How c-yn/adair works
AdaIR is an all-in-one image-restoration research codebase: it trains one adaptive model across denoising, deraining, dehazing, deblurring, and enhancement, and provides task-specific and multi-degradation evaluation. The main documented workflow is configured training data → degraded/clean patches → model restoration and loss optimization; evaluation uses a selected task mode and checkpoint. The model’s frequency-mining, modulation, and progressive-restoration stages are documented in the README. The evaluation implementation and some model internals were not sampled, so those links reflect README statements rather than inspected calls.
Main entry points: train.py Training loop · dataset_utils.py Training dataset · test.py Task evaluation · options.py Run options
Components 16
Configuration
- Run options
options.py
Data preparation
- Training dataset
utils/dataset_utils.py - Degradation synthesis
utils/degradation_utils.py - Image transforms
utils/image_utils.py
Restoration model
- Feature encoder
net/model.py - Frequency mining
net/model.py - Bidirectional modulation
net/model.py - Progressive restoration
net/model.py
Training and evaluation
- Training loop
train.py - Learning-rate schedule
utils/schedulers.py - Task evaluation
test.py
Other components
- Researcher
- Training images
- Model checkpoint
- Restored image
- Evaluation images
Connections 19
- Researcher → Run options: sets options
options.py - Researcher → Training images: prepares data
README.md - Run options → Training loop: configures
train.py - Training loop → Training dataset: constructs dataset
train.py - Training dataset → Training images: reads samples
- Training dataset → Degradation synthesis: uses
utils/dataset_utils.py - Training dataset → Image transforms: uses transforms
utils/dataset_utils.py - Training loop → Feature encoder: invokes model
train.py - Training loop → Learning-rate schedule: configures schedule
train.py - Feature encoder → Frequency mining: extracts features
README.md - Frequency mining → Bidirectional modulation: feeds modulation
README.md - Bidirectional modulation → Progressive restoration: guides restoration
README.md - Progressive restoration → Restored image: produces
README.md - Training loop → Model checkpoint: saves checkpoints
train.py - Researcher → Evaluation images: prepares inputs
README.md - Researcher → Task evaluation: selects task mode
README.md - Task evaluation → Model checkpoint: loads checkpoint
README.md - Task evaluation → Evaluation images: evaluates images
README.md - Task evaluation → Restored image: restores images