nvidia--tensorrt/tools/onnx-graphsurgeon/examples/03_isolating_a_subgraph

Isolating A Subgraph

Introduction

This example first generates a basic model, then extracts a subgraph from this model.

Isolating a subgraph from a model is as simple as modifying the inputs and outputs of the graph, running graph.cleanup(), and then re-exporting the graph.

We do not need to know which nodes, initializers, or intermediate tensors we want - marking the inputs and outputs is sufficient for ONNX GraphSurgeon to be able to determine the other information automatically.

Running the example

1. Generate a model with several nodes and save it to model.onnx by running:

   python3 generate.py
   

   The generated model computes Y = x0 + (a * x1 + b):

   

2. Isolate the subgraph that computes (a * x1 + b) and save it to subgraph.onnx by running:

   python3 isolate.py
   

   The resulting model computes add_out = (a * x1 + b):