How junlinhan/cwr works
Project type and purpose
• This is a research‐oriented PyTorch project implementing “Contrastive UnderWater Restoration (CWR),” a deep learning pipeline for unpaired (and paired) underwater image restoration and related low‐level vision tasks (dehazing, deraining).
• It provides training and testing scripts, dataset loaders, model definitions (generator + contrastive losses), utilities (visualization, HTML logging), and an experiments entry point.
File structure
• Top-level scripts:
• train.py – launches the training loop
• test.py – runs inference and saves results
• experiments/__main__.py – alternative launcher for experiments
• Modules by responsibility:
• options/ – CLI argument parsing (base, train, test)
• data/ – abstract dataset classes and concrete unaligned/paired datasets
• datasets/ – local data storage (HICRD folder with train/test splits)
• models/ – model definitions (networks.py, spectralNormalization, PatchNCE, StyleGAN networks, base & template models)
• util/ – helpers for data loading, HTML generation, image pooling, logging, visualization via Visdom
• Configuration & environment: requirements.txt, environment.yml
• Checkpoints & results conventions: saved under ./checkpoints/<name> and ./results/<name>
README for additional insights
• Confirms architecture: unpaired image‐to‐image translation with contrastive loss (built upon CUT and CycleGAN codebases)
• Training uses Visdom server for real‐time loss curves and HTML pages for visual outputs
• External dependencies: PyTorch, torchvision, visdom, dominate, gputil
1. CLI Layer
• train.py / test.py / experiments
• Options Parser (options/base_options.py, train_options.py, test_options.py)
2. Data Layer
• Dataset Factory (data/__init__.py)
• Dataset Classes (unaligned_dataset, single_dataset, image_folder, etc.)
• Local Storage (datasets/HICRD/*)
• DataLoader (PyTorch DataLoader via util/get_data.py)
3. Model Layer
• Network Architectures (models/networks.py, stylegan_networks.py)
• Contrastive Modules (models/patchnce.py, spectralNormalization.py)
• Model Orchestration (models/base_model.py, template_model.py, cwr_model.py)
4. Training Engine
• Training Loop (train.py)
• Loss Computation (contrastive, adversarial, cycle, identity etc.)
• Optimizers & Schedulers
• Checkpoint Manager (saving under ./checkpoints)
5. Testing/Inference Engine
• Inference Loop (test.py)
• Image Pre/post‐processing (util/util.py, util/image_pool.py)
• Result Export (HTML pages via util/html.py)
6. Visualization & Logging
• Visdom Server (util/visualizer.py)
• HTML Logger (util/html.py)
7. External Systems
• Local Filesystem (datasets, checkpoints, results)
• Visdom Server (http://localhost:8097)
b. Relationships and interactions
• CLI scripts → parse options → instantiate Data Layer & Model Layer
• Data Layer → yields batches → Training Engine / Testing Engine
• Training Engine ↔ Model Layer (forward/backward passes)
• Training Engine → loss/output → Visualization & Logging
• Training Engine → checkpoint saving to filesystem
• Testing Engine → model.load() → inference → HTML export
c. Architectural patterns & design principles
• Modular separation of concerns: options, data, models, utils
• Factory pattern for dataset instantiation
• Template Method / Inheritance for Model classes (base_model → cwr_model, template_model)
• Pipeline pattern: data → model → loss → optimizer → log → checkpoint
d. Key technologies/frameworks
• PyTorch & torchvision for model and DataLoader
• Visdom + dominate for real‐time and static visualization
• Conda / pip for environment reproducibility
• Based on pytorch-CycleGAN-and-pix2pix and CUT implementations
Final Advice
• Essential modules (options, data classes, model classes, utils).
• Data flow and dependency direction clearly.