Fix efficient ad (#2015)

* Remove batch_size from config, make it hardcoded in the model (it is batch size for the imagenet dataloader, which should not be changed).

* Better description

* Removed Imagenet normalization, added check-up for train_batch_size

* Fix train_batch_size for testing EfficientAd

* Updated usage example

configs/model/efficient_ad.yaml

@@ -7,12 +7,11 @@ model:
     weight_decay: 1.0e-05
     padding: false
     pad_maps: true
-    batch_size: 1
 
 metrics:
   pixel:
     - AUROC
 
 trainer:
-  max_epochs: 200
+  max_epochs: 1000
   max_steps: 70000

src/anomalib/models/image/efficient_ad/README.md

@@ -18,7 +18,7 @@ Anomalies are detected as the difference in output feature maps between the teac
 
 ## Usage
 
-`python tools/train.py --model efficient_ad`
+`anomalib train --model EfficientAd --data anomalib.data.MVTec --data.train_batch_size 1`
 
 ## Benchmark
 

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

@@ -58,8 +58,6 @@ class EfficientAd(AnomalyModule):
         pad_maps (bool): relevant if padding is set to False. In this case, pad_maps = True pads the
             output anomaly maps so that their size matches the size in the padding = True case.
             Defaults to ``True``.
-        batch_size (int): batch size for imagenet dataloader
-            Defaults to ``1``.
     """
 
     def __init__(
@@ -71,7 +69,6 @@ def __init__(
         weight_decay: float = 0.00001,
         padding: bool = False,
         pad_maps: bool = True,
-        batch_size: int = 1,
     ) -> None:
         super().__init__()
 
@@ -83,7 +80,7 @@ def __init__(
             padding=padding,
             pad_maps=pad_maps,
         )
-        self.batch_size = batch_size
+        self.batch_size = 1  # imagenet dataloader batch_size is 1 according to the paper
         self.lr = lr
         self.weight_decay = weight_decay
 
@@ -237,9 +234,18 @@ def configure_optimizers(self) -> torch.optim.Optimizer:
     def on_train_start(self) -> None:
         """Called before the first training epoch.
 
-        First sets up the pretrained teacher model, then prepares the imagenette data, and finally calculates or
-        loads the channel-wise mean and std of the training dataset and push to the model.
+        First check if EfficientAd-specific parameters are set correctly (train_batch_size of 1
+        and no Imagenet normalization in transforms), then sets up the pretrained teacher model,
+        then prepares the imagenette data, and finally calculates or loads
+        the channel-wise mean and std of the training dataset and push to the model.
         """
+        if self.trainer.datamodule.train_batch_size != 1:
+            msg = "train_batch_size for EfficientAd should be 1."
+            raise ValueError(msg)
+        if self._transform and any(isinstance(transform, Normalize) for transform in self._transform.transforms):
+            msg = "Transforms for EfficientAd should not contain Normalize."
+            raise ValueError(msg)
+
         sample = next(iter(self.trainer.train_dataloader))
         image_size = sample["image"].shape[-2:]
         self.prepare_pretrained_model()
@@ -314,11 +320,10 @@ def learning_type(self) -> LearningType:
         return LearningType.ONE_CLASS
 
     def configure_transforms(self, image_size: tuple[int, int] | None = None) -> Transform:
-        """Default transform for Padim."""
+        """Default transform for EfficientAd. Imagenet normalization applied in forward."""
         image_size = image_size or (256, 256)
         return Compose(
             [
                 Resize(image_size, antialias=True),
-                Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
             ],
         )

tests/integration/model/test_models.py

@@ -204,7 +204,8 @@ def _get_objects(
                 root=dataset_path / "mvtec",
                 category="dummy",
                 task=task_type,
-                train_batch_size=2,
+                # EfficientAd requires train batch size 1
+                train_batch_size=1 if model_name == "efficient_ad" else 2,
             )
 
         model = get_model(model_name, **extra_args)