Feature/rd++

[cjy513203427]

📝 Description

• Provide a clear summary of the changes and the issue that has been addressed.
• 🛠️ Fixes # (issue number)

✨ 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:

• [ x ] 📋 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).
• [ x ] 🧪 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.

I re-implemented part of the RD++ algorithm. It is based on reverse distillation. I added self-supervised optimal transport loss, reconstruction loss, contrast loss and Multiscale projection layers based on📄 Paper
🧑‍💻 Code

Some things are missing: The authors use a customised dataloader and noise.py for MVTEC dataset. However, I don't find any noise definition in Anomalib. Perlin noise is something else. Should I add a noise function in mvtec.py which could affect other functions or create a customised dataloader for RD++?

[cjy513203427]

This is test code (I used the recommended epochs from the authors.):

import logging
from anomalib import TaskType
from anomalib.data import MVTec
from anomalib.engine import Engine
from anomalib.models import RevisitingReverseDistillation
from lightning.pytorch.callbacks import EarlyStopping, ModelCheckpoint

# configure logger
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# Define the number of epochs for each category
epoch_mapping = {
    'carpet': 10,
    'leather': 10,
    'grid': 260,
    'tile': 260,
    'wood': 100,
    'cable': 240,
    'capsule': 300,
    'hazelnut': 160,
    'metal_nut': 160,
    'screw': 280,
    'toothbrush': 280,
    'transistor': 300,
    'zipper': 300,
    'pill': 200,
    'bottle': 200,
}

# datasets = ['screw', 'pill', 'capsule', 'carpet', 'grid', 'tile', 'wood', 'zipper', 'cable', 'toothbrush', 'transistor',
#             'metal_nut', 'bottle', 'hazelnut', 'leather']
# datasets = ['carpet']
datasets = ['bottle', 'hazelnut', 'leather']

for dataset in datasets:
    logger.info(f"================== Processing dataset: {dataset} ==================")
    task = TaskType.SEGMENTATION
    datamodule = MVTec(
        root="../datasets/MVTec",
        category=dataset,
        image_size=256,
        train_batch_size=32,
        eval_batch_size=32,
        num_workers=0,
        task=task,
    )

    model = RevisitingReverseDistillation()

    callbacks = [
        ModelCheckpoint(
            mode="max",
            monitor="pixel_AUROC",
        ),
        EarlyStopping(
            monitor="pixel_AUROC",
            mode="max",
            patience=3,
        ),
    ]

    # Get the number of epochs for the current dataset
    num_epochs = epoch_mapping.get(dataset, 100)  # Default to 100 if not found

    logger.info(f"Using {num_epochs} epochs for dataset: {dataset}")

    engine = Engine(
        max_epochs=num_epochs,
        check_val_every_n_epoch=3,
        callbacks=callbacks,
        pixel_metrics=["AUROC", "PRO"], image_metrics=["AUROC", "PRO"],
        accelerator="auto",  # <"cpu", "gpu", "tpu", "ipu", "hpu", "auto">,
        devices=1,
        logger=False,
    )

    logger.info(f"================== Start training for dataset: {dataset} ==================")
    engine.fit(datamodule=datamodule, model=model)

    logger.info(f"================== Start testing for dataset: {dataset} ==================")
    engine.test(datamodule=datamodule, model=model)

This is the result compared to RD:

	Screw	Pill	Capsule	Carpet	Grid	Tile	Wood	Zipper	Cable	Toothbrush	Transistor	Metal Nut	Bottle	Hazelnut	Leather
RD	98.03	97.63	97.93	99.36	95.49	100.00	99.39	97.16	95.45	91.39	97.87	100.00	100.00	100.00	100.00
RD++	94.71	94.44	90.65	99.80	90.98	99.96	99.39	85.95	94.79	92.22	97.29	99.22	100	100	100

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[samet-akcay]

@cjy513203427 thanks a lot for your contribution!

regarding the noise, it could be a torchvision transform such as SimplexNoise, and could be added as a train_transform from DataModule or from the model. Since this is a model specific transform, I think it might make more sense to define this as a model specific transform such as done here:

anomalib/src/anomalib/models/image/patchcore/lightning_model.py

Lines 131 to 143 in c00e101

	def configure_transforms(image_size: tuple[int, int] | None = None) -> Transform:
	"""Default transform for Padim."""
	image_size = image_size or (256, 256)
	# scale center crop size proportional to image size
	height, width = image_size
	center_crop_size = (int(height * (224 / 256)), int(width * (224 / 256)))
	return Compose(
	[
	Resize(image_size, antialias=True),
	CenterCrop(center_crop_size),
	Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
	],
	)