fix: lazy load torch to prevent windows DLL crash on startup

app.py

@@ -234,15 +234,27 @@ def browse_dataset_path():
                     )
 
         # Load model action in sidebar
-        from perceptionmetrics.models.torch_detection import TorchImageDetectionModel
+        try:
+            from perceptionmetrics.models.torch_detection import TorchImageDetectionModel
+            _torch_available = True
+        except (ImportError, OSError):
+            _torch_available = False
         import json, tempfile
 
+        if not _torch_available:
+            st.warning(
+                "⚠️ PyTorch غير متاح على هذا النظام. "
+                "تبويب Inference لن يعمل. "
+                "يُرجى تثبيت PyTorch بشكل صحيح."
+            )
+
         load_model_btn = st.button(
             "Load Model",
             type="primary",
             width="stretch",
             help="Load and save the model for use in the Inference tab",
             key="sidebar_load_model_btn",
+            disabled=not _torch_available,
         )
 
         if load_model_btn:

perceptionmetrics/models/__init__.py

@@ -8,22 +8,22 @@
 
     REGISTRY["torch_image_segmentation"] = TorchImageSegmentationModel
     REGISTRY["torch_lidar_segmentation"] = TorchLiDARSegmentationModel
-except ImportError:
+except (ImportError, OSError):
     print("Torch not available")
 
 try:
     from perceptionmetrics.models.torch_detection import TorchImageDetectionModel
 
     REGISTRY["torch_image_detection"] = TorchImageDetectionModel
-except ImportError:
+except (ImportError, OSError):
     print("Torch detection not available")
 
 try:
     from perceptionmetrics.models.tf_segmentation import TensorflowImageSegmentationModel
 
     REGISTRY["tensorflow_image_segmentation"] = TensorflowImageSegmentationModel
-except ImportError:
+except (ImportError, OSError):
     print("Tensorflow not available")
 
 if not REGISTRY:
-    raise Exception("No valid deep learning framework found")
+    raise ImportError("No valid deep learning framework found")

tabs/inference.py

@@ -3,7 +3,6 @@
 import streamlit as st
 import json
 from PIL import Image
-import torch
 
 
 def draw_detections(image: Image, predictions: dict, label_map: Optional[dict] = None):
@@ -18,6 +17,7 @@ def draw_detections(image: Image, predictions: dict, label_map: Optional[dict] =
     :return: np.ndarray with detections drawn (for st.image)
     :rtype: np.ndarray
     """
+    import torch
     from perceptionmetrics.utils import image as ui
 
     boxes = predictions.get("boxes", torch.empty(0)).cpu().numpy()
@@ -103,6 +103,7 @@ def inference_tab():
                     st.markdown("#### Detection Details")
 
                     # Convert predictions to JSON format
+                    import torch
                     detection_results = []
                     boxes = predictions.get("boxes", torch.empty(0)).cpu().numpy()
                     labels = predictions.get("labels", torch.empty(0)).cpu().numpy()