Nanos models onnx exportation inconsistent with previous model

[lucasjinreal]

Search before asking

• [X] I have searched the DAMO-YOLO issues and found no similar feature requests.

Description

Previous model have outputs concatenated, but new one separated a lot.

This is not a problem, but caused very serious inconsistency for deploy code which could be same without even change a line to deploy all DAMO series models

Old:

New:

Can we just merge them all exactly same as previous and same like yolov5/yolov7/yolov8 so that deploy code can compatible at maximum?

Use case

No response

Additional

No response

Are you willing to submit a PR?

• [ ] Yes I'd like to help by submitting a PR!

[lucasjinreal]

Also, I even doubt the new nanos models are wrong, since I can not really figure out where is box....

[XianzheXu]

Hi, the new model's in model zoo is exported with --benchmark, which means the post-processing is ignored. You can export onnx model without --benchmark to keep consistency with your deploy codes.

In the onnx output, tensor shaped as [1, 32, 52, 52], [1, 32, 26, 26], [1, 32, 13, 13] is the box before post-process. This is for the benchmark consistency with picodet.

[XianzheXu]

In order to provide convenience for developers and maintain consistency with our previous code, we have updated the ONNX download link. You can now download the ONNX model that includes post-processing.

[lucasjinreal]

@XianzheXu thank u so much!

[lucasjinreal]

@XianzheXu hi, I downloaded damoyolo_nano_small.onnx, it still same.

[XianzheXu]

@XianzheXu hi, I downloaded damoyolo_nano_small.onnx, it still same.

Please retry it, the download link is updated now.