MA/01

Deep Learning / Case study

Fashion-MNIST CNN in PyTorch

Two PyTorch CNN variants for Fashion-MNIST, including a batch-normalized architecture and tracked validation performance.

View repository

Problem

Build and compare convolutional baselines for ten-class grayscale fashion-image classification.

Solution

Implemented standard and batch-normalized CNNs with torchvision preprocessing, SGD optimization, and per-epoch cost and validation tracking.

Architecture

28×28 image → convolution → batch normalization → max pooling → convolution → batch normalization → max pooling → 10-class linear head.

Technical implementation

Uses Fashion-MNIST’s 70,000 grayscale images across ten clothing categories as part of IBM deep-learning coursework.

Key challenges

Compare a baseline CNN with a batch-normalized variant through a consistent training loop.

Lessons learned

The project makes architectural changes and their training behavior directly inspectable.