How schobs/medimarker works
Project Type and Purpose
• The project is a research/production pipeline for anatomical landmark localization (using models such as U-Net, PHD-Net, and Convolutional Gaussian Processes). It supports uncertainty estimation, ensembling, patch‐based training, and a comprehensive configuration system using .yaml files.
File Structure
• There are clear modular divisions by functionality:
• Configuration: Directories such as “configs” (with many subdirectories for dev, GP, legacy, and local) indicate a strong emphasis on easy customization via YAML.
• Data Management and Preprocessing: The “datasets”, “preprocessing”, and “transforms” directories handle data ingestion, JSON generation, and image transformation.
• Model Implementations: Under “models” you find architectures for U-Net, PHD-Net, and GP models grouped in subdirectories.
• Training and Inference: “trainer” contains various training routines for different model types while “inference” holds scripts for inference tasks.
• Evaluation: The “evaluation” folder holds scripts for analyzing model outputs.
• Utilities and Logging: “utils” and “losses” support logging, TensorFlow helper scripts, and loss definitions.
• Deployment/Scripting: “scripts” contains various HPC and training scripts.
• The pattern is clearly modular with a separation between configuration, data, model definitions, training/inference, evaluation, and utilities.