在 C++ 中建構圖形

C++ 圖形建構工具是一款強大的工具,適用於:

  • 建構複雜的圖形
  • 參數化圖表 (例如在 InferenceCalculator 上設定委派代表、 啟用/停用圖形部分)
  • 重複的圖表 (例如 pbtxt 中的 CPU 和 GPU 專屬圖表) 您只要使用單一程式碼就能建構所需圖表 )
  • 支援選用圖形輸入/輸出
  • 依平台自訂圖表

基本用法

我們來看看 C++ 圖形建構工具如何用於簡單的圖表:

# Graph inputs.
input_stream: "input_tensors"
input_side_packet: "model"

# Graph outputs.
output_stream: "output_tensors"

node {
  calculator: "InferenceCalculator"
  input_stream: "TENSORS:input_tensors"
  input_side_packet: "MODEL:model"
  output_stream: "TENSORS:output_tensors"
  options: {
    [drishti.InferenceCalculatorOptions.ext] {
      # Requesting GPU delegate.
      delegate { gpu {} }
    }
  }
}

建構上述 CalculatorGraphConfig 的函式可能如下所示:

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>();

  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();
}

簡短摘要:

  • 使用 Graph::In/SideIn,以 Stream/SidePacket 的形式取得圖表輸入內容
  • 使用 Node::Out/SideOut 取得做為 Stream/SidePacket 的節點輸出內容
  • 使用 Stream/SidePacket::ConnectTo 將串流和側邊封包連線至 節點輸入 (Node::In/SideIn) 和圖形輸出 (Graph::Out/SideOut)
    • 設有「捷徑」運算子 >>,可用於取代 ConnectTo 函式 (例如 x >> node.In("IN"))。
  • Stream/SidePacket::Cast」已用於投放「AnyType」的串流或側邊封包 (例如 Stream<AnyType> in = graph.In(0);) 為特定類型
    • 使用實際類型而非 AnyType,有助於您找出 展現圖表建立工具的能力並改善圖表 且可讀性高

進階用法

Utility 函式

現在將推論建構程式碼擷取到專屬的公用函式中 有助於閱讀和重複使用程式碼:

// Updates graph to run inference.
Stream<std::vector<Tensor>> RunInference(
    Stream<std::vector<Tensor>> tensors, SidePacket<TfLiteModelPtr> model,
    const InferenceCalculatorOptions::Delegate& delegate, Graph& graph) {
  auto& inference_node = graph.AddNode("InferenceCalculator");
  auto& inference_opts =
      inference_node.GetOptions<InferenceCalculatorOptions>();
  *inference_opts.mutable_delegate() = delegate;
  tensors.ConnectTo(inference_node.In("TENSORS"));
  model.ConnectTo(inference_node.SideIn("MODEL"));
  return inference_node.Out("TENSORS").Cast<std::vector<Tensor>>();
}

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>();

  InferenceCalculatorOptions::Delegate delegate;
  delegate.mutable_gpu();
  Stream<std::vector<Tensor>> output_tensors =
      RunInference(input_tensors, model, delegate, graph);

  // Graph outputs.
  output_tensors.SetName("output_tensors").ConnectTo(graph.Out(0));

  return graph.GetConfig();
}

因此,RunInference 提供清楚的介面,指出 輸入/輸出及其類型

很容易重複使用,例如:但若您想執行額外的程式碼 模型推論:

  // Run first inference.
  Stream<std::vector<Tensor>> output_tensors =
      RunInference(input_tensors, model, delegate, graph);
  // Run second inference on the output of the first one.
  Stream<std::vector<Tensor>> extra_output_tensors =
      RunInference(output_tensors, extra_model, delegate, graph);

你也不需要重複的名稱和標記 (InferenceCalculatorTENSORSMODEL) 或在這裡導入專用常數,這些常數 系統會將詳細資料本地化為 RunInference 函式。

公用程式類別

而且,除了功能以外,在某些情況下 介紹公用程式類別,可協助您製作圖形建構程式碼 更容易閱讀,也較不容易出錯

MediaPipe 提供的 PassThroughCalculator 計算機功能 移除 YAML 檔案

input_stream: "float_value"
input_stream: "int_value"
input_stream: "bool_value"

output_stream: "passed_float_value"
output_stream: "passed_int_value"
output_stream: "passed_bool_value"

node {
  calculator: "PassThroughCalculator"
  input_stream: "float_value"
  input_stream: "int_value"
  input_stream: "bool_value"
  # The order must be the same as for inputs (or you can use explicit indexes)
  output_stream: "passed_float_value"
  output_stream: "passed_int_value"
  output_stream: "passed_bool_value"
}

我們來看看建立上方圖表的簡單 C++ 建構程式碼:

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Graph inputs.
  Stream<float> float_value = graph.In(0).SetName("float_value").Cast<float>();
  Stream<int> int_value = graph.In(1).SetName("int_value").Cast<int>();
  Stream<bool> bool_value = graph.In(2).SetName("bool_value").Cast<bool>();

  auto& pass_node = graph.AddNode("PassThroughCalculator");
  float_value.ConnectTo(pass_node.In("")[0]);
  int_value.ConnectTo(pass_node.In("")[1]);
  bool_value.ConnectTo(pass_node.In("")[2]);
  Stream<float> passed_float_value = pass_node.Out("")[0].Cast<float>();
  Stream<int> passed_int_value = pass_node.Out("")[1].Cast<int>();
  Stream<bool> passed_bool_value = pass_node.Out("")[2].Cast<bool>();

  // Graph outputs.
  passed_float_value.SetName("passed_float_value").ConnectTo(graph.Out(0));
  passed_int_value.SetName("passed_int_value").ConnectTo(graph.Out(1));
  passed_bool_value.SetName("passed_bool_value").ConnectTo(graph.Out(2));

  // Get `CalculatorGraphConfig` to pass it into `CalculatorGraph`
  return graph.GetConfig();
}

雖然 pbtxt 表示法可能很容易出錯 (如果有多項輸入內容要傳遞 ),C++ 程式碼看起來則更糟:重複的空白標記和 Cast 呼叫。我們 瞭解我們可以如何改進,只要導入 PassThroughNodeBuilder

class PassThroughNodeBuilder {
 public:
  explicit PassThroughNodeBuilder(Graph& graph)
      : node_(graph.AddNode("PassThroughCalculator")) {}

  template <typename T>
  Stream<T> PassThrough(Stream<T> stream) {
    stream.ConnectTo(node_.In(index_));
    return node_.Out(index_++).Cast<T>();
  }

 private:
  int index_ = 0;
  GenericNode& node_;
};

現在,圖形建構程式碼看起來會像這樣:

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Graph inputs.
  Stream<float> float_value = graph.In(0).SetName("float_value").Cast<float>();
  Stream<int> int_value = graph.In(1).SetName("int_value").Cast<int>();
  Stream<bool> bool_value = graph.In(2).SetName("bool_value").Cast<bool>();

  PassThroughNodeBuilder pass_node_builder(graph);
  Stream<float> passed_float_value = pass_node_builder.PassThrough(float_value);
  Stream<int> passed_int_value = pass_node_builder.PassThrough(int_value);
  Stream<bool> passed_bool_value = pass_node_builder.PassThrough(bool_value);

  // Graph outputs.
  passed_float_value.SetName("passed_float_value").ConnectTo(graph.Out(0));
  passed_int_value.SetName("passed_int_value").ConnectTo(graph.Out(1));
  passed_bool_value.SetName("passed_bool_value").ConnectTo(graph.Out(2));

  // Get `CalculatorGraphConfig` to pass it into `CalculatorGraph`
  return graph.GetConfig();
}

現在你無法在路途中放置錯誤順序或索引 而藉由猜測 PassThrough 中的 Cast 類型,即可省下一些輸入內容 。

建議做法

盡可能從一開始就定義圖表輸入內容

在下列程式碼中:

  • 要猜測圖表中有多少輸入資料,可能會不容易。
  • 整體而言可能出錯,而且難以維護 (例如 是否應該正確顯示索引?名稱?如果部分輸入內容遭到移除或選用,會有什麼影響? 等)。
  • 由於其他圖表可能會有差異,因此RunSomething重複使用的情形有限 輸入來源

錯誤做法 — 錯誤程式碼的例子。

Stream<D> RunSomething(Stream<A> a, Stream<B> b, Graph& graph) {
  Stream<C> c = graph.In(2).SetName("c").Cast<C>();  // Bad.
  // ...
}

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  Stream<A> a = graph.In(0).SetName("a").Cast<A>();
  // 10/100/N lines of code.
  Stream<B> b = graph.In(1).SetName("b").Cast<B>()  // Bad.
  Stream<D> d = RunSomething(a, b, graph);
  // ...

  return graph.GetConfig();
}

請改為在圖表建構工具的開頭定義圖表輸入內容:

正確 - 良好程式碼範例。

Stream<D> RunSomething(Stream<A> a, Stream<B> b, Stream<C> c, Graph& graph) {
  // ...
}

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Inputs.
  Stream<A> a = graph.In(0).SetName("a").Cast<A>();
  Stream<B> b = graph.In(1).SetName("b").Cast<B>();
  Stream<C> c = graph.In(2).SetName("c").Cast<C>();

  // 10/100/N lines of code.
  Stream<D> d = RunSomething(a, b, c, graph);
  // ...

  return graph.GetConfig();
}

如果您的輸入串流或側邊封包不屬於這些範圍,請使用 std::optional 務必將其定義為開頭,並將它放在開頭處:

正確 - 良好程式碼範例。

std::optional<Stream<A>> a;
if (needs_a) {
  a = graph.In(0).SetName(a).Cast<A>();
}

定義最後的圖表輸出內容

在下列程式碼中:

  • 很難猜出圖表中的輸出內容數量。
  • 整體而言可能出錯,而且難以維護 (例如 是否應該正確顯示索引?名稱?如果部分 Pod 遭移除或設定為可選,該怎麼辦? 等)。
  • 其他圖表的輸出內容可能不同,因此「RunSomething」重複使用的情形有限

錯誤做法 — 錯誤程式碼的例子。

void RunSomething(Stream<Input> input, Graph& graph) {
  // ...
  node.Out("OUTPUT_F")
      .SetName("output_f").ConnectTo(graph.Out(2));  // Bad.
}

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // 10/100/N lines of code.
  node.Out("OUTPUT_D")
      .SetName("output_d").ConnectTo(graph.Out(0));  // Bad.
  // 10/100/N lines of code.
  node.Out("OUTPUT_E")
      .SetName("output_e").ConnectTo(graph.Out(1));  // Bad.
  // 10/100/N lines of code.
  RunSomething(input, graph);
  // ...

  return graph.GetConfig();
}

請改為在圖表建構工具的最結尾定義圖表輸出內容:

正確 - 良好程式碼範例。

Stream<F> RunSomething(Stream<Input> input, Graph& graph) {
  // ...
  return node.Out("OUTPUT_F").Cast<F>();
}

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // 10/100/N lines of code.
  Stream<D> d = node.Out("OUTPUT_D").Cast<D>();
  // 10/100/N lines of code.
  Stream<E> e = node.Out("OUTPUT_E").Cast<E>();
  // 10/100/N lines of code.
  Stream<F> f = RunSomething(input, graph);
  // ...

  // Outputs.
  d.SetName("output_d").ConnectTo(graph.Out(0));
  e.SetName("output_e").ConnectTo(graph.Out(1));
  f.SetName("output_f").ConnectTo(graph.Out(2));

  return graph.GetConfig();
}

保持節點彼此分離

在 MediaPipe 中,封包串流和側邊封包與處理資料的意義一樣重要 節點。透過任何節點輸入需求和輸出產品清楚表達 可獨立使用容器耗用的串流和側邊封包 相當容易

錯誤做法 — 錯誤程式碼的例子。

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Inputs.
  Stream<A> a = graph.In(0).Cast<A>();

  auto& node1 = graph.AddNode("Calculator1");
  a.ConnectTo(node1.In("INPUT"));

  auto& node2 = graph.AddNode("Calculator2");
  node1.Out("OUTPUT").ConnectTo(node2.In("INPUT"));  // Bad.

  auto& node3 = graph.AddNode("Calculator3");
  node1.Out("OUTPUT").ConnectTo(node3.In("INPUT_B"));  // Bad.
  node2.Out("OUTPUT").ConnectTo(node3.In("INPUT_C"));  // Bad.

  auto& node4 = graph.AddNode("Calculator4");
  node1.Out("OUTPUT").ConnectTo(node4.In("INPUT_B"));  // Bad.
  node2.Out("OUTPUT").ConnectTo(node4.In("INPUT_C"));  // Bad.
  node3.Out("OUTPUT").ConnectTo(node4.In("INPUT_D"));  // Bad.

  // Outputs.
  node1.Out("OUTPUT").SetName("b").ConnectTo(graph.Out(0));  // Bad.
  node2.Out("OUTPUT").SetName("c").ConnectTo(graph.Out(1));  // Bad.
  node3.Out("OUTPUT").SetName("d").ConnectTo(graph.Out(2));  // Bad.
  node4.Out("OUTPUT").SetName("e").ConnectTo(graph.Out(3));  // Bad.

  return graph.GetConfig();
}

在上述程式碼中:

  • 節點彼此相連,例如:node4 知道輸入內容的位置 資源來自 (node1node2node3) 且導致重構變複雜 維護和程式碼重複使用
    • 這種使用模式會從 proto 表示法降級, 系統預設會分離
  • 系統會重複 node#.Out("OUTPUT") 呼叫,並造成可讀性的影響 可以改用更簡潔的名稱,並提供實際類型

因此,如要修正上述問題,您可以編寫下列圖形建構程式碼:

正確 - 良好程式碼範例。

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Inputs.
  Stream<A> a = graph.In(0).Cast<A>();

  // `node1` usage is limited to 3 lines below.
  auto& node1 = graph.AddNode("Calculator1");
  a.ConnectTo(node1.In("INPUT"));
  Stream<B> b = node1.Out("OUTPUT").Cast<B>();

  // `node2` usage is limited to 3 lines below.
  auto& node2 = graph.AddNode("Calculator2");
  b.ConnectTo(node2.In("INPUT"));
  Stream<C> c = node2.Out("OUTPUT").Cast<C>();

  // `node3` usage is limited to 4 lines below.
  auto& node3 = graph.AddNode("Calculator3");
  b.ConnectTo(node3.In("INPUT_B"));
  c.ConnectTo(node3.In("INPUT_C"));
  Stream<D> d = node3.Out("OUTPUT").Cast<D>();

  // `node4` usage is limited to 5 lines below.
  auto& node4 = graph.AddNode("Calculator4");
  b.ConnectTo(node4.In("INPUT_B"));
  c.ConnectTo(node4.In("INPUT_C"));
  d.ConnectTo(node4.In("INPUT_D"));
  Stream<E> e = node4.Out("OUTPUT").Cast<E>();

  // Outputs.
  b.SetName("b").ConnectTo(graph.Out(0));
  c.SetName("c").ConnectTo(graph.Out(1));
  d.SetName("d").ConnectTo(graph.Out(2));
  e.SetName("e").ConnectTo(graph.Out(3));

  return graph.GetConfig();
}

現在,您可以視需要移除 node1,將 b 設為圖表輸入內容,而非 更新 node2node3node4 需要更新 (與 proto 表示法相同) 因為兩者是分離的

總體上,上述程式碼更緊密地複製 proto 圖表:

input_stream: "a"

node {
  calculator: "Calculator1"
  input_stream: "INPUT:a"
  output_stream: "OUTPUT:b"
}

node {
  calculator: "Calculator2"
  input_stream: "INPUT:b"
  output_stream: "OUTPUT:C"
}

node {
  calculator: "Calculator3"
  input_stream: "INPUT_B:b"
  input_stream: "INPUT_C:c"
  output_stream: "OUTPUT:d"
}

node {
  calculator: "Calculator4"
  input_stream: "INPUT_B:b"
  input_stream: "INPUT_C:c"
  input_stream: "INPUT_D:d"
  output_stream: "OUTPUT:e"
}

output_stream: "b"
output_stream: "c"
output_stream: "d"
output_stream: "e"

除此之外,您現在可以擷取公用程式函式,以便在其他圖表中重複使用:

正確 - 良好程式碼範例。

Stream<B> RunCalculator1(Stream<A> a, Graph& graph) {
  auto& node = graph.AddNode("Calculator1");
  a.ConnectTo(node.In("INPUT"));
  return node.Out("OUTPUT").Cast<B>();
}

Stream<C> RunCalculator2(Stream<B> b, Graph& graph) {
  auto& node = graph.AddNode("Calculator2");
  b.ConnectTo(node.In("INPUT"));
  return node.Out("OUTPUT").Cast<C>();
}

Stream<D> RunCalculator3(Stream<B> b, Stream<C> c, Graph& graph) {
  auto& node = graph.AddNode("Calculator3");
  b.ConnectTo(node.In("INPUT_B"));
  c.ConnectTo(node.In("INPUT_C"));
  return node.Out("OUTPUT").Cast<D>();
}

Stream<E> RunCalculator4(Stream<B> b, Stream<C> c, Stream<D> d, Graph& graph) {
  auto& node = graph.AddNode("Calculator4");
  b.ConnectTo(node.In("INPUT_B"));
  c.ConnectTo(node.In("INPUT_C"));
  d.ConnectTo(node.In("INPUT_D"));
  return node.Out("OUTPUT").Cast<E>();
}

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Inputs.
  Stream<A> a = graph.In(0).Cast<A>();

  Stream<B> b = RunCalculator1(a, graph);
  Stream<C> c = RunCalculator2(b, graph);
  Stream<D> d = RunCalculator3(b, c, graph);
  Stream<E> e = RunCalculator4(b, c, d, graph);

  // Outputs.
  b.SetName("b").ConnectTo(graph.Out(0));
  c.SetName("c").ConnectTo(graph.Out(1));
  d.SetName("d").ConnectTo(graph.Out(2));
  e.SetName("e").ConnectTo(graph.Out(3));

  return graph.GetConfig();
}

將節點分開來提高可讀性

錯誤做法 — 錯誤程式碼的例子。

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Inputs.
  Stream<A> a = graph.In(0).Cast<A>();
  auto& node1 = graph.AddNode("Calculator1");
  a.ConnectTo(node1.In("INPUT"));
  Stream<B> b = node1.Out("OUTPUT").Cast<B>();
  auto& node2 = graph.AddNode("Calculator2");
  b.ConnectTo(node2.In("INPUT"));
  Stream<C> c = node2.Out("OUTPUT").Cast<C>();
  auto& node3 = graph.AddNode("Calculator3");
  b.ConnectTo(node3.In("INPUT_B"));
  c.ConnectTo(node3.In("INPUT_C"));
  Stream<D> d = node3.Out("OUTPUT").Cast<D>();
  auto& node4 = graph.AddNode("Calculator4");
  b.ConnectTo(node4.In("INPUT_B"));
  c.ConnectTo(node4.In("INPUT_C"));
  d.ConnectTo(node4.In("INPUT_D"));
  Stream<E> e = node4.Out("OUTPUT").Cast<E>();
  // Outputs.
  b.SetName("b").ConnectTo(graph.Out(0));
  c.SetName("c").ConnectTo(graph.Out(1));
  d.SetName("d").ConnectTo(graph.Out(2));
  e.SetName("e").ConnectTo(graph.Out(3));

  return graph.GetConfig();
}

在上述程式碼中,很難掌握每個節點的起始位置 結束。如要改善這個問題並協助程式碼閱讀器,只需留空 請使用下列程式碼:

正確 - 良好程式碼範例。

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Inputs.
  Stream<A> a = graph.In(0).Cast<A>();

  auto& node1 = graph.AddNode("Calculator1");
  a.ConnectTo(node1.In("INPUT"));
  Stream<B> b = node1.Out("OUTPUT").Cast<B>();

  auto& node2 = graph.AddNode("Calculator2");
  b.ConnectTo(node2.In("INPUT"));
  Stream<C> c = node2.Out("OUTPUT").Cast<C>();

  auto& node3 = graph.AddNode("Calculator3");
  b.ConnectTo(node3.In("INPUT_B"));
  c.ConnectTo(node3.In("INPUT_C"));
  Stream<D> d = node3.Out("OUTPUT").Cast<D>();

  auto& node4 = graph.AddNode("Calculator4");
  b.ConnectTo(node4.In("INPUT_B"));
  c.ConnectTo(node4.In("INPUT_C"));
  d.ConnectTo(node4.In("INPUT_D"));
  Stream<E> e = node4.Out("OUTPUT").Cast<E>();

  // Outputs.
  b.SetName("b").ConnectTo(graph.Out(0));
  c.SetName("c").ConnectTo(graph.Out(1));
  d.SetName("d").ConnectTo(graph.Out(2));
  e.SetName("e").ConnectTo(graph.Out(3));

  return graph.GetConfig();
}

此外,上述表示法與 CalculatorGraphConfig proto 相符 表示法

如果您將節點擷取至公用程式函式,這些節點會在函式中限定範圍 都有明確的起點和終點 包含:

正確 - 良好程式碼範例。

CalculatorGraphConfig BuildGraph() {
  Graph graph;

  // Inputs.
  Stream<A> a = graph.In(0).Cast<A>();

  Stream<B> b = RunCalculator1(a, graph);
  Stream<C> c = RunCalculator2(b, graph);
  Stream<D> d = RunCalculator3(b, c, graph);
  Stream<E> e = RunCalculator4(b, c, d, graph);

  // Outputs.
  b.SetName("b").ConnectTo(graph.Out(0));
  c.SetName("c").ConnectTo(graph.Out(1));
  d.SetName("d").ConnectTo(graph.Out(2));
  e.SetName("e").ConnectTo(graph.Out(3));

  return graph.GetConfig();
}