Hi Team,
I was using the hugging face utility from latest transformer version to convert a SQUAD based Roberta model to ONNX. After conversion, I observed the accuracy has dipped significantly even though I never quantized the model. Any suggestions or advice what could have been the reason. Does the ONNX conversion results in loss in terms of predictions? If so, is there any parameter which can be experimented with to reduce and yet improve the run time performance.