🔨 Replace imgaug with Native PyTorch Transforms

[samet-akcay]

📝 Description

• Removed imgaug dependency from project requirements
• Replaced Augmenter class with new PerlinAnomalyGenerator using torchvision transforms
• Refactored perlin noise generation to use PyTorch native operations
• Added new MultiRandomChoice transform class for flexible augmentation pipelines
• Updated all dependent modules (DRAEM, DSR) to use the new implementation
• 🛠️ 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

Examples

Before (using imgaug):

from imgaug import augmenters as iaa

# Define augmentations
augmenters = [
  iaa.GammaContrast((0.5, 2.0), per_channel=True),
  iaa.MultiplyAndAddToBrightness(mul=(0.8, 1.2), add=(-30, 30)),
  iaa.pillike.EnhanceSharpness(),
  # ... more imgaug augmenters
]

# Apply rotation
rot = iaa.Sequential([iaa.Affine(rotate=(-90, 90))])
perlin_noise = rot(image=perlin_noise_np)

After (Using torch vision)

from torchvision.transforms import v2

# Define augmentations using MultiRandomChoice

augmenters = MultiRandomChoice(
    transforms=[
        v2.ColorJitter(contrast=(0.5, 2.0)),
        v2.Compose([
        v2.Lambda(lambda x: x torch.empty(1).uniform(0.8, 1.2).item()),
        v2.Lambda(lambda x: torch.clamp(x + torch.empty(1).uniform_(-30/255, 30/255).item(), 0, 1)),]),
        v2.RandomAdjustSharpness(sharpness_factor=2.0, p=1.0),
        # ... more torchvision transforms
    ],
    num_transforms=3,
    fixed_num_transforms=True
)

# Apply rotation
generator = PerlinAnomalyGenerator(
    anomaly_source_path="path/to/images",
    blend_factor=(0.1, 0.9),
    probability=0.5,
    rotation_range=(-90, 90)
)
augmented_image, mask = generator(image)

✅ 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.

[codecov]

Codecov Report

Attention: Patch coverage is 87.59690% with 16 lines in your changes missing coverage. Please review.

Please upload report for BASE (feature/v2@c16f51e). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
src/anomalib/models/image/dsr/anomaly_generator.py	27.27%	8 Missing ⚠️
...rc/anomalib/data/transforms/multi_random_choice.py	82.60%	4 Missing ⚠️
src/anomalib/data/utils/generators/perlin.py	95.00%	4 Missing ⚠️

Additional details and impacted files

@@              Coverage Diff              @@
##             feature/v2    #2436   +/-   ##
=============================================
  Coverage              ?   78.28%           
=============================================
  Files                 ?      307           
  Lines                 ?    13115           
  Branches              ?        0           
=============================================
  Hits                  ?    10267           
  Misses                ?     2848           
  Partials              ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ashwinvaidya17]

One minor comment

[ashwinvaidya17]

Maybe for a different PR but it would be nice if we could remove inline methods.

[ashwinvaidya17]

Is this correct?

[samet-akcay]

with numpy 2.0, yes

[djdameln]

Did you confirm that the generated anomalous images look similar, visually, as the images generated with the previous approach?

[samet-akcay]

rated anomalous images look similar, visually, as the images generated with the previous approach?

Here is a comparison

Some stats

Image_MSE:
  mean: 0.006824
  std: 0.003222
  min: 0.002471
  max: 0.016470

Image_PSNR:
  mean: 22.084712
  std: 1.897144
  min: 17.833178
  max: 26.071310

Image_SSIM:
  mean: 0.949429
  std: 0.020979
  min: 0.900574
  max: 0.976912

Mask_MSE:
  mean: 0.097563
  std: 0.025687
  min: 0.050312
  max: 0.152783

Mask_IoU:
  mean: 0.020791
  std: 0.015581
  min: 0.000000
  max: 0.051682