YOLOX OpenVINO Model Batch Inference issue

[LiuYiShan613]

I am currently conducting batch inference on the YOLOX model using the OpenVINO IR format. When the batch size is set to 1, the IR model's output shape is (1, 3549, 85). However, when the batch size is set to 'n', the output shape becomes (1, 3549 * n, 85). I anticipate the output shape to be (n, 3549, 85) to align with the requirements of the demo_postprocess function.

Consequently, I attempt to reshape the IR model's output from (1, 3549 * n, 85) to (n, 3549, 85). Unfortunately, the bounding boxes are not predicted correctly. How should I process this output data to ensure it fits the demo_postprocess function?

[tsubasahasumi]

Hello,

I followed the instructions provided in this guide to create an IR model for OpenVINO. However, like you, I encountered an issue where the model structure became corrupted when setting the input batch size to 'n'.

In my case, when running tools/export_onnx.py to create the ONNX model, I added the --no-onnxsim option, which seemed to resolve the problem for me. While I can't say for certain what caused the issue, it seems that the ONNX Simplifier (onnxsim) might be the culprit during the conversion process.

Hope it helps a bit.