512 → 256 → 128 → 10

---

Quick Start

1 – Install dependencies

python3 -m venv .venv
source .venv/bin/activate
pip3 install -r requirements.txt

2 – Train

python3 train_numpy.py

This will:

1. Download Fashion MNIST automatically
2. Run 5-fold cross-validation on 60,000 training samples (20 epochs per fold)
3. Retrain on the full training set (50 epochs)
4. Report accuracy on the 10,000-sample test set

---

Hyperparameters

All hyperparameters are set inside train_numpy.py in the build_model() function:

Parameter	Value
Hidden layers	784 → 512 → 512 → 256 → 128 → 10
L2 regularization	5e-4
Dropout rate	0.05
Learning rate	0.0005
LR decay	1e-5
Adam β₁ / β₂	0.9 / 0.999
Batch size	128
CV folds	5

---

Results

Metric	Value
Test Accuracy	~89.5%
Optimizer	Adam with LR decay
Validation strategy	5-fold cross-validation

---

Fashion MNIST Classes

Label	Class
0	T-shirt/top
1	Trouser
2	Pullover
3	Dress
4	Coat
5	Sandal
6	Shirt
7	Sneaker
8	Bag
9	Ankle boot

---

Stack

Concern	Tool
Numerics / forward & backward pass	NumPy
Dataset download	torchvision (utility only)
Python version	3.11+

---

License

MIT