🔨 Replace imgaug with Native PyTorch Transforms (#2436)

* Add multi random choice transform

* Add DRAEMAugmenter class and Perlin noise generation to new_perlin.py

- Introduced DRAEMAugmenter for advanced image augmentations using torchvision v2.
- Implemented various augmentation techniques including ColorJitter, RandomAdjustSharpness, and custom transformations.
- Added functionality for comparing augmentation methods and visualizing results.
- Included utility functions for metrics computation and image processing.
- Established logging for better traceability of operations.

This commit enhances the image processing capabilities within the Anomalib framework, facilitating more robust anomaly detection workflows.

* Add the new perlin noise

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Add the new perlin noise

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* add generate_perlin_noise relative import

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* add tiffile as a dependency

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Remove upper bound from wandb

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Added skimage

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* add scikit-learn as a dependency

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* limit ollama to < 0.4.0 as it has breaking changes

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Fix data generators in test helpers

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Update the perlin augmenters

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Fix numpy validator tests caused by numpy upgrade

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Fix CS-Flow tests

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

* Fix the tests

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

---------

Signed-off-by: Samet Akcay <samet.akcay@intel.com>

pyproject.toml

@@ -42,11 +42,12 @@ core = [
     "av>=10.0.0",
     "einops>=0.3.2",
     "freia>=0.2",
-    "imgaug==0.4.0",
     "kornia>=0.6.6",
     "matplotlib>=3.4.3",
     "opencv-python>=4.5.3.56",
     "pandas>=1.1.0",
+    "scikit-image",  # NOTE: skimage should be removed as part of dependency cleanup
+    "tifffile",
     "timm",
     "lightning>=2.2",
     "torch>=2",
@@ -57,12 +58,12 @@ core = [
     "open-clip-torch>=2.23.0,<2.26.1",
 ]
 openvino = ["openvino>=2024.0", "nncf>=2.10.0", "onnx>=1.16.0"]
-vlm = ["ollama", "openai", "python-dotenv","transformers"]
+vlm = ["ollama<0.4.0", "openai", "python-dotenv","transformers"]
 loggers = [
     "comet-ml>=3.31.7",
     "gradio>=4",
     "tensorboard",
-    "wandb>=0.12.17,<=0.15.9",
+    "wandb",
     "mlflow >=1.0.0",
 ]
 notebooks = ["gitpython", "ipykernel", "ipywidgets", "notebook"]

src/anomalib/data/transforms/__init__.py

@@ -4,5 +4,6 @@
 # SPDX-License-Identifier: Apache-2.0
 
 from .center_crop import ExportableCenterCrop
+from .multi_random_choice import MultiRandomChoice
 
-__all__ = ["ExportableCenterCrop"]
+__all__ = ["ExportableCenterCrop", "MultiRandomChoice"]

src/anomalib/data/transforms/multi_random_choice.py

@@ -0,0 +1,82 @@
+"""Multi random choice transform."""
+
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+from collections.abc import Callable, Sequence
+
+import torch
+from torchvision.transforms import v2
+
+
+class MultiRandomChoice(v2.Transform):
+    """Apply multiple transforms randomly picked from a list.
+
+    This transform does not support torchscript.
+
+    Args:
+        transforms (sequence or torch.nn.Module): List of transformations to choose from.
+        probabilities (list[float] | None, optional): Probability of each transform being picked.
+            If None (default), all transforms have equal probability. If provided, probabilities
+            will be normalized to sum to 1.
+        num_transforms (int): Maximum number of transforms to apply at once.
+            Defaults to ``1``.
+        fixed_num_transforms (bool): If ``True``, always applies exactly ``num_transforms`` transforms.
+            If ``False``, randomly picks between 1 and ``num_transforms``.
+            Defaults to ``False``.
+
+    Examples:
+        >>> import torchvision.transforms.v2 as v2
+        >>> transforms = [
+        ...     v2.RandomHorizontalFlip(p=1.0),
+        ...     v2.ColorJitter(brightness=0.5),
+        ...     v2.RandomRotation(10),
+        ... ]
+        >>> # Apply 1-2 random transforms with equal probability
+        >>> transform = MultiRandomChoice(transforms, num_transforms=2)
+
+        >>> # Always apply exactly 2 transforms with custom probabilities
+        >>> transform = MultiRandomChoice(
+        ...     transforms,
+        ...     probabilities=[0.5, 0.3, 0.2],
+        ...     num_transforms=2,
+        ...     fixed_num_transforms=True
+        ... )
+    """
+
+    def __init__(
+        self,
+        transforms: Sequence[Callable],
+        probabilities: list[float] | None = None,
+        num_transforms: int = 1,
+        fixed_num_transforms: bool = False,
+    ) -> None:
+        if not isinstance(transforms, Sequence):
+            msg = "Argument transforms should be a sequence of callables"
+            raise TypeError(msg)
+
+        if probabilities is None:
+            probabilities = [1.0] * len(transforms)
+        elif len(probabilities) != len(transforms):
+            msg = f"Length of p doesn't match the number of transforms: {len(probabilities)} != {len(transforms)}"
+            raise ValueError(msg)
+
+        super().__init__()
+
+        self.transforms = transforms
+        total = sum(probabilities)
+        self.probabilities = [probability / total for probability in probabilities]
+
+        self.num_transforms = num_transforms
+        self.fixed_num_transforms = fixed_num_transforms
+
+    def forward(self, *inputs: torch.Tensor) -> torch.Tensor | tuple[torch.Tensor, ...]:
+        """Apply randomly selected transforms to the input."""
+        # First determine number of transforms to apply
+        num_transforms = (
+            self.num_transforms if self.fixed_num_transforms else int(torch.randint(self.num_transforms, (1,)) + 1)
+        )
+        # Get transforms
+        idx = torch.multinomial(torch.tensor(self.probabilities), num_transforms).tolist()
+        transform = v2.Compose([self.transforms[i] for i in idx])
+        return transform(*inputs)

src/anomalib/data/utils/__init__.py

@@ -3,10 +3,9 @@
 # Copyright (C) 2022-2024 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
-from .augmenter import Augmenter
 from .boxes import boxes_to_anomaly_maps, boxes_to_masks, masks_to_boxes
 from .download import DownloadInfo, download_and_extract
-from .generators import random_2d_perlin
+from .generators import generate_perlin_noise
 from .image import (
     generate_output_image_filename,
     get_image_filenames,
@@ -30,7 +29,7 @@
     "generate_output_image_filename",
     "get_image_filenames",
     "get_image_height_and_width",
-    "random_2d_perlin",
+    "generate_perlin_noise",
     "read_image",
     "read_mask",
     "read_depth_image",
@@ -42,7 +41,6 @@
     "TestSplitMode",
     "LabelName",
     "DirType",
-    "Augmenter",
     "masks_to_boxes",
     "boxes_to_masks",
     "boxes_to_anomaly_maps",

src/anomalib/data/utils/generators/__init__.py

@@ -3,6 +3,6 @@
 # Copyright (C) 2022 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
-from .perlin import random_2d_perlin
+from .perlin import PerlinAnomalyGenerator, generate_perlin_noise
 
-__all__ = ["random_2d_perlin"]
+__all__ = ["PerlinAnomalyGenerator", "generate_perlin_noise"]