Issue 323 - Changes made in preprocessing images for torch detection models.

detectionmetrics/models/torch_detection.py

@@ -243,16 +243,21 @@ def __init__(
         # Build input transforms (resize, normalize, etc.)
         self.transform_input = []
 
+        # Default resize to 640x640 if not specified
         if "resize" in self.model_cfg:
-            self.transform_input += [
-                transforms.Resize(
-                    size=(
-                        self.model_cfg["resize"].get("height", None),
-                        self.model_cfg["resize"].get("width", None),
-                    ),
-                    interpolation=transforms.InterpolationMode.BILINEAR,
-                )
-            ]
+            resize_height = self.model_cfg["resize"].get("height", 640)
+            resize_width = self.model_cfg["resize"].get("width", 640)
+        else:
+            # Default to 640x640 when no resize is specified
+            resize_height = 640
+            resize_width = 640
+
+        self.transform_input += [
+            transforms.Resize(
+                size=(resize_height, resize_width),
+                interpolation=transforms.InterpolationMode.BILINEAR,
+            )
+        ]
 
         if "crop" in self.model_cfg:
             crop_size = (
@@ -403,7 +408,27 @@ def eval(
                     print("Skipping batch: empty image tensor detected.")
                     continue
 
+                # Move images to device and ensure consistent shapes for batching
                 images = [img.to(self.device) for img in images]
+
+                # For batch processing, we need to stack tensors, but they must have the same shape
+                # Even with resize transforms, there might be slight differences
+                if len(images) > 1:
+                    # Get the target shape from the first image
+                    target_shape = images[0].shape
+                    # Ensure all images have the same shape
+                    for i, img in enumerate(images):
+                        if img.shape != target_shape:
+                            # Resize to match the first image's shape
+                            images[i] = torch.nn.functional.interpolate(
+                                img.unsqueeze(0), 
+                                size=target_shape[-2:],  # [H, W]
+                                mode='bilinear', 
+                                align_corners=False
+                            ).squeeze(0)
+
+                # Stack images for batch processing
+                images = torch.stack(images)
                 predictions = self.model(images)
 
                 for i in range(len(images)):