TensorRT / ONNX output only one box

[defrag-bambino]

Hello,

I have trained RT-DETR using Paddle. Then I exported it as described in the README and in various previous issues (e.g. #95). However, I get a strange output:

these are the trt bindings:

m_shape: shape: (1, 2), dtype: <class 'numpy.float32'>, mode: TensorIOMode.INPUT
image: shape: (-1, 4, 640, 640), dtype: <class 'numpy.float32'>, mode: TensorIOMode.INPUT
scale_factor: shape: (1, 2), dtype: <class 'numpy.float32'>, mode: TensorIOMode.INPUT
tile_3.tmp_0: shape: (), dtype: <class 'numpy.int32'>, mode: TensorIOMode.OUTPUT
reshape2_56.tmp_0: shape: (-1, 6), dtype: <class 'numpy.float32'>, mode: TensorIOMode.OUTPUT

notice how tile_3.tmp_0 has shape (). The reshape2_56.tmp_0 hast shape (batch, 6), which seems to be correct.

The output of the network for a batch size of 9 is a single (label, score, bbox) for each of the 9 images:

The tile_3.tmp_0 output produces a single value of "300". The single bbox seems to be the one with the highest confidence.

Any ideas what may cause this?

Viewing the ONNX with www.netron.app shows this: