Namespace Unity.Sentis

Classes

CPUTensorData

Represents Burst-specific internal data storage for a Tensor.

CommandBufferWorkerExtensions

Provides extension methods for scheduling the worker from a CommandBuffer.

ComputeTensorData

Represents data storage for a Tensor as a compute buffer, for GPUCompute backend.

Constant

Represents a constant in a model.

Functional

Represents the static functional methods for model building and compilation.

FunctionalExtensions

Represents extension functions for the Sentis functional API.

FunctionalGraph

Represents a model graph using the functional API.

Input functional tensors can be added to the graph, then manipulated using the functional API methods.

The functional graph can be compiled to return an optimized Sentis runtime model.

FunctionalTensor

Represents a tensor that is a result of tensor operations.

Layer

Represents the base class for all model layers.

Model

Represents a Sentis neural network.

ModelAsset

Represents a Sentis model asset.

ModelLoader

Provides methods for loading models.

ModelQuantizer

Provides methods for quantizing models.

ModelWriter

Provides methods for saving models.

NativeTensorArray

Represents an area of native memory that's exposed to managed code.

NativeTensorArrayFromManagedArray

Represents an area of managed memory that's exposed as if it's native memory.

Random

Represents a pseudo-random number generator used by Sentis.

Tensor

Represents data in a multidimensional array-like structure.

Ownership and lifetime:

  • Disposed needs to be called on the main thread.
  • Ownership is always to the owner of the object.

Data Representation:

  • TensorShape represents the data layout of the tensor
  • Data is held by a tensorData (ITensorData) which can be on a given backend
  • Data is stored in a flattened row major format
  • Data can be pending (ie computation is being done in parallel)
    • call CompleteAllPendingOperations for a blocking call to finish computing the tensor's data Data can be in a non readable type (GPU/NPU)
    • Call CompleteAllPendingOperations to finish computing the tensor's data
    • Call ReadbackAndClone or ReadBackAndCloneAsync to allow reading the tensor's data

Data manipulation

  • ToReadOnlyArray returns a copy of the tensor's data
  • dataOnBackend can be manipulated directly to avoid a unnecessary copy see ComputeTensorData/CPUTensorData for info

Tensor<T>

Represents data in a multidimensional array-like structure.

Ownership and lifetime:

  • Disposed needs to be called on the main thread.
  • Ownership is always to the owner of the object.

Data Representation:

  • TensorShape represents the data layout of the tensor
  • Data is held by a tensorData (ITensorData) which can be on a given backend
  • Data is stored in a flattened row major format
  • Data can be pending (ie computation is being done in parallel)
    • call CompleteAllPendingOperations for a blocking call to finish computing the tensor's data Data can be in a non readable type (GPU/NPU)
    • Call CompleteAllPendingOperations to finish computing the tensor's data
    • Call ReadbackAndClone or ReadBackAndCloneAsync to allow reading the tensor's data

Data manipulation

  • ToReadOnlyArray returns a copy of the tensor's data
  • dataOnBackend can be manipulated directly to avoid a unnecessary copy see ComputeTensorData/CPUTensorData for info

TextureConverter

Provides methods for converting between textures and tensors.

TextureTensorData

Represents the data storage for a 'Tensor' as a render texture, for backends that use GPU pixel shaders.

Sentis packs the tensor data into the pixels of an RGBA float4 texture.

Sentis chooses a single tensor dimension as the blocked axis, across which data is chunked in float4 blocks.

Tensor dimensions don't map directly to texture dimensions. Sentis creates the texture with dimensions large enough to fit all the data and pixel shaders index the data based on both the tensor and texture dimensions (see example below).

Worker

Represents a worker that allows you to execute neural networks (models).

Worker abstracts implementation details on different hardware devices such as the CPU and the GPU. Worker lets you do the following:

  • Specify inputs.
  • Schedule the work.
  • Get outputs.

Internally, Worker translates the neural network from a Model into a set of operations, then sends the operations to the hardware device for asynchronous execution.

Structs

DynamicTensorShape

Represents the shape of an input tensor, or the predicted shape of a tensor before Sentis executes.

Model.Input

Represents an input to a model.

Model.Output

Represents an output to a model.

ONNXModelMetadata

Contains additional metadata about the ONNX model, stored in the ONNX file.

TensorShape

Represents the shape of a tensor.

TextureTransform

Represents settings for converting between textures and tensors.

Create an instance of TextureTransform using the constructor, then use the TextureTransform object as a parameter in TextureConverter methods.

For example: TextureTransform settings = new TextureTransform().SetDimensions(256, 256, 4).SetTensorLayout(TensorLayout.NHWC);

Interfaces

IONNXMetadataImportCallbackReceiver

Implement this interface to receive model metadata during ONNX import.

ITensorData

An interface that represents a device-dependent representation of the data in a tensor.

Enums

BackendType

Types of backend that Sentis uses to execute a neural network.

Channel

Set the position of each color channel in a texture tensor.

ChannelSwizzle

Options for the order of the color channels in a texture tensor.

CoordOrigin

Options for the position of the origin (0, 0) in the x and y dimensions of a texture tensor.

DataType

Options for the data type of a Tensor.

DeviceType

Types of devices that Sentis uses to execute a neural network.

QuantizationType

Type of quantization for float weights.

TensorLayout

Options for the dimension layout in a texture tensor.