^j dZdddiiZddidddddZddiddiddd d d d d dddddddddZddiddiddd dddddZddidddddddddddZy)u Augmentation presets and default configuration for RF-DETR training. Import a preset and pass it as ``aug_config`` to your training call: ```python from rfdetr.datasets.aug_configs import AUG_CONSERVATIVE, AUG_AGGRESSIVE, AUG_AERIAL, AUG_INDUSTRIAL model.train(dataset_dir="...", aug_config=AUG_CONSERVATIVE) model.train(dataset_dir="...", aug_config=AUG_AGGRESSIVE) # Disable all augmentations model.train(dataset_dir="...", aug_config={}) # Fully custom model.train(dataset_dir="...", aug_config={"HorizontalFlip": {"p": 0.5}}) ``` ## Available presets | Preset | Best for | | -------------- | ------------------------------------------------ | | ``AUG_CONSERVATIVE`` | Small datasets (under 500 images) | | ``AUG_AGGRESSIVE`` | Large datasets (2000+ images) | | ``AUG_AERIAL`` | Satellite / overhead imagery | | ``AUG_INDUSTRIAL`` | Manufacturing / inspection data | ## Transform Categories **Geometric transforms** (automatically transform bounding boxes): - Flips: HorizontalFlip, VerticalFlip - Rotations: Rotate, Affine, ShiftScaleRotate - Crops: RandomCrop, CenterCrop, RandomResizedCrop - Perspective: Perspective, ElasticTransform, GridDistortion **Pixel-level transforms** (preserve bounding boxes): - Color: ColorJitter, HueSaturationValue, RandomBrightnessContrast - Blur/Noise: GaussianBlur, GaussNoise, Blur - Enhancement: CLAHE, Sharpen, Equalize ## Best Practices 1. **Start conservative**: Use moderate probabilities (p=0.3-0.5) and small parameter ranges 2. **Geometric caution**: Extreme rotations (>45°) or crops may remove too many boxes 3. **Performance**: Fewer transforms = faster training; prioritize transforms that match your domain 4. **Validation**: Monitor validation mAP - excessive augmentation can hurt performance 5. **Domain-specific**: Enable augmentations that reflect real-world variations in your data ## Adding Custom Transforms For geometric transforms not in GEOMETRIC_TRANSFORMS set, add them in transforms.py: ```python GEOMETRIC_TRANSFORMS = { ... "YourCustomTransform", # Add here } ``` ## Kornia GPU Backend When ``augmentation_backend="auto"`` or ``"gpu"`` is set in ``TrainConfig``, augmentations run on the GPU via Kornia instead of Albumentations. **Supported transforms** (all presets): | Preset key | Kornia equivalent | Notes | |---|---|---| | ``HorizontalFlip`` | ``K.RandomHorizontalFlip`` | Direct | | ``VerticalFlip`` | ``K.RandomVerticalFlip`` | Direct | | ``Rotate`` | ``K.RandomRotation`` | ``limit`` may be scalar or tuple | | ``Affine`` | ``K.RandomAffine`` | ``translate_percent`` treated as fraction | | ``ColorJitter`` | ``K.ColorJiggle`` | Same multiplicative semantics | | ``RandomBrightnessContrast`` | ``K.ColorJiggle`` | ``brightness_limit`` / ``contrast_limit`` direct | | ``GaussianBlur`` | ``K.RandomGaussianBlur`` | ``blur_limit`` rounded up to odd; ``sigma=(0.1, 2.0)`` | | ``GaussNoise`` | ``K.RandomGaussianNoise`` | Upper bound of ``std_range`` used as fixed std | **Phase 1 limitation**: Segmentation models (``segmentation_head=True``) skip GPU augmentation; CPU Albumentations are used instead. Mask support is planned for Phase 2. HorizontalFlippg?皙?g333333?)brightness_limitcontrast_limitr)rRandomBrightnessContrast-)limitr)g?g333333?)gr)i))scaletranslate_percentrotateshearrg?) brightnesscontrast saturationhuer)r VerticalFlipRotateAffine ColorJitter)Zrg333333?g?)rrrr) blur_limitr)g{Gz?g?) std_ranger)rr GaussianBlur GaussNoiseN)__doc__ AUG_CONFIGAUG_CONSERVATIVEAUG_AGGRESSIVE AUG_AERIALAUG_INDUSTRIALf/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/datasets/aug_configs.pyr(s MhsCj Cj !Cj#J%(    ,Cj#J s+  !  Cj ! $%3/ ,37 r&