Adds "Building Graphs in C++" initial page and updates "Graph" page to link to that section showcasing alternative C++ graph representation.

PiperOrigin-RevId: 508517348
This commit is contained in:
MediaPipe Team 2023-02-09 17:18:07 -08:00 committed by Copybara-Service
parent 2163920ee8
commit be0681c61d
2 changed files with 129 additions and 1 deletions

View File

@ -0,0 +1,98 @@
---
layout: default
title: Building Graphs in C++
parent: Graphs
nav_order: 1
---
# Building Graphs in C++
{: .no_toc }
1. TOC
{:toc}
---
## C++ Graph Builder
C++ graph builder is a powerful tool for:
* Building complex graphs
* Parametrizing graphs (e.g. setting a delegate on
`InferenceCalculator`, enabling/disabling parts of the graph)
* Deduplicating graphs (e.g. instead of CPU and GPU dedicated graphs in pbtxt
you can have a single code that constructs required graphs, sharing as much
as possible)
* Supporting optional graph inputs/outputs
* Customizing graphs per platform
### Basic Usage
Let's see how C++ graph builder can be used for a simple graph:
```proto
// Graph inputs.
input_stream: "input_tensors"
input_side_packet: "model"
// Graph outputs.
output_stream: "output_tensors"
// Nodes.
node {
calculator: "InferenceCalculator"
input_stream: "TENSORS:input_tensors"
input_side_packet: "MODEL:model"
output_stream: "TENSORS:output_tensors"
node_options: {
[type.googleapis.com/mediapipe.InferenceCalculatorOptions] {
# Requesting GPU delegate.
delegate { gpu {} }
}
}
}
```
Function to build the above `CalculatorGraphConfig` may look like:
```c++
CalculatorGraphConfig BuildGraph() {
Graph graph;
// Graph inputs.
Stream<std::vector<Tensor>> input_tensors =
graph.In(0).SetName("input_tensors").Cast<std::vector<Tensor>>();
SidePacket<TfLiteModelPtr> model =
graph.SideIn(0).SetName("model").Cast<TfLiteModelPtr>();
// Nodes.
auto& inference_node = graph.AddNode("InferenceCalculator");
auto& inference_opts =
inference_node.GetOptions<InferenceCalculatorOptions>();
// Requesting GPU delegate.
inference_opts.mutable_delegate()->mutable_gpu();
input_tensors.ConnectTo(inference_node.In("TENSORS"));
model.ConnectTo(inference_node.SideIn("MODEL"));
Stream<std::vector<Tensor>> output_tensors =
inference_node.Out("TENSORS").Cast<std::vector<Tensor>>();
// Graph outputs.
output_tensors.SetName("output_tensors").ConnectTo(graph.Out(0));
// Get `CalculatorGraphConfig` to pass it into `CalculatorGraph`
return graph.GetConfig();
}
```
Short summary:
* Use `Graph::In/SideIn` to get graph inputs as `Stream/SidePacket`
* Use `Node::Out/SideOut` to get node outputs as `Stream/SidePacket`
* Use `Stream/SidePacket::ConnectTo` to connect streams and side packets to node
inputs (`Node::In/SideIn`) and graph outputs (`Graph::Out/SideOut`)
* There's a "shortcut" operator `>>` that you can use instead of
`ConnectTo` function (E.g. `x >> node.In("IN")`).
* `Stream/SidePacket::Cast` is used to cast stream or side packet of `AnyType` (E.g. `Stream<AnyType> in = graph.In(0);`) to a particular type
* Using actual types instead of `AnyType` sets you on a better path for unleashing graph
builder capabilities and improving your graphs readability.

View File

@ -36,6 +36,7 @@ passthrough calculators :
# This graph named main_pass_throughcals_nosubgraph.pbtxt contains 4
# passthrough calculators.
input_stream: "in"
output_stream: "out"
node {
calculator: "PassThroughCalculator"
input_stream: "in"
@ -54,10 +55,39 @@ node {
node {
calculator: "PassThroughCalculator"
input_stream: "out3"
output_stream: "out4"
output_stream: "out"
}
```
MediaPipe offers an alternative `C++` representation for complex graphs (e.g. ML pipelines, handling model metadata, optional nodes, etc.). The above graph may look like:
```c++
CalculatorGraphConfig BuildGraphConfig() {
Graph graph;
// Graph inputs
Stream<AnyType> in = graph.In(0).SetName("in");
auto pass_through_fn = [](Stream<AnyType> in,
Graph& graph) -> Stream<AnyType> {
auto& node = graph.AddNode("PassThroughCalculator");
in.ConnectTo(node.In(0));
return node.Out(0);
};
Stream<AnyType> out1 = pass_through_fn(in, graph);
Stream<AnyType> out2 = pass_through_fn(out1, graph);
Stream<AnyType> out3 = pass_through_fn(out2, graph);
Stream<AnyType> out4 = pass_through_fn(out3, graph);
// Graph outputs
out4.SetName("out").ConnectTo(graph.Out(0));
return graph.GetConfig();
}
```
See more details in [Building Graphs in C++](building_graphs_cpp.md)
## Subgraph
To modularize a `CalculatorGraphConfig` into sub-modules and assist with re-use