How yongkyung-oh/taod-net works
Project type and purpose
• This is an ML research codebase (PyTorch) for TAOD-Net, a transformer-based multi-label image classification system that detects key inflammatory signs of Thyroid-Associated Orbitopathy (TAO) in facial imagery.
• It leverages domain-knowledge cropping, a two-stage training strategy (pretraining + fine-tuning), and custom loss functions to standardize TAO diagnosis.
File structure
• Top level:
• assets/ – static images (e.g., “overview_new.png”) for documentation
• experiments/– CLI entrypoints for experiments
– binary.py (baseline multi-label classification)
– pretraining.py (pretraining phase)
– model_run.py (combined, multi, and binary training & evaluation)
• src/ – core library modules
– utils.py (data I/O, preprocessing, cropping)
– model.py (TAOD-Net architecture based on ViT)
– losses.py (custom multi-label loss definitions)
• Configuration and metadata: .gitignore, LICENSE, README.md
README for additional insights
• Confirms two-stage training and domain-knowledge cropping
• Specifies that the core model is a Vision Transformer (ViT) backbone with custom heads
• Describes data (2,040 anonymized images) and evaluation modalities