How rmpku/cin works
Project type and purpose
• This is a research‐oriented deep‐learning codebase (“CIN”): it implements an invertible neural network combined with non-invertible modules to perform blind watermarking or robust image restoration under noise.
• The README references datasets (COCO2017, DIV2K), a noise pool, pre-trained models, and training details.
File structure
• Top level:
· codes/ ← all Python code
· pth/ ← pretrained model weights
· README.md, .gitignore
• Under codes/:
· main.py ← entry point that parses options, sets up data, runs train/test
· models/ ← model definition package
– CIN.py, NSM_module.py, DEM.py, FSM.py, IM.py, Network.py, Noise_pool.py
– modules/ ← building blocks: Conv1x1, DenseBlock, the invertible architecture, noise‐injection options, loss functions
· options/ ← configuration loader & YAML file (opt.yml)
· utils/ ← checkpointing, general utilities, YAML parsing
README for additional insights
• Architecture overview screenshots under “Framework”, “Visualization”, “invertibleNet”
• Describes noise pool, training details, environment dependencies
Elements
• Blue = configuration & I/O
• Green = data processing
• Orange = model internals (invertible vs non-invertible blocks)
• Gray = utilities (checkpoint, logging)
Detailed capture of essential architectural elements
• Do not simply lump everything into one block; separate:
· opt.yml reader
· dataset ingestion
· noise‐pool (list the noise options submodules)
· invertible architecture (InvArch, InvDownscaling)
· non-invertible branch (NSM_module, DEM, FSM, IM)
· loss function calculations
· optimizer / training loop
· checkpoint save/load