C++ API¶

Abstruct¶

This chapter describes the C++ API

There is no detailed documentation available at this time, and the information to get an overview for readers is provided.

The BinaryBrain code has the name bb in the namespace.

Model class¶

Basic class¶

All layers are generated by deriving from the Model class.

Model class¶

Although abstract classes cannot be generated directly, they are the basis for each layer and define operations. It has the following methods.

SendCommand(): send command string to layer
GetClassName(): get class name
set instance name.: set instance name
GetParameters(): get parameters for optimizer
GetGradients(): get gradients for optimizer
SetInputShape(): set shape of the input-data
Forward(): run forward
Backward(): run backward
PrintInfo(): print structure of network

Activation layers¶

Binarize class¶

forward operation is binarize.backward operation is hard-tanh.

ReLU class¶

ReLU function.If you call SendCommand with “binary true”, it behaves as a Binarize layer

Sigmoid class¶

Sigmoid function.If you call SendCommand with “binary true”, it behaves as a Binarize layer

Operation layers¶

SparseLutN class¶

Layer to learn LUT using original model

StochasticLutN class¶

This is a layer to train the part corresponding to LUT of LUT-Network based on Stochastic model.Although Stochastic binary data is limited to learning to objects that have Stochastic properties, it is possible to train faster than SparseLut can.

MicroMlp class¶

This is a layer that uses the perceptron to learn the part corresponding to the LUT of the LUT-Network. The interior consists of three layers: MicroMlpAffine + BatchNormalization + Activation layer. Activation layer defaults to ReLU However, it can be changed with the template argument.

MicroMlpAffine class¶

It is a component of MicroMlp, a small MLP (Multi Layer Perceptron) aggregate with loose coupling with a limited number of inputs such as 6 and a hidden layer inside. The number of inputs and the number of hidden layers can be changed with template arguments.

DenseAffine class¶

Dense affine operation

BatchNormalization class¶

BatchNormalization operation

MaxPooling class¶

MaxPooling operation

LutLayer class¶

LUT Layer

BinaryLutN class¶

binary LUT operation (like FPGA)

Utility classes¶

Sequential class¶

constructer for sequential network

LoweringConvolution class¶

Lowering layer

Lowering for convolution

ConvolutionIm2 class¶

Im2Col operation layer

ConvolutionCol2Im class¶

Col2Im operation layer

BinaryModulation class¶

Binary modulation wrapper. (add RealToBinary(preprocess) and BinaryToReal(postprocess))

RealToBinary class¶

Binarize with modulation.

BinaryToReal class¶

Binary to Real with integration

Other classes¶

Loss function classes¶

LossSoftmaxCrossEntropy class¶

Softmax-CrossEntropy operation

LossMeanSquaredError class¶

calclate Mean Squared Error

Metrics classes¶

MetricsCategoricalAccuracy class¶

calcrate CategoricalAccuracy for Metrics.

MetricsMeanSquaredError class¶

calclate Mean Squared Error

Optimizer¶

OptimizerSgd class¶

stochastic gradient descent

OptimizerAdam class¶

Adam

Assist for run¶

Runner class¶

Assists in executing fitting and evaluation of the built model.

Containers¶

Tensor class¶

Container for N-dimensional array

Variables class¶

Container for Tensors

FrameBuffer class¶

Container for Tensors with frames

functions¶

export RTL for FPGA¶

ExportVerilog_LutLayers function¶

output verilog-RTL from LutLayer

ExportVerilog_LutCnnLayersAxi4s function¶

output verilog-RTL(AXI4-Stream I/F) from CNN-Layer”