Update Detect() grid init for loop

[glenn-jocher]

May resolve threaded inference issue in #9425 (comment) by avoiding memory sharing on init.

Signed-off-by: Glenn Jocher glenn.jocher@ultralytics.com

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved initialization of grid and anchor grid variables in YOLOv5 model.

📊 Key Changes

• Changed the way grid variables are initialized from a list containing a single empty tensor to a list comprehension creating separate empty tensors for each detection layer.
• Applied analogous changes to the anchor grid variable initialization.

🎯 Purpose & Impact

• These changes prevent potential issues with multiple layers referencing the same tensor, enabling proper individual layer operations.
• By ensuring each detection layer has its own unique grid and anchor grid, the update enhances the model's reliability and correctness, which might lead to subtle performance improvements for users of the YOLOv5 object detection model. 🚀

[glenn-jocher]

Problem statement that this PR resolves:

import torch
x = [torch.tensor(1)] * 10
x
Out[8]: 
[tensor(1),
 tensor(1),
 tensor(1),
 tensor(1),
 tensor(1),
 tensor(1),
 tensor(1),
 tensor(1),
 tensor(1),
 tensor(1)]
x[3] *=2
x
Out[10]: 
[tensor(2),
 tensor(2),
 tensor(2),
 tensor(2),
 tensor(2),
 tensor(2),
 tensor(2),
 tensor(2),
 tensor(2),
 tensor(2)]