Set static input size for rtmdet conversion to onnx for onnxruntime backend

[mcusgithub]

I'm able to convert rtmdet to onnx as described in the documentation for the onnxruntime backend with:

1. dynamic input shape:
   python3 tools/deploy.py configs/mmdet/detection/detection_onnxruntime_dynamic.py /mmdetection/configs/rtmdet/rtmdet_s_8xb32-300e_coco.py /mmdetection/configs/rtmdet/rtmdet_s_8xb32-300e_coco_20220905_161602-387a891e.pth demo/resources/det.jpg --work-dir ./work_dir --device cpu

2. static input shape:
   python3 tools/deploy.py configs/mmdet/detection/detection_onnxruntime_static.py /mmdetection/configs/rtmdet/rtmdet_s_8xb32-300e_coco.py /mmdetection/configs/rtmdet/rtmdet_s_8xb32-300e_coco_20220905_161602-387a891e.pth demo/resources/det.jpg --work-dir ./work_dir --device cpu

I can't however figure out how to specify a static input shape by modifying the mmdet .py configuration file

• I know it can be done for the tensorrt backend using configs/mmdet/detection/detection_tensorrt_static-640x640.py for instance
• I know it can be done in mmyolo with projects/easydeploy/tools/export.py ... --img-size 320 640 --batch 1 ...

I tried to change the default static configuration in various ways, like for instance:

onnx_config = dict(input_shape=[320, 640])

backend_config = dict(
    model_inputs=[
        dict(
            input_shapes=dict(
                input=dict(
                    min_shape=[1, 3, 320, 640],
                    opt_shape=[1, 3, 320, 640],
                    max_shape=[1, 3, 320, 640])))
    ])

but the parameters are always ignored.

Can someone help me figuring it out?
Thanks a lot.

[irexyc]

@mcusgithub

I tried to change the default static configuration in various ways, like for instance:

For onnxruntime backend, you should use deploy configs for tensorrt.

When you use deploy.py and want a custom static input shape. For tensorrt, you can use dynamic deploy config but set min/opt/max shape to the same value. For onnxruntime, as mmdeploy will use test_pipeline to create input for model convertion (https://github.com/open-mmlab/mmdeploy/blob/main/mmdeploy/codebase/mmdet/deploy/object_detection.py#L166-L223), the key it to create a input with custom size. Take rtmdet_s_8xb32-300e_coco.py for example, as it inherit test_pipeline in rtmdet_l_8xb32-300e_coco.py, you can edit the these two lines to create the input shape you want.