Goal: To gain hands-on experience in designing, training, and evaluating deep convolutional models for visual recognition tasks.
This project involved building a deep convolutional neural network (CNN) to classify images from the FashionMNIST dataset, which contains grayscale images of clothing items across 10 categories (e.g., T-shirts, shoes, bags).
Model Architecture:
- Two convolutional layers with ReLU and max pooling
- Two fully connected layers
- Output layer with 10 units (softmax)
- Xavier uniform initialisation
- Trained with SGD optimizer, learning rate 0.1 for 30 epochs
Evaluation:
- Achieved strong performance on both training and validation sets
- Plotted accuracy and loss curves per epoch
- Analysed training speed and overfitting behaviour
Tech stack: Python, PyTorch, torchvision, matplotlib