CASE STUDY 09
Fashion-MNIST Classification
PythonPyTorchDeep LearningCNNVision Transformer
Role
ML Engineer
Year
2024
Category
Machine Learning
Context
Academic Project
01.
The Problem
Understanding the trade-offs between different machine learning approaches is crucial. This project aimed to rigorously benchmark simple MLPs, CNNs, and Vision Transformers on a standard dataset to visualize performance vs. complexity.
I implemented a modular PyTorch pipeline to train and evaluate multiple architectures: PCA + MLP, standard CNNs, and patch-based Vision Transformers. I used consistent validation/test splits and metrics (Accuracy, Macro F1) to ensure fair comparison.
02.
The Approach
03. The Result
The study highlighted that while Transformers are powerful, CNNs remain highly efficient for specific image scales. The project serves as a solid flexible codebase for future computer vision experiments.