DetectMultiBackend() default stride=32

[rglkt]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Optimization of default stride setting for YOLOv5 models.

📊 Key Changes

• Changed the default stride value from 64 to 32 in the model initialization.

🎯 Purpose & Impact

• 🎯 Purpose: This change likely aims to improve the default configuration for better compatibility or performance.
• 💡 Impact: Users might experience improved detection accuracy or speed due to the more optimal stride setting. This could be especially beneficial for real-time applications.

[github-actions]

👋 Hello @rglkt, thank you for submitting a YOLOv5 🚀 PR! To allow your work to be integrated as seamlessly as possible, we advise you to:

• ✅ Verify your PR is up-to-date with upstream/master. If your PR is behind upstream/master an automatic GitHub Actions merge may be attempted by writing /rebase in a new comment, or by running the following code, replacing 'feature' with the name of your local branch:

git remote add upstream https://github.com/ultralytics/yolov5.git
git fetch upstream
# git checkout feature  # <--- replace 'feature' with local branch name
git merge upstream/master
git push -u origin -f

• ✅ Verify all Continuous Integration (CI) checks are passing.
• ✅ Reduce changes to the absolute minimum required for your bug fix or feature addition. "It is not daily increase but daily decrease, hack away the unessential. The closer to the source, the less wastage there is." -Bruce Lee

[rglkt]

hello, when i try to use engine model, it has some wrong. i set my imgsz 224， the multipule of 32, but the common.py set it as 64 automatically and result to wrong. i think may be the default stride should be 32, when the n/s/m/l/x stride is the multiple of 32.

[zhiqwang]

Seems the P6 series need stride 64, so it would be better to set this parameter optional here.

[rglkt]

sorry, i can't realize the meaning of P6, do you mean as follow? i think the data configuration is used to config something about dataset, it's not suitable to set stride on it?

[rglkt]

sorry，i know， set it on cmd argument， i try it

[zhiqwang]

P6 is yolov5n6, yolov5s6, yolov5m6 and yolov5l6, introduced from tag 5.0.

[rglkt]

thanks！and i have added argument --stride

[glenn-jocher]

@rglkt some model formats (like PyTorch) can access attached metadata (like stride) to dynamically select the minimum viable stride, while other formats (like ONNX) do not have any metadata and thus lack stride information, in which case stride 64 is assumed since it is valid for both P5 models (minimum stride 32) and P6 models (minimum stride 64).

The only alternative here is to run an image and compare the input to the output to empirically attempt a stride determination at inference time, which would add delay on DetectMultiBackend init. I'll think this over a bit and see what we can do here.

[glenn-jocher]

@rglkt I've implemented ONNX metadata saving and reading now which resolves any stride issues for ONNX models. Unfortunately the rest of the formats are not resolved for this issue, like the engine issue in #7164.

One limitation in this PR is that detect.py is only one entrypoint for inference, we also have val.py and PyTorch Hub model inference.

I don't see a good solution to this other than to try to keep attaching custom metadata to each export format. So far I have PyTorch, TorchScript, and now ONNX available with stride and names metadata in DetectMultiBackend.

[rglkt]

Thank you for your work, I am a senior who is doing graduation design. This code helped me a lot, I think my pr is not the best solution, but I want to report the problem that other people will also encounter in the code and hope to contribute some power to the code. Thanks again for the open source code

[glenn-jocher]

@rglkt PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐

I think is probably a slight improvement over master, as now errors will appear on incorrect --imgsz rather than appearing on correct --imgsz.

The problem's not solved of course, as now errors will appear when users with P6 exported models attempt inference on stride 32 images, but at least the error will appear appropriately on an incorrect argument value.