Why is the default "stride" of common.DetectMultiBackend set to 64

[MarineCao]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

Hi, I'm trying to run detect.py with my onnx model, while I find that the image size is changed from (416, 416) to (448,448) by the check_img_size function, since model.stride=64. However, my model is a P5 model. Then, I find that the default stride is set to 64 in DetectMultiBackend as follows:
stride, names = 64, [f'class{i}' for i in range(1000)] # assign defaults
If the input is a .pt model, the stride will be changed by
stride = max(int(model.stride.max()), 32) # model stride
while, if the input is a .onnx model, the stride will never be changed. Then, if I exported the onnx mdoel by python export.py --device 0 --img 416, during inference, I got this error:

onnxruntime.capi.onnxruntime_pybind11_state.InvalidArgument: [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Got invalid dimensions for input: images for the following indices
 index: 2 Got: 448 Expected: 416
 index: 3 Got: 448 Expected: 416
 Please fix either the inputs or the model.

Process finished with exit code 1

Of course, if I use --dynamic to export .onnx model, the detect.py will run without error, but the image size is still changed to (448,448) and the inference speed is lower. And the the output anchors will be misplaced.
If I modify the default stride to 32, then all the outputs are correct.
So I'm confused about the default setting of stride. Is this a bug？ Or, is there a bug in my code?

Environment:
Win11 + anaconda + cuda 11.0 + python 3.8 + pytorch 1.7.1 + yolov5-6.0

Additional

No response

[github-actions]

👋 Hello @MarineCao, thank you for your interest in YOLOv5 🚀! Please visit our ⭐️ Tutorials to get started, where you can find quickstart guides for simple tasks like Custom Data Training all the way to advanced concepts like Hyperparameter Evolution.

If this is a 🐛 Bug Report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset images, training logs, screenshots, and a public link to online W&B logging if available.

For business inquiries or professional support requests please visit https://ultralytics.com or email [email protected].

Requirements

Python>=3.7.0 with all requirements.txt installed including PyTorch>=1.7. To get started:

git clone https://github.com/ultralytics/yolov5  # clone
cd yolov5
pip install -r requirements.txt  # install

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Google Colab and Kaggle notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training (train.py), validation (val.py), inference (detect.py) and export (export.py) on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

@MarineCao thanks for the question! 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]

@MarineCao good news 😃! Your original issue may now be fixed ✅ in PR #7353. This PR attaches stride and names as ONNX model metadata which are then read and used during inference.

To receive this update:

• Git – git pull from within your yolov5/ directory or git clone https://github.com/ultralytics/yolov5 again
• PyTorch Hub – Force-reload model = torch.hub.load('ultralytics/yolov5', 'yolov5s', force_reload=True)
• Notebooks – View updated notebooks
• Docker – sudo docker pull ultralytics/yolov5:latest to update your image

Thank you for spotting this issue and informing us of the problem. Please let us know if this update resolves the issue for you, and feel free to inform us of any other issues you discover or feature requests that come to mind. Happy trainings with YOLOv5 🚀!

[MarineCao]

@glenn-jocher Thanks for your quick reply. It works for me！