Fashion-MNIST CNN Classifier

A Keras/TensorFlow convolutional neural network for classifying Fashion-MNIST apparel images.

What I Built

I developed a CNN architecture that achieves 91.44% test accuracy on the Fashion-MNIST dataset. The model features:

• Architecture: 3 convolutional blocks (32→64→128 filters) with MaxPooling2D, Dropout regularization (0.2-0.25), and a dense classifier (512 units)
• Training: 5-fold cross-validation with SGD optimizer (lr=0.01, momentum=0.9), ModelCheckpoint for best weights, and EarlyStopping
• Monitoring: TensorBoard integration for real-time loss/accuracy visualization
• Persistence: Model saved as .h5 format with best weights from validation

Dataset

Fashion-MNIST - 70,000 grayscale images (28×28 pixels) across 10 apparel classes:

• T-shirt/top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot

Results

• Test Accuracy: 91.44%
• Cross-validation Mean: 91.94% ± 0.32%
• Training: 10 epochs per fold with batch size 32

How to Run

Notebook

pip install -r requirements.txt
jupyter notebook fashion_mnist_cnn.ipynb

Docker

Prerequisites

First, train and export the model:

pip install -r requirements.txt
python train_model.py

This creates models/on_the_go/fashion_mnist_api_final.keras required for serving.

Build and Run

# Build and start all services
docker-compose up --build

# Scale web containers (e.g., 3 replicas)
docker-compose up --scale web=3

# Scale down
docker-compose up --scale web=1

Test Endpoints

# Health check
curl http://localhost/health

# Predict with image upload
curl -X POST http://localhost/predict -F "image=@test_image.png"

# Predict with pixel array (JSON)
curl -X POST http://localhost/predict_array \
  -H "Content-Type: application/json" \
  -d '{"pixels": [0,0,0,...]}'

# View API documentation
open http://localhost/docs

TensorBoard

Access training visualizations at http://localhost:6006

Tech Stack

• TensorFlow 2.x / Keras
• NumPy, Pandas, Matplotlib, Seaborn
• Scikit-learn (cross-validation)
• FastAPI (serving API)
• Docker & Docker Compose
• Nginx (load balancer)