help me to run mmyolo models on cpu

[Vinaygoudasp7]

hi @okhovan @adk9 @dmatveev @cabelo
I'm using MMYOLO and have successfully trained a model that gives good results. Now, I want to run the model on CPU by converting it to OpenVINO. I followed the conversion pipeline: MMYOLO → ONNX → OpenVINO (using the OpenVINO CLI).

However, after conversion, the inference results are not as expected. Could you please suggest possible solutions or guide me on how to resolve this?

Thanks in advance!

[Vinaygoudasp7]

hi any one please ans this

[Vinaygoudasp7]

Hi @praasz @alexsu52 @faizan102418

can you please help me to solve this

[Ganesh-403]

@Vinaygoudasp7 When a model converts successfully but the inference results are "not as expected", the first thing I'd try to determine is where the discrepancy is introduced:

• during MMYOLO → ONNX export, or
• during ONNX → OpenVINO conversion.

A few suggestions:

1. Validate the ONNX model first

   Before converting to OpenVINO, run inference with ONNX Runtime on the same test image and compare it against the original MMYOLO/PyTorch output.

   If the ONNX model already produces different detections, the issue is likely in the export step rather than OpenVINO. :contentReference[oaicite:0]{index=0}

2. Compare raw outputs, not just final detections

   If the ONNX outputs match PyTorch but OpenVINO does not, compare the raw output tensors before NMS/post-processing. That helps determine whether the difference is in the network itself or in the decoding/post-processing pipeline.

3. Consider using MMDeploy instead of a manual ONNX → OpenVINO pipeline

   MMYOLO officially recommends MMDeploy for deployment. It integrates the post-processing (including NMS and decoding) into the deployment pipeline and provides dedicated deployment configurations for supported models and backends. A manual ONNX export can sometimes lead to differences in post-processing or model configuration. :contentReference[oaicite:1]{index=1}

4. Share the model details

   It would help to know:

   • Which MMYOLO model are you using? (YOLOv5, YOLOX, RTMDet, etc.)
   • Which MMYOLO version?
   • Which OpenVINO version?
   • Which export command did you use to generate the ONNX model?
   • Are the outputs completely incorrect, or are they simply lower-confidence / missing detections?

Without that information it's difficult to tell whether this is:

• an export issue,
• a post-processing mismatch,
• or an OpenVINO conversion issue.

If the ONNX model matches the original PyTorch model but the OpenVINO model does not, I'd be interested in seeing a minimal reproducible example (model, ONNX export command, and a sample image), as that would make it much easier to determine whether this is a conversion regression.

If this solves your problem, feel free to mark it as the accepted answer so others can find it easily.