🔧 refactor(data): Polish BMAD dataset

CHANGELOG.md

@@ -8,6 +8,8 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
+- 🚀 Add BMAD dataset by @code-dev05 in https://github.com/open-edge-platform/anomalib/pull/2900
+
 ### Removed
 
 ### Changed

docs/source/markdown/guides/reference/data/datamodules/image.md

@@ -7,6 +7,13 @@ Image datamodules in Anomalib are designed to handle image-based anomaly detecti
 ```{grid} 3
 :gutter: 2
 
+:::{grid-item-card} BMAD
+:link: anomalib.data.datamodules.image.bmad
+:link-type: doc
+
+BMAD dataset datamodule for medical anomaly detection.
+:::
+
 :::{grid-item-card} BTech
 :link: anomalib.data.datamodules.image.BTech
 :link-type: doc
@@ -89,7 +96,7 @@ Visual Anomaly dataset.
 
 ```{eval-rst}
 .. automodule:: anomalib.data
-   :members: BTech, Datumaro, Folder, Kolektor, MVTecAD, MVTecAD2, MVTecLOCO, RealIAD, Tabular, VAD, Visa
+   :members: BMAD, BTech, Datumaro, Folder, Kolektor, MVTecAD, MVTecAD2, MVTecLOCO, RealIAD, Tabular, VAD, Visa
    :undoc-members:
    :show-inheritance:
 ```

docs/source/markdown/guides/reference/data/datamodules/image/bmad.md

@@ -0,0 +1,7 @@
+# BMAD Datamodule
+
+```{eval-rst}
+.. automodule:: anomalib.data.datamodules.image.bmad
+   :members:
+   :show-inheritance:
+```

docs/source/markdown/guides/reference/data/datamodules/image/index.md

@@ -7,6 +7,13 @@ Anomalib provides various datamodules for handling image-based anomaly detection
 ```{grid} 3
 :gutter: 2
 
+:::{grid-item-card} BMAD
+:link: bmad
+:link-type: doc
+
+BMAD dataset datamodule for medical anomaly detection.
+:::
+
 :::{grid-item-card} BTech
 :link: btech
 :link-type: doc
@@ -54,6 +61,7 @@ Visual Anomaly (VisA) dataset datamodule.
 :hidden:
 :maxdepth: 1
 
+bmad
 btech
 datumaro
 folder

docs/source/markdown/guides/reference/data/datamodules/index.md

@@ -7,6 +7,13 @@ Anomalib provides various datamodules for different types of data modalities. Th
 ```{grid} 3
 :gutter: 2
 
+:::{grid-item-card} BMAD
+:link: image/bmad
+:link-type: doc
+
+BMAD dataset datamodule for medical anomaly detection.
+:::
+
 :::{grid-item-card} BTech
 :link: image/btech
 :link-type: doc

examples/configs/README.md

@@ -8,6 +8,7 @@ The configurations in this folder are organized as follows.
 configs/
 ├── data
 │   ├── avenue.yaml
+│   ├── bmad.yaml
 │   ├── btech.yaml
 │   ├── folder_3d.yaml
 │   ├── folder.yaml

examples/configs/data/bmad.yaml

@@ -0,0 +1,12 @@
+class_path: anomalib.data.BMAD
+init_args:
+  root: ./datasets/BMAD
+  category: Brain
+  train_batch_size: 32
+  eval_batch_size: 32
+  num_workers: 8
+  test_split_mode: from_dir
+  test_split_ratio: 0.2
+  val_split_mode: from_dir
+  val_split_ratio: null
+  seed: null

src/anomalib/data/datamodules/image/__init__.py

@@ -6,6 +6,7 @@
 This module contains data modules for loading and processing image datasets for
 anomaly detection. The following data modules are available:
 
+- ``BMAD``: BMAD Dataset for Medical Anomaly Detection
 - ``BTech``: BTech Surface Defect Dataset
 - ``Datumaro``: Dataset in Datumaro format (Intel Geti™ export)
 - ``Folder``: Custom folder structure with normal/abnormal images
@@ -50,6 +51,7 @@ class ImageDataFormat(str, Enum):
 
         The following dataset formats are supported:
 
+    - ``BMAD``: BMAD Dataset for Medical Anomaly Detection
     - ``BTECH``: BTech Surface Defect Dataset
     - ``DATUMARO``: Dataset in Datumaro format
     - ``FOLDER``: Custom folder structure
@@ -66,6 +68,7 @@ class ImageDataFormat(str, Enum):
     - ``VISA``: Visual Anomaly Dataset
     """
 
+    BMAD = "bmad"
     BTECH = "btech"
     DATUMARO = "datumaro"
     FOLDER = "folder"

src/anomalib/data/datamodules/image/bmad.py

@@ -1,4 +1,4 @@
-# Copyright (C) 2022-2025 Intel Corporation
+# Copyright (C) 2025 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
 """BMAD Data Module.
@@ -14,20 +14,22 @@
 Example:
     Create a BMAD datamodule::
 
-        >>> from your_module import BMADDataModule
-        >>> datamodule = BMADDataModule(
+        >>> from anomalib.data import BMAD
+        >>> datamodule = BMAD(
         ...     root="./datasets/BMAD",
-        ...     dataset="brain",       # options: "Brain", "Chest", "Histopathology", "Liver", "Retina_OCT2017",
+        ...     dataset="Brain",       # options: "Brain", "Chest", "Histopathology", "Liver", "Retina_OCT2017",
                                                     "Retina_RESC"
         ... )
 
 Notes:
     The dataset will be automatically downloaded and reorganized upon first usage.
     Directory structure after preparation may look like:
 
+    .. code-block:: text
+
         datasets/
         └── BMAD/
-            ├── brain/
+            ├── Brain/
             │   ├── train/
             │   │   └── good/
             │   ├── valid/
@@ -67,12 +69,10 @@
 logger = logging.getLogger(__name__)
 
 DOWNLOAD_INFO = DownloadInfo(
-    name="mvtecad",
-    url="https://www.mydrive.ch/shares/38536/3830184030e49fe74747669442f0f282/"
-    "download/420938113-1629952094/mvtec_anomaly_detection.tar.xz",
-    hashsum="cf4313b13603bec67abb49ca959488f7eedce2a9f7795ec54446c649ac98cd3d",
+    name="bmad",
+    url="https://huggingface.co/datasets/code-dev05/BMAD/resolve/main/bmad.zip",
+    hashsum="df655def31f3f638a91c567550de54d6e45a74b2368f666c13a6a3052c063165",
 )
-# TODO(Devansh Agarwal): Update the download url.  # noqa: TD003
 
 
 class BMAD(AnomalibDataModule):
@@ -149,8 +149,8 @@ def __init__(
         augmentations: Transform | None = None,
         test_split_mode: TestSplitMode | str = TestSplitMode.FROM_DIR,
         test_split_ratio: float = 0.2,
-        val_split_mode: ValSplitMode | str = ValSplitMode.SAME_AS_TEST,
-        val_split_ratio: float = 0.5,
+        val_split_mode: ValSplitMode | str = ValSplitMode.FROM_DIR,
+        val_split_ratio: float | None = None,
         seed: int | None = None,
     ) -> None:
         super().__init__(
@@ -177,11 +177,12 @@ def _setup(self, _stage: str | None = None) -> None:
             root=self.root,
             category=self.category,
         )
-        self.validation_data = BMADDataset(
-            split="valid",
-            root=self.root,
-            category=self.category,
-        )
+        if self.val_split_mode == ValSplitMode.FROM_DIR:
+            self.val_data = BMADDataset(
+                split="valid",
+                root=self.root,
+                category=self.category,
+            )
         self.test_data = BMADDataset(
             split=Split.TEST,
             root=self.root,
@@ -208,7 +209,7 @@ def prepare_data(self) -> None:
 
                 datasets/
                 └── BMAD/
-                    ├── brain/
+                    ├── Brain/
                     ├── Liver/
                     └── ...
         """

src/anomalib/data/datasets/image/__init__.py

@@ -6,6 +6,7 @@
 This module provides dataset implementations for various image anomaly detection
 datasets:
 
+- ``BMADDataset``: BMAD dataset containing medical images
 - ``BTechDataset``: BTech dataset containing industrial objects
 - ``DatumaroDataset``: Dataset in Datumaro format (Intel Geti™ export)
 - ``FolderDataset``: Custom dataset from folder structure

src/anomalib/data/datasets/image/bmad.py

@@ -1,4 +1,4 @@
-# Copyright (C) 2024 Intel Corporation
+# Copyright (C) 2025 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
 """BMAD Dataset.
@@ -109,12 +109,34 @@ def make_bmad_dataset(path: Path, split: str | Split | None = None) -> DataFrame
     """Create BMAD samples by parsing the dataset structure.
 
     The files are expected to follow the structure:
-        ``path/to/dataset/category/split/image_filename.png``
-        ``path/to/dataset/category/split/mask_filename.png``
+
+    .. code-block:: text
+
+        path/to/dataset
+        ├── category
+        │   ├── train
+        │   │   ├── good
+        │   │   │   ├── image_filename.png
+        │   │   ├── Ungood
+        │   │   │   ├── img
+        │   │   │   │   ├── image_filename.png
+        │   │   │   │   └── label
+        │   │   │   │       └── image_filename.png
+        │   │   └── test
+        │   │       ├── good
+        │   │       │   ├── img
+        │   │       │   │   ├── image_filename.png
+        │   │       │   │   └── label
+        │   │       │   │       └── image_filename.png
+        │   │       └── Ungood
+        │   │           ├── img
+        │   │           │   ├── image_filename.png
+        │   │           │   └── label
+
+
     Args:
-        root (Path | str): Path to dataset root directory.
+        path (Path | str): Path to dataset root directory.
         split (str | Split | None, optional): Dataset split (train/test/val). Defaults to ``None``.
-        extensions (Sequence[str] | None, optional): Valid image extensions. Defaults to ``None``.
 
     Returns:
         DataFrame with columns:
@@ -175,7 +197,7 @@ def make_bmad_dataset(path: Path, split: str | Split | None = None) -> DataFrame
     samples["mask_path"] = None
     if len(mask_samples):
         samples.loc[
-            (samples.split == "test") | (samples.split == "valid"),
+            ((samples.split == "test") | (samples.split == "valid")) & (samples.label_index == LabelName.ABNORMAL),
             "mask_path",
         ] = mask_samples.image_path.to_numpy()
 
@@ -190,14 +212,13 @@ def make_bmad_dataset(path: Path, split: str | Split | None = None) -> DataFrame
         ):
             msg = (
                 "Mismatch between anomalous images and ground truth masks. Make sure "
-                "mask files in 'ground_truth' folder follow the same naming "
+                "mask files in 'Ungood/label/' folder follow the same naming "
                 "convention as the anomalous images (e.g. image: '000.png', "
-                "mask: '000.png' or '000_mask.png')."
+                "mask: '000.png')."
             )
             raise MisMatchError(msg)
 
     samples.attrs["task"] = "classification" if (samples["mask_path"] == "").all() else "segmentation"
-    split = "train"
     if split:
         samples = samples[samples.split == split].reset_index(drop=True)
 

tests/helpers/data.py

@@ -415,6 +415,36 @@ def _generate_dummy_tabular_dataset(self) -> None:
         """Generate dummy folder structure for tabular dataset in a temporary directory."""
         self._generate_dummy_folder_dataset()
 
+    def _generate_dummy_bmad_dataset(self) -> None:
+        """Generate dummy BMAD dataset in directory."""
+        dataset_category = "dummy"
+        # train split. Images are in train/good
+        split_path = self.dataset_root / dataset_category / "train" / self.normal_category
+        for i in range(self.num_train):
+            label = LabelName.NORMAL
+            image_filename = split_path / f"{i:03}.png"
+            self.image_generator.generate_image(label=label, image_filename=image_filename)
+        # Good images are in subset/normal_category/img/i000.png
+        for split in ("test", "valid"):
+            split_path = self.dataset_root / dataset_category / split / self.normal_category / "img"
+            for i in range(self.num_test):
+                label = LabelName.NORMAL
+                image_filename = split_path / f"{i:03}.png"
+                self.image_generator.generate_image(label=label, image_filename=image_filename)
+        # Abnormal images are in subset/abnormal_category/img/i000.png
+        # and subset/abnormal_category/label/i000.png
+        for split in ("test", "valid"):
+            split_path = self.dataset_root / dataset_category / split / self.abnormal_category
+            for i in range(self.num_test):
+                label = LabelName.ABNORMAL
+                image_filename = split_path / "img" / f"{i:03}.png"
+                mask_filename = split_path / "label" / f"{i:03}.png"
+                self.image_generator.generate_image(
+                    label=label,
+                    image_filename=image_filename,
+                    mask_filename=mask_filename,
+                )
+
     def _generate_dummy_btech_dataset(self) -> None:
         """Generate dummy BeanTech dataset in directory using the same convention as BeanTech AD."""
         # BeanTech AD follows the same convention as MVTec AD.

tests/unit/data/datamodule/image/test_bmad.py

@@ -0,0 +1,40 @@
+# Copyright (C) 2025 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+"""Unit Tests - BMAD Datamodule."""
+
+from pathlib import Path
+
+import pytest
+from torchvision.transforms.v2 import Resize
+
+from anomalib.data import BMAD
+from anomalib.data.utils import ValSplitMode
+from tests.unit.data.datamodule.base.image import _TestAnomalibImageDatamodule
+
+
+class TestBMAD(_TestAnomalibImageDatamodule):
+    """BMAD Datamodule Unit Tests."""
+
+    @pytest.fixture()
+    @staticmethod
+    def datamodule(dataset_path: Path) -> BMAD:
+        """Create and return a BMAD datamodule."""
+        datamodule_ = BMAD(
+            root=dataset_path / "bmad",
+            category="dummy",
+            train_batch_size=4,
+            eval_batch_size=4,
+            val_split_mode=ValSplitMode.FROM_DIR,
+            augmentations=Resize((256, 256)),
+        )
+        datamodule_.prepare_data()
+        datamodule_.setup()
+
+        return datamodule_
+
+    @pytest.fixture()
+    @staticmethod
+    def fxt_data_config_path() -> str:
+        """Return the path to the test data config."""
+        return "examples/configs/data/bmad.yaml"