What a difference in tensort api and onnx to trt?

Could someone tell me what a difference in build with tensort api and build with onnx? When i test in 2d net work, the speed of them seem to be same and the speed in 3d point clound network may be some different. I like to build net work with tensorRT api but i don’t know whether it has some advantage than onnx.Thanks!

i just want to known when to use trt api and use onnx if all of them can be work…


TensorRT API and trtexec tool both use the ONNX model to build the TensorRT engine(they internally use the ONNX parser to read the model).
In case your purpose is just to build the engine, you can use either of them.

API can be used for additional functionalities like,

  • Build a custom plugin needed
  • Add optimization profiles for Dynamic shape inputs
  • Build Network from scratch
  • Run inference, etc.

We recommend you to please go through the following docs for more information.

Thank you.