Add comprehensive debugging and tensor error handling

- Enhanced tensor shape verification with detailed logging - Added failsafe tensor interpolation for shape mismatches - Comprehensive error handling with fallback options - Added reload_node.py script for development - Created TROUBLESHOOTING.md with step-by-step fixes - Detailed console output for debugging tensor issues - Prevents crashes by returning single object as fallback
2025-08-22 13:21:36 -07:00
parent c6176beaf5
commit e331af2a48
3 changed files with 145 additions and 2 deletions
@@ -0,0 +1,79 @@
+# Troubleshooting Guide
+
+## ❌ "Sizes of tensors must match" Error
+
+If you're getting this error:
+```
+Error in object detection: Sizes of tensors must match except in dimension 0. Expected size 732 but got size 450 for tensor number 1 in the list.
+```
+
+### 🔧 Solution Steps:
+
+#### 1. **Restart ComfyUI** (Most Important)
+ComfyUI needs to be completely restarted to load updated custom nodes:
+- Close ComfyUI completely
+- Restart ComfyUI 
+- Load your workflow again
+
+#### 2. **Update from GitHub**
+Make sure you have the latest version:
+```bash
+cd ComfyUI/custom_nodes/ComfyUI_Detection_List
+git pull origin main
+```
+
+#### 3. **Check Debug Output**
+After restarting, you should see these debug messages in the console:
+```
+YOLOv8 Detection: Model inference complete
+YOLOv8 Detection: Found X objects, resizing to (width, height)
+Object 1 tensor shape: torch.Size([1, height, width, 3])
+Object 2 tensor shape: torch.Size([1, height, width, 3])
+Successfully created batch tensor with shape: torch.Size([X, height, width, 3])
+```
+
+If you **don't see these messages**, ComfyUI is still using the old version.
+
+#### 4. **Manual Reload** (Alternative)
+If restarting doesn't work, try:
+```bash
+cd ComfyUI/custom_nodes/ComfyUI_Detection_List
+python reload_node.py
+```
+
+### 🎯 What Was Fixed:
+
+- **Tensor Standardization**: All cropped objects are now resized to the same dimensions
+- **Shape Verification**: Added checks to ensure all tensors match before concatenating  
+- **Fallback Handling**: If concatenation still fails, returns single object instead of crashing
+- **Debug Logging**: Added detailed console output to track the process
+
+### 🔍 Verification:
+
+The node is working correctly when you see:
+1. ✅ Objects detected (e.g., "1 person, 1 tie")
+2. ✅ Debug messages showing tensor shapes
+3. ✅ "Successfully created batch tensor" message
+4. ✅ Multiple images output in the batch
+
+### 🚫 Common Issues:
+
+1. **Old Node Version**: ComfyUI not restarted after update
+2. **Installation Issues**: Node not properly installed in custom_nodes folder
+3. **Dependencies**: Missing ultralytics or torch packages
+
+### 📞 Still Having Issues?
+
+1. Check the [GitHub Issues](https://github.com/mangobyed/ComfyUI_Detection_List/issues)
+2. Create a new issue with:
+   - ComfyUI console output
+   - Your workflow JSON
+   - Python/ComfyUI versions
+   - Full error message
+
+### 💡 Pro Tips:
+
+- Always restart ComfyUI after updating custom nodes
+- Use lower confidence thresholds (0.1-0.3) to detect more objects
+- Increase max_size parameter for higher quality crops
+- Check console output for debug information
@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+"""
+Simple script to help reload the YOLOv8 detection node in ComfyUI
+Run this script after making changes to the node code.
+"""
+
+import importlib
+import sys
+import os
+
+def reload_node():
+    """Attempt to reload the YOLOv8 detection node"""
+    try:
+        # Add current directory to Python path
+        current_dir = os.path.dirname(os.path.abspath(__file__))
+        if current_dir not in sys.path:
+            sys.path.insert(0, current_dir)
+        
+        # Try to reload the module
+        module_name = 'yolov8_object_detection_node'
+        if module_name in sys.modules:
+            print(f"Reloading {module_name}...")
+            importlib.reload(sys.modules[module_name])
+            print("✅ Module reloaded successfully!")
+        else:
+            print(f"Module {module_name} not found in sys.modules")
+            print("Available modules:", [m for m in sys.modules.keys() if 'yolo' in m.lower()])
+        
+        # Also try reloading the main init module
+        init_module = '__init__'
+        if init_module in sys.modules:
+            importlib.reload(sys.modules[init_module])
+            print("✅ Init module reloaded!")
+            
+    except Exception as e:
+        print(f"❌ Error reloading module: {e}")
+        print("You may need to restart ComfyUI to load the updated node.")
+
+if __name__ == "__main__":
+    print("🔄 Attempting to reload YOLOv8 detection node...")
+    reload_node()
+    print("\n📝 If this doesn't work, please restart ComfyUI completely.")
@@ -189,12 +189,34 @@ class YOLOv8ObjectDetectionNode:
                print(f"YOLOv8 Detection: Found {len(cropped_objects)} objects, resizing to {target_size}")
                
                # Second pass: resize all objects to the same size and convert to tensors
-                for cropped_object in cropped_objects:
+                for i, cropped_object in enumerate(cropped_objects):
                    object_tensor = self.pil_to_tensor(cropped_object, target_size)
+                    print(f"Object {i+1} tensor shape: {object_tensor.shape}")
                    detected_images.append(object_tensor)
                
+                # Verify all tensors have the same shape before concatenating
+                if len(detected_images) > 1:
+                    first_shape = detected_images[0].shape
+                    for i, tensor in enumerate(detected_images):
+                        if tensor.shape != first_shape:
+                            print(f"WARNING: Tensor {i} shape mismatch: {tensor.shape} vs expected {first_shape}")
+                            # Force resize the tensor to match
+                            detected_images[i] = torch.nn.functional.interpolate(
+                                tensor.permute(0, 3, 1, 2), 
+                                size=(first_shape[1], first_shape[2]), 
+                                mode='bilinear', 
+                                align_corners=False
+                            ).permute(0, 2, 3, 1)
+                
                # Concatenate all detected objects into a batch
-                batch_tensor = torch.cat(detected_images, dim=0)
+                try:
+                    batch_tensor = torch.cat(detected_images, dim=0)
+                    print(f"Successfully created batch tensor with shape: {batch_tensor.shape}")
+                except Exception as cat_error:
+                    print(f"Error concatenating tensors: {cat_error}")
+                    # Fallback: return first detected object only
+                    batch_tensor = detected_images[0]
+                    print(f"Fallback: returning single object with shape: {batch_tensor.shape}")
                class_names_str = ", ".join(class_names)
                detection_info_str = "; ".join(detection_info_list)
                object_count = len(detected_images)