🐞 fix(data): Perlin noise generator masks

pyproject.toml

@@ -37,7 +37,7 @@ dependencies = [
     "rich_argparse",        # CLI help-formatter
     # Torch and Lightning dependencies
     "lightning>=2.2",
-    "torchmetrics>=1.3.2",
+    "torchmetrics>=1.8.2",
     "lightning-utilities",
     # Deep learning models and utilities
     "timm",

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

@@ -269,6 +269,11 @@ def generate_perturbation(
         """
         # Generate perlin noise
         perlin_noise = generate_perlin_noise(height, width, device=device)
+        # in some cases the perlin noise is all less than 0.5, so we need to rescale it between 0 and 1
+        if not (perlin_noise > 0.5).any():
+            perlin_noise = (perlin_noise - perlin_noise.min()) / (perlin_noise.max() - perlin_noise.min())
+            # rescale to [-1, 1] range
+            perlin_noise = (perlin_noise * 2) - 1
 
         # Create rotated noise pattern
         perlin_noise = perlin_noise.unsqueeze(0)  # [1, H, W]

src/anomalib/models/image/dsr/anomaly_generator.py

@@ -75,6 +75,11 @@ def generate_anomaly(self, height: int, width: int) -> Tensor:
         # Generate perlin noise using the new function
         perlin_noise = generate_perlin_noise(height, width, scale=(perlin_scalex, perlin_scaley))
 
+        # Rescale with threshold at center if all values are below the threshold
+        if not (perlin_noise > threshold).any():
+            perlin_noise = (perlin_noise - perlin_noise.min()) / (perlin_noise.max() - perlin_noise.min())
+            perlin_noise = (perlin_noise * 2) - 1
+
         # Apply random rotation
         perlin_noise = perlin_noise.unsqueeze(0)  # Add channel dimension for transform
         perlin_noise = self.rot(perlin_noise).squeeze(0)  # Remove channel dimension

src/anomalib/models/image/supersimplenet/anomaly_generator.py

@@ -63,6 +63,13 @@ def generate_perlin(self, batches: int, height: int, width: int) -> torch.Tensor
             # keep power of 2 here for reproduction purpose, although this function supports power2 internally
             perlin_noise = generate_perlin_noise(height=perlin_height, width=perlin_width)
 
+            # Rescale with threshold at center if all values are below the threshold
+            if not (perlin_noise > self.threshold).any():
+                # First normalize to [0,1] range by min-max normalization
+                perlin_noise = (perlin_noise - perlin_noise.min()) / (perlin_noise.max() - perlin_noise.min())
+                # Then rescale to [-1, 1] range
+                perlin_noise = (perlin_noise * 2) - 1
+
             # original is power of 2 scale, so fit to our size
             perlin_noise = F.interpolate(
                 perlin_noise.reshape(1, 1, perlin_height, perlin_width),

src/anomalib/models/image/vlm_ad/backends/chat_gpt.py

@@ -36,12 +36,6 @@
 
 from lightning_utilities.core.imports import module_available
 
-# Optional import for dotenv
-try:
-    from dotenv import load_dotenv
-except ImportError:
-    load_dotenv = None
-
 from anomalib.models.image.vlm_ad.utils import Prompt
 
 from .base import Backend
@@ -51,6 +45,11 @@
 else:
     OpenAI = None
 
+if module_available("dotenv"):
+    from dotenv import load_dotenv
+else:
+    load_dotenv = None
+
 if TYPE_CHECKING:
     from openai.types.chat import ChatCompletion
 

src/anomalib/models/image/winclip/torch_model.py

@@ -39,6 +39,9 @@
 from anomalib.data import InferenceBatch
 from anomalib.models.components import BufferListMixin, DynamicBufferMixin
 
+from .prompting import create_prompt_ensemble
+from .utils import class_scores, harmonic_aggregation, make_masks, visual_association_score
+
 # Optional import for open_clip
 try:
     import open_clip
@@ -47,8 +50,6 @@
     open_clip = None
     tokenize = None
 
-from .prompting import create_prompt_ensemble
-from .utils import class_scores, harmonic_aggregation, make_masks, visual_association_score
 
 BACKBONE = "ViT-B-16-plus-240"
 PRETRAINED = "laion400m_e31"