🚀 Add PreProcessor to AnomalyModule

[samet-akcay]

📝 Description

The PreProcessor class serves as both a PyTorch module and a Lightning callback, handling transforms during different stages of training, validation, testing and prediction. This PR demonstrates how to create and use custom pre-processors.

Key Components

The pre-processor functionality is implemented in:

class PreProcessor(nn.Module, Callback):
    """Anomalib pre-processor.

    This class serves as both a PyTorch module and a Lightning callback, handling
    the application of transforms to data batches during different stages of
    training, validation, testing, and prediction.

    Args:
        train_transform (Transform | None): Transform to apply during training.
        val_transform (Transform | None): Transform to apply during validation.
        test_transform (Transform | None): Transform to apply during testing.
        transform (Transform | None): General transform to apply if stage-specific
            transforms are not provided.

    Raises:
        ValueError: If both `transform` and any of the stage-specific transforms
            are provided simultaneously.

    Notes:
        If only `transform` is provided, it will be used for all stages (train, val, test).

        Priority of transforms:
        1. Explicitly set PreProcessor transforms (highest priority)
        2. Datamodule transforms (if PreProcessor has no transforms)
        3. Dataloader transforms (if neither PreProcessor nor datamodule have transforms)
        4. Default transforms (lowest priority)

    Examples:
        >>> from torchvision.transforms.v2 import Compose, Resize, ToTensor
        >>> from anomalib.pre_processing import PreProcessor

        >>> # Define transforms
        >>> train_transform = Compose([Resize((224, 224)), ToTensor()])
        >>> val_transform = Compose([Resize((256, 256)), CenterCrop((224, 224)), ToTensor()])

        >>> # Create PreProcessor with stage-specific transforms
        >>> pre_processor = PreProcessor(
        ...     train_transform=train_transform,
        ...     val_transform=val_transform
        ... )

        >>> # Create PreProcessor with a single transform for all stages
        >>> common_transform = Compose([Resize((224, 224)), ToTensor()])
        >>> pre_processor_common = PreProcessor(transform=common_transform)

        >>> # Use in a Lightning module
        >>> class MyModel(LightningModule):
        ...     def __init__(self):
        ...         super().__init__()
        ...         self.pre_processor = PreProcessor(...)
        ...
        ...     def configure_callbacks(self):
        ...         return [self.pre_processor]
        ...
        ...     def training_step(self, batch, batch_idx):
        ...         # The pre_processor will automatically apply the correct transform
        ...         processed_batch = self.pre_processor(batch)
        ...         # Rest of the training step
    """

And used by the base AnomalyModule in:

    def _resolve_pre_processor(self, pre_processor: PreProcessor | bool) -> PreProcessor:
        """Resolve and validate which pre-processor to use..

        Args:
            pre_processor: Pre-processor configuration
                - True -> use default pre-processor
                - False -> no pre-processor
                - PreProcessor -> use the provided pre-processor

        Returns:
            Configured pre-processor
        """
        if isinstance(pre_processor, PreProcessor):
            return pre_processor
        if isinstance(pre_processor, bool):
            return self.configure_pre_processor()
        msg = f"Invalid pre-processor type: {type(pre_processor)}"
        raise TypeError(msg)

Usage Examples

1. Using Default Pre-Processor

The simplest way is to use the default pre-processor which resizes images to 256x256 and normalizes using ImageNet statistics:

from anomalib.models import PatchCore

# Uses default pre-processor
model = PatchCore()

2. Custom Pre-Processor with Different Transforms

Create a pre-processor with custom transforms for different stages:

from torchvision.transforms.v2 import Compose, Resize, CenterCrop, RandomHorizontalFlip, Normalize
from anomalib.pre_processing import PreProcessor

# Define stage-specific transforms
train_transform = Compose([
    Resize((256, 256), antialias=True),
    RandomHorizontalFlip(p=0.5),
    Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

val_transform = Compose([
    Resize((256, 256), antialias=True), 
    CenterCrop((224, 224)),
    Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

# Create pre-processor with different transforms per stage
pre_processor = PreProcessor(
    train_transform=train_transform,
    val_transform=val_transform,
    test_transform=val_transform  # Use same transform as validation for testing
)

# Use custom pre-processor in model
model = PatchCore(pre_processor=pre_processor)

3. Disable Pre-Processing

To disable pre-processing entirely:

model = PatchCore(pre_processor=False)

4. Override Default Pre-Processor in Custom Model

Custom models can override the default pre-processor configuration:

from anomalib.models.components.base import AnomalyModule

class CustomModel(AnomalyModule):
    @classmethod
    def configure_pre_processor(cls, image_size=(224, 224)) -> PreProcessor:
        transform = Compose([
            Resize(image_size, antialias=True),
            Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
        ])
        return PreProcessor(transform=transform)

Notes

• Pre-processor transforms are applied in order of priority:
  • Explicitly set PreProcessor transforms (highest)
  • Datamodule transforms
  • Dataloader transforms
  • Default transforms (lowest)
• The pre-processor automatically handles both image and mask transforms during training
• Custom transforms should maintain compatibility with both image and segmentation mask inputs

Testing

• Added unit tests to verify:
• Default pre-processor behavior
• Custom transform application
• Transform priority order
• Mask transformation handling

✨ Changes

Select what type of change your PR is:

• 🐞 Bug fix (non-breaking change which fixes an issue)
• 🔨 Refactor (non-breaking change which refactors the code base)
• 🚀 New feature (non-breaking change which adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📚 Documentation update
• 🔒 Security update

✅ Checklist

Before you submit your pull request, please make sure you have completed the following steps:

• 📋 I have summarized my changes in the CHANGELOG and followed the guidelines for my type of change (skip for minor changes, documentation updates, and test enhancements).
• 📚 I have made the necessary updates to the documentation (if applicable).
• 🧪 I have written tests that support my changes and prove that my fix is effective or my feature works (if applicable).

For more information about code review checklists, see the Code Review Checklist.

[jpcbertoldo]

A sub-feature request that would fit here: (optionally?) keep both the transformed and original image/mask in the batch.

So instead of

            image, gt_mask = self.XXX_transform(batch.image, batch.gt_mask)
            batch.update(image=image, gt_mask=gt_mask)

something like

            batch.update(image_original=batch.image, gt_mask_original=batch.gt_mask)
            image, gt_mask = self.XXX_transform(batch.image, batch.gt_mask)
            batch.update(image=image, gt_mask=gt_mask)

It's quite practical to have these when using the API (i've re-implemented this in my local copy 100 times haha).

[samet-akcay]

A sub-feature request that would fit here: (optionally?) keep both the transformed and original image/mask in the batch.

So instead of

            image, gt_mask = self.XXX_transform(batch.image, batch.gt_mask)
            batch.update(image=image, gt_mask=gt_mask)

something like

            batch.update(image_original=batch.image, gt_mask_original=batch.gt_mask)
            image, gt_mask = self.XXX_transform(batch.image, batch.gt_mask)
            batch.update(image=image, gt_mask=gt_mask)

It's quite practical to have these when using the API (i've re-implemented this in my local copy 100 times haha).

yeah, the idea is to keep batch.image and batch.gt_mask original outside the model. It is not working that way though :)

[jpcbertoldo]

yeah, the idea is to keep batch.image and batch.gt_mask original outside the model

exactly, makes sense : )

but it's also useful to be able to access the transformed one (eg. when using augmentations)

it is not working that way though :)

didnt get this. cause it's not backcompatible?

[samet-akcay]

yeah, the idea is to keep batch.image and batch.gt_mask original outside the model

exactly, makes sense : )

but it's also useful to be able to access the transformed one (eg. when using augmentations)

it is not working that way though :)

didnt get this. cause it's not backcompatible?

oh I meant, it is currently not working, I need to fix it :)

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[ashwinvaidya17]

Thanks. I have a few minor comments

[ashwinvaidya17]

Should we add a check to ascertain whether train_transform is present? Models like VlmAD might not have train_transforms passed to them. I feel it should pick up val or pred transform is train is not available.

[djdameln]

Thanks! I think we could merge this