Some problems about the onnx-tensorrt source code.

[hy846130226]

Describe the issue

Hello.
I am learning the onnx-tensorrt code, I have some questions:

I have got the files

But I could not open it in netron.app

so my question is:
is this file a standard onnx mode? and what's the difference beteen this model and my source model

I could see some variant in compile method:

input_map
output_map
input_indexes
output_indexes

I see after the code build the engine, it will look up the input_map to build the input_indexes.

so my question is:
Is there a need to do this search all over again?
Because the tensorrt model is from my orignal onnx model, Shouldn't his two inputs and outputs be the same?

To reproduce

None

Urgency

No response

Platform

Windows

OS Version

WIN10

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

v1.16.3

ONNX Runtime API

C++

Architecture

X64

Execution Provider

TensorRT

Execution Provider Library Version

No response

[hy846130226]

And another question:

I see in tensorrt_execution_provider.cc, it has aleady created IBuilder.

Why it need create again in compile method?
Why can't you reuse an IBuilder object?

[hy846130226]

Could Anyone help me?

[chilo-ms]

so my question is:
is this file a standard onnx mode? and what's the difference beteen this model and my source model

This is a standard onnx model and you should be able to open with Netron.
Make sure model_proto->SerializeToOstream(dump) is complete and then check the file otherwise you might get the corrupt onnx file.
This onnx model might be slightly different than your original model.
original model -> ORT graph optimization -> ORT TRT generated onnx model -> TRT parser to get support list.
So, this model is optimized model that will be fed into TRT parser to get support list.

[chilo-ms]

so my question is:
Is there a need to do this search all over again?
Because the tensorrt model is from my orignal onnx model, Shouldn't his two inputs and outputs be the same?

That's a good question.
Theoretically, in most cases, the name or number of inputs/outputs should be the same for onnx model and TRT network.
I wonder if there are some historical reasons or in some cases they are not the same.

[chilo-ms]

And another question:

I see in tensorrt_execution_provider.cc, it has aleady created IBuilder.

Why it need create again in compile method? Why can't you reuse an IBuilder object?

We did. Please see ORT 1.17.0 or main branch

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.