How awsaf49/covxnet works
Project type and purpose
• This is a Python‑based machine‑learning library (packaged as covxnet) for building, training, and deploying the CovXNet CNN architecture to detect COVID‑19 and other pneumonia from chest X‑ray images.
• It provides a set of reusable “blocks” (stem, shifter, residual) and a model builder in covxnet/network/model.py, plus a stacking ensemble and Grad‑CAM integration for localization.
File structure
Top level:
• setup.py, requirements.txt, LICENSE, README.md – packaging, dependencies, documentation
• .github/workflows/publish_to_pypi.yml – CI/CD for building and publishing the package to PyPI
• covxnet/
‑ version.py – package version
‑ __init__.py – top‑level importer
‑ blocks/ – primitive convolutional units (stem, residual, shifter)
‑ network/model.py – CovXNet class definitions (CovXNet128, 224, 320, etc.), training & inference logic
• images/ – architecture sketches for the paper & README
README for architecture insights
• It calls out two phases:
a. Base training on normal/pneumonia X‑rays
b. Transfer learning & fine‑tuning on limited COVID‑19 data
• A stacking algorithm and a gradient‑based localization (Grad‑CAM) layer are applied post‑training.
A. Data Pipeline
• External chest X‑ray datasets (Normal, Pneumonia, COVID‑19)
• Preprocessing (resizing to 128/224/320, normalization)
B. Model Definition Layer
• Blocks module (stem, residual, shifter)
C. Training Pipeline
• Phase 1: Train base CovXNet on large Normal+Pneumonia set
• Phase 2: Transfer & fine‑tune on small COVID‑19 dataset
D. Ensemble & Optimization
• Multiple CovXNet variants (128, 224, 320) → stacking ensemble
E. Localization & Explanation
• Grad‑CAM integration to highlight ROI in X‑rays
F. Packaging & Deployment
• setup.py / requirements.txt for pip installation
• CI/CD workflow (.github/workflows) for publish_to_pypi
Data Pipeline → Model Definition → Training → Ensemble → Inference & Grad‑CAM → Client code / Notebook.
Architectural patterns & design principles
• Modular library design: separation of low‑level blocks vs. high‑level model logic
• Composition over inheritance: CovXNet assembles block objects
• Transfer learning pattern: base model + fine‑tuning layers
• Ensemble learning: stacking multiple model variants
• CI/CD automation for packaging & release
Technologies & external services
• Core: Python, TensorFlow / Keras (implicit from model code)
• Packaging: setuptools, pip
• CI/CD: GitHub Actions, PyPI repository