Hi,
This is duplicate to the TensorRT optimisation. Please check above topic for the suggestion.
Thanks.