Overview
A systematic comparison of convolutional architectures on the COVID-19 Radiography dataset: two custom CNNs, a BasicBlock-based residual CNN, ResNet50 from scratch and a pretrained fine-tuned ResNet50.
What I built
A custom dataset loader with batch visualization, five switchable model configurations, early stopping, and run logs with plots. The training CLI estimates VRAM usage up front (training 13.62 GB / validation 4.54 GB in the reference configuration) and can pin datasets into GPU memory — treating memory as a first-class engineering constraint rather than letting runs fail late.
The same repository carries logistic-regression baselines, including a training-progress GIF, forming a complete progression from linear baselines to deep networks.
Technical highlights
- Scratch and pretrained ResNet50 paths against custom CNNs
- Explicit VRAM budgeting and GPU-resident datasets
- Reproducible run logs and plots per configuration
Outcome
A complete comparison study with documented results and full reproducibility.