Class FunctionalGraph

Represents a model graph built with the Functional API.

Inheritance

object

FunctionalGraph

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Namespace: Unity.InferenceEngine

Assembly: Unity.InferenceEngine.dll

Syntax

[MovedFrom("Unity.Sentis")]
public class FunctionalGraph

Remarks

FunctionalGraph defines a computation graph by adding inputs, applying Functional operations to FunctionalTensor values, and declaring outputs. Use AddInput(DataType, DynamicTensorShape, string) or AddInputs(Model) to define graph inputs, then chain operations from the Functional class. Call AddOutput(FunctionalTensor, string) or AddOutputs(params FunctionalTensor[]) to declare which tensors the graph produces, then Compile() to build an optimized runtime Model.

You can create new models from scratch, wrap or extend existing models with Forward(Model, params FunctionalTensor[]), or build preprocessing and postprocessing pipelines.

Additional resources

Examples

Create a functional graph from scratch. Build a graph with inputs, operations, and outputs, then compile to a model.

var graph = new FunctionalGraph();
var x = graph.AddInput<float>(new TensorShape(6), "input_x");
var y = graph.AddInput<float>(new TensorShape(6), "input_y");
var prod = x * y;
var reduce = Functional.ReduceSum(prod, dim: 0, keepdim: false);
graph.AddOutput(reduce, "output");
Model model = graph.Compile();

 // Modify an existing model.
 // Declare a functional graph.
 var graph = new FunctionalGraph();
 // Get the input functional tensor from the graph with input data type and shape matching that of the original model input.
var RGB = graph.AddInput(sourceModel, 0);
 // Apply f(x) = x^(1/2.2) element-wise to transform from RGB to sRGB.
 var sRGB = Functional.Pow(RGB, Functional.Constant(1 / 2.2f));
 // Apply f(x) = x * 2 - 1 element-wise to transform values from the range [0, 1] to the range [-1, 1].
 var sRGB_normalised = sRGB * 2 - 1;
 // Apply the forward method of the source model to the transformed functional input and add the outputs to the graph.
 var outputs = Functional.Forward(sourceModel, sRGB_normalised);
 graph.AddOutputs(outputs);
 // Compile the graph to return the final model.
 m_RuntimeModel = graph.Compile();

Methods

AddInput(DataType, DynamicTensorShape, string)

Appends an input to the graph with the specified data type and dynamic shape.

Declaration

public FunctionalTensor AddInput(DataType dataType, DynamicTensorShape shape, string name = null)

Parameters

Type	Name	Description
DataType	dataType	The data type of the input.
DynamicTensorShape	shape	The dynamic shape of the input.
string	name	The name of the input. If null, defaults to `input_` plus the input index.

Returns

Type	Description
FunctionalTensor	The functional tensor input.

Remarks

Use this overload when the input has dynamic dimensions (for example, variable batch size or sequence length).

Examples

Add an input with dynamic shape for variable batch or sequence length.

var shape = DynamicTensorShape.DynamicOfRank(2);
var input = graph.AddInput(DataType.Float, shape, "input");

AddInput(DataType, TensorShape, string)

Appends an input to the graph with the specified data type and static shape.

Declaration

public FunctionalTensor AddInput(DataType dataType, TensorShape shape, string name = null)

Parameters

Type	Name	Description
DataType	dataType	The data type of the input.
TensorShape	shape	The static shape of the input.
string	name	The name of the input. If null, defaults to `input_` plus the input index.

Returns

Type	Description
FunctionalTensor	The functional tensor input.

Remarks

Use this overload when all dimensions of the input are known at compile time.

Examples

Add an input with static shape.

var input = graph.AddInput(DataType.Float, new TensorShape(1, 3, 224, 224), "image");

AddInput(Model, int, string)

Appends an input to the graph that matches a specific input of an existing model.

Declaration

public FunctionalTensor AddInput(Model model, int index, string name = null)

Parameters

Type	Name	Description
Model	model	The model whose input to match.
int	index	The index of the input in the model.
string	name	The name for this input. If null, uses the name from the model.

Returns

Type	Description
FunctionalTensor	The functional tensor input.

Remarks

Use this overload when extending or wrapping an existing model. The new input has the same data type and shape as the model input at the given index. Pass the resulting tensor(s) to Forward(Model, params FunctionalTensor[]) to run the model.

Examples

Add an input matching a model input, then run the model and add its output.

var input = graph.AddInput(sourceModel, 0);
var outputs = Functional.Forward(sourceModel, new[] { input });
graph.AddOutput(outputs[0]);

AddInput<T>(DynamicTensorShape, string)

Appends an input to the graph with type inferred from T and a dynamic shape.

Declaration

public FunctionalTensor AddInput<T>(DynamicTensorShape shape, string name = null) where T : unmanaged

Parameters

Type	Name	Description
DynamicTensorShape	shape	The dynamic shape of the input.
string	name	The name of the input. If null, defaults to `input_` plus the input index.

Returns

Type	Description
FunctionalTensor	The functional tensor input.

Type Parameters

Name	Description
T	The element type of the input (for example, `float` or `int`).

Remarks

The data type is derived from the generic parameter (for example, float for AddInput<float>). Use when you prefer type inference over explicit DataType.

Examples

Add an input with type inferred from the generic parameter.

var input = graph.AddInput<float>(DynamicTensorShape.DynamicOfRank(1));

AddInput<T>(TensorShape, string)

Appends an input to the graph with type inferred from T and a static shape.

Declaration

public FunctionalTensor AddInput<T>(TensorShape shape, string name = null) where T : unmanaged

Parameters

Type	Name	Description
TensorShape	shape	The static shape of the input.
string	name	The name of the input. If null, defaults to `input_` plus the input index.

Returns

Type	Description
FunctionalTensor	The functional tensor input.

Type Parameters

Name	Description
T	The element type of the input (for example, `float` or `int`).

Remarks

The data type is derived from the generic parameter (for example, float for AddInput<float>). Use when you prefer type inference over explicit DataType.

Examples

Add inputs with type inferred from the generic parameter.

var x = graph.AddInput<float>(new TensorShape(6), "input_x");
var y = graph.AddInput<int>(new TensorShape(3), "indices");

AddInputs(Model)

Appends inputs to the graph that match all inputs of an existing model.

Declaration

public FunctionalTensor[] AddInputs(Model model)

Parameters

Type	Name	Description
Model	model	The model whose inputs to match.

Returns

Type	Description
FunctionalTensor[]	The array of functional tensor inputs, one per model input.

Remarks

Use this overload when wrapping or extending a model. Each input matches the corresponding model input by index. Pass the returned array to Forward(Model, params FunctionalTensor[]) to run the model.

Examples

Add inputs matching all model inputs, run the model, and add a postprocessed output.

var inputs = graph.AddInputs(sourceModel);
var softmax = Functional.Softmax(outputs[0]);
graph.AddOutput(softmax);

AddOutput(FunctionalTensor, string)

Appends an output to the graph from a functional tensor.

Declaration

public void AddOutput(FunctionalTensor output, string name = null)

Parameters

Type	Name	Description
FunctionalTensor	output	The functional tensor to use as an output.
string	name	The name for the output. If null, defaults to `output_` plus the output index, or is inferred from the original model when the tensor is from a forward pass.

Remarks

The output tensor must be derived from inputs or constants in the graph. If the tensor comes from an input or constant, it is cloned before being added. You must add at least one output before calling Compile().

Examples

Add an output from a functional tensor.

var x = graph.AddInput<float>(new TensorShape(6));
var result = Functional.Relu(x);
graph.AddOutput(result, "output");

AddOutputs(params FunctionalTensor[])

Appends multiple outputs to the graph from the given functional tensors.

Declaration

public void AddOutputs(params FunctionalTensor[] outputs)

Parameters

Type	Name	Description
FunctionalTensor[]	outputs	The functional tensors to use as outputs.

Remarks

Equivalent to calling AddOutput(FunctionalTensor, string) for each tensor in the array. Use when the graph produces multiple outputs. You must add at least one output before calling Compile()

Examples

Add multiple outputs at once.

var x = graph.AddInput<float>(new TensorShape(6));
var y = graph.AddInput<float>(new TensorShape(6));
var reduce = Functional.ReduceSum(x * y, dim: 0, keepdim: false);
var prod = x * y;
graph.AddOutputs(reduce, prod);

Compile()

Compiles a graph with the given outputs and returns an optimized runtime model.

Declaration

public Model Compile()

Returns

Type	Description
Model	The compiled runtime model.

Remarks

You must add at least one output to the graph using AddOutput(FunctionalTensor, string) or AddOutputs(params FunctionalTensor[]) before calling this method. The returned model is optimized and ready to run with a Worker.

Examples

Compile the graph to an optimized model.

var x = graph.AddInput<float>(new TensorShape(6));
var result = Functional.Relu(x);
graph.AddOutput(result);
Model model = graph.Compile();

Compile(params FunctionalTensor[])

Compiles a graph with the given outputs and returns an optimized runtime model.

Declaration

public Model Compile(params FunctionalTensor[] outputs)

Parameters

Type	Name	Description
FunctionalTensor[]	outputs	The functional tensors to use as outputs.

Returns

Type	Description
Model	The compiled runtime model.

Remarks

Adds the specified tensors as outputs and then compiles. With this overload, don't add outputs to the graph separately or AddOutputs(params FunctionalTensor[]). The returned model is optimized and ready to run with a Worker. Throws an exception if outputs were already added to the graph.

Examples

Compile with inline outputs (no prior AddOutput call).

var x = graph.AddInput<float>(new TensorShape(6));
var y = graph.AddInput<float>(new TensorShape(6));
var result = x * y;
Model model = graph.Compile(result);

Exceptions

Type	Condition
AssertionException	Thrown when outputs have already been added to the graph.