How lime-j/rdnet works
Project type and purpose
• This is a research‐oriented deep‐learning pipeline (“Reversible Decoupling Network”) implemented in PyTorch. Its goal is single‐image reflection removal: given an image with unwanted reflections, the network decouples transmission and reflection features and reconstructs a clean image.
File structure
• Top‐level scripts:
• train.py – orchestrates training
• test_sirs.py – runs evaluation with pretrained checkpoints
• engine.py – core training/validation loops and scheduling
• script.py – auxiliary or legacy entrypoint
• data/ – dataset definitions and preprocessing
• models/ – model definitions
• arch/ – building blocks (encoders, decoders, reversible units, prompts)
• base_model.py, networks.py, losses.py – abstractions for training
• options/ – CLI and configuration (argparsing, omegaconf)
• tools/ – logging, visualization, TensorBoard writer, model utilities
• util/ – HTML reporting, scheduler helpers, general utilities
• pretrained/ – place for downloaded checkpoints
• figures/ – static assets (for README)
Insights from README
• Describes two main workflows: TRAIN (train.py) and TEST (test_sirs.py)
• Dependencies: PyTorch ≥2.0, einops, timm, tensorboard, etc.
• Emphasis on reversible encoder, prompt generator, and use of TPU/Cloud.
Essential details to capture
• Option parsing → component instantiation
• Data module feeding into reversible encoder + prompt generator
• Loss computation and optimizer scheduling
• Checkpoint read/write and evaluation path
• Logging hooks (TensorBoard, HTML reports)