Fashion Image Classification with Neural Networks

This project is a machine learning task focused on classifying grayscale images (32×32 pixels) derived from the Fashion-MNIST dataset. It was developed as part of a university assignment to explore and compare neural network architectures for image classification.

🗂️ Dataset

• train.csv — contains labeled training data.
  • Features: flattened grayscale images (1024 pixels).
  • Target: label column (categorical classes).
• evaluate.csv — contains unlabeled test data for final prediction.

🎯 Project Objectives

1. Data Exploration

   • Load and inspect the structure of the dataset.
   • Visualize sample images.
   • Analyze label distribution and pixel value characteristics.

2. Model Development

   • Build and train feedforward neural networks (FNNs):

     • Vary number of layers and neurons.
     • Experiment with normalization, optimizers (SGD, Adam), and regularization (Dropout, L2).
     • Compare performance and document findings.

   • Build and train convolutional neural networks (CNNs):

     • Use multiple convolutional and pooling layers.
     • Apply batch normalization, Dropout, and optimizer tuning.
     • Assess suitability for image data and evaluate test accuracy.

3. Model Evaluation & Selection

   • Use validation subsets to assess model generalization.
   • Select a final model based on accuracy and stability.

4. Prediction on Test Set

   • Use the final trained model to predict labels in evaluate.csv.
   • Save predictions in the format below and export as results.csv:

🧪 Tools & Libraries

• Language: Python
• Libraries: NumPy, Pandas, Matplotlib, torch, scikit-learn, seaborn, torchvision

📈 Outcomes

• Comparison of different model architectures and training strategies.
• Understanding of CNNs for image recognition.
• Final test set predictions stored in results.csv.

---

This project showcases practical skills in deep learning, model tuning, and real-world image classification workflows.