How ruoyuguo/mda-net works
Project type and purpose
• This is a research‐oriented PyTorch implementation of the “Multi‐degradation‐adaptation Network (MDA‐Net)” for fundus image enhancement (Medical Image Analysis, 2024).
• It provides scripts to generate/degrade datasets, define/train deep networks, and evaluate saved models on benchmark data.
File structure
• Top-level files and scripts:
• train.py / train_from_scratch.sh: orchestrate model training and checkpointing
• eval.py: loads a saved checkpoint and runs inference/evaluation
• dataset/ directory:
• Raw CSVs for labels (Label_EyeQ_*.csv, MIL_*.csv)
• Data-generation and degradation scripts (gen_data_stage1/2, degrad_de.py, degrad_de_visual.py)
• Utility modules for parameter loading and augmentation
• network/ directory:
• Model definitions (DANet.py, DCouple.py, kpn_network.py) and their helper functions
• clf.py: classification or feature branch
• utils/ directory:
• Dataset wrappers (drive_dataset.py, eyeq_dataset.py, refuge_dataset.py, etc.)
• Logging, metric calculations, loss functions
Extract additional insights from README
• Training can be launched via a single shell script
• eval.py supports selecting any saved checkpoint by index/epoch and specifying dataset paths
• Outputs (enhanced images, metrics) are written under ./myoutput/<run_id>/