The huggingface/transformers provides DINOv2

app.py

@@ -1,32 +1,29 @@
-import gradio as gr
 import torch
-from patchcore import patchcore
-from patchcore.datasets import MVTecDataset
+from transformers import AutoModel, AutoImageProcessor
 from PIL import Image
-import numpy as np
+import gradio as gr
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-# Load pre-trained PatchCore model
-# Example: using pretrained model for 'fabric' from MVTec dataset
-model = patchcore.PatchCore(device=device)
-model.load_from_file("pretrained/patchcore_fabric.pth")  # You need to provide the correct model file
+# Load DINOv2 model
+model_name = "facebook/dinov2-base"
+model = AutoModel.from_pretrained(model_name).to(device)
+processor = AutoImageProcessor.from_pretrained(model_name)
 
-def detect_anomaly(img):
-    img = np.array(img.convert("RGB"))
+# Embed a normal sample (you can do this during inference without storing)
+# You can even hardcode embeddings later if you want
 
-    # Preprocess and predict
-    score, anomaly_map = model.predict(img)
-
-    # Simple threshold
-    threshold = 0.5  # tune this depending on your dataset
-    if score > threshold:
-        result = f"Anomaly Detected | Score: {score:.2f}"
-    else:
-        result = f"Normal | Score: {score:.2f}"
-        
-    return result
+def detect_anomaly(img):
+    inputs = processor(images=img, return_tensors="pt").to(device)
+    with torch.no_grad():
+        features = model(**inputs).last_hidden_state.mean(dim=1)
+    
+    # Fake logic for now (just show feature norm)
+    # Later you can compare with normal samples' embeddings
+    score = torch.norm(features).item()
+    
+    return f"Feature Norm (use for anomaly logic): {score:.2f}"
 
 gr.Interface(fn=detect_anomaly,
              inputs=gr.Image(),
-             outputs="text").launch()
+             outputs="text").launch()