Replacing A Subgraph

Introduction

This example first generates a model consisting of a Min op followed by a Max, and then uses the graph.layer() and graph.register() APIs seen in example 07 to create a function that can be used to replace this subgraph with a Clip op.

This can be useful, for example, to enable TensorRT plugins with ONNX.

Subgraph Replacement Basics

The process of replacing a subgraph involves 3 steps. For example, for a graph with the following structure:

     Tensor0
        |
      Node0
        |
     Tensor1    Tensor2
           \    /
            Node1
              |
           Tensor3
              |
            Node2

In order to replace the subgraph consisting of [Node0, Node1], we need to:

Disconnect the outputs of the subgraph inputs: Tensor0 and Tensor2

That means we need to delete the edge between Tensor0 and Node0, and between Tensor2 and Node1.
Disconnect the inputs of the subgraph outputs: Tensor3

That means we need to delete the edge between Node1 and Tensor3.

This will leave us with a graph like this:

     Tensor0     Tensor2

           Tensor3
              |
            Node2

And the now disconnected subgraph:

      Node0
        |
     Tensor1
           \
            Node1

3. Lastly, we need to insert our node such that it has inputs: [Tensor0, Tensor2] and outputs: [Tensor3, ].

After this step, we have our final graph (cleanup() will remove the dangling subgraph):

     Tensor0     Tensor2
           \     /
          MyNewNode0
              |
           Tensor3
              |
            Node2

Running the example

Generate a the model including Min and Max ops by running:
bash
```
python3 generate.py
```
The generated model will compute max(min(x, 6), 0) and look like this:
Replace the subgraph with a Clip op by running:
bash
```
python3 replace.py
```
The final model will include a clip(x, min=0, max=6) and look like this: