How runner-xc/rdamu-net works
Project type and purpose
• This is a research-oriented deep-learning image segmentation toolkit focused on shale microstructure segmentation. It provides multiple U-Net-style and related architectures (including the novel RDAMU-Net), end-to-end training, evaluation, logging, and slide-based prediction pipelines.
File structure
• Top-level “train.py” and “predict.py” scripts: entry points for training and inference.
• “model/”: definitions of all segmentation networks (SegNet, U-Net variants, DeepLabV3, PSPNet, RDAMU-Net).
• “model/utils/”: building-block modules (attention, backbones, utility layers).
• “utils/”: data preparation, augmentation (“transforms.py”), dataset splitting, custom losses, metrics, logging, TensorBoard integration, experiment-management helpers, slide-based inference.
• “cam.py”: class-activation mapping for model interpretability.
• Shell script “train_models.sh” for batch training.
Key insights from README
• Project is invoked via CLI arguments to “train.py” (data paths, model choice).
• Data structure: raw images + masks + CSV index files.
• Core contribution is the RDAMU-Net architecture, combining dilated residuals, multi-scale fusion, and multi-dimensional attention.