org.apache.ignite.ml.nn

Class MultilayerPerceptron

java.lang.Object

org.apache.ignite.ml.nn.MultilayerPerceptron

All Implemented Interfaces:

Serializable, AutoCloseable, IgniteModel<Matrix,Matrix>, Model<Matrix,Matrix>, Parametrized<MultilayerPerceptron>, SmoothParametrized<MultilayerPerceptron>

public final class MultilayerPerceptron
extends Object
implements SmoothParametrized<MultilayerPerceptron>

Class encapsulating logic of multilayer perceptron.

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
MultilayerPerceptron(MLPArchitecture arch) Construct MLP from given architecture.
MultilayerPerceptron(MLPArchitecture arch, MLPInitializer initializer) Construct MLP from given architecture and parameters initializer.

Method Summary

Modifier and Type	Method and Description
MultilayerPerceptron	add(MultilayerPerceptron above) Create MLP where this MLP output is fed as input to added MLP.
MLPArchitecture	architecture() Get architecture of this MLP.
double	bias(int layerIdx, int neuronIdx) Get the bias of given neuron in given layer.
Vector	biases(int layerIdx) Get biases of layer with given index.
MLPState	computeState(Matrix val) Perform forward pass and return state of outputs of each layer.
Vector	differentiateByParameters(IgniteFunction<Vector,IgniteDifferentiableVectorToDoubleFunction> loss, Matrix inputsBatch, Matrix truthBatch) Compose function in the following way: feed output of this model as input to second argument to loss function.
Matrix	forwardPass(Matrix val, MLPState state, boolean writeState) Perform forward pass and if needed write state of outputs of each layer.
boolean	hasBiases(int layerIdx) Checks if layer with given index has biases.
int	layersCount() Get count of layers in this MLP.
Vector	parameters() Get parameters vector.
int	parametersCount() Get count of parameters of this model.
Matrix	predict(Matrix val) Makes a prediction for the given objects.
MultilayerPerceptron	setBias(int layerIdx, int neuronIdx, double val) Set the bias of given neuron in given layer.
MultilayerPerceptron	setBiases(int layerIdx, Vector bias) Sets the biases of layer with a given index.
MultilayerPerceptron	setParameters(Vector vector) Set parameters.
MultilayerPerceptron	setWeight(int layerIdx, int fromNeuron, int toNeuron, double val) Set the weight of neuron with given index in previous layer to neuron with given index in given layer.
MultilayerPerceptron	setWeights(int layerIdx, Matrix weights) Sets the weighs of layer with a given index.
String	toString()
String	toString(boolean pretty)
double	weight(int layerIdx, int fromNeuron, int toNeuron) Get the weight of neuron with given index in previous layer to neuron with given index in given layer.
Matrix	weights(int layerIdx) Get weights of layer with given index.
default M	withParameters(Vector vector) Return new model with given parameters vector.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Methods inherited from interface org.apache.ignite.ml.IgniteModel

andBefore, andThen, andThen, close, combine

Constructor Detail

MultilayerPerceptron

public MultilayerPerceptron(MLPArchitecture arch,
                            MLPInitializer initializer)

Construct MLP from given architecture and parameters initializer.

Parameters:

arch - Architecture.

initializer - Parameters initializer.

MultilayerPerceptron

public MultilayerPerceptron(MLPArchitecture arch)

Construct MLP from given architecture.

Parameters:

arch - Architecture.

Method Detail

computeState

public MLPState computeState(Matrix val)

Perform forward pass and return state of outputs of each layer.

Parameters:

val - Value to perform computation on.

Returns:

MLP state after computation.

forwardPass

public Matrix forwardPass(Matrix val,
                          MLPState state,
                          boolean writeState)

Perform forward pass and if needed write state of outputs of each layer.

Parameters:

val - Value to perform computation on.

state - State object to write state into.

writeState - Flag indicating need to write state.

predict

public Matrix predict(Matrix val)

Makes a prediction for the given objects.

Specified by:

predict in interface Model<Matrix,Matrix>

Parameters:

val - Matrix containing objects.

Returns:

Matrix with predicted vectors.

add

public MultilayerPerceptron add(MultilayerPerceptron above)

Create MLP where this MLP output is fed as input to added MLP.

Parameters:

above - Added MLP.

Returns:

New MLP where this MLP output is fed as input to added MLP.

weights

public Matrix weights(int layerIdx)

Get weights of layer with given index. Proper indexes are in [1, layersCount).

Parameters:

layerIdx - Layer index.

Returns:

Weights of layer with given index.

biases

public Vector biases(int layerIdx)

Get biases of layer with given index. Proper indexes are in [1, layersCount).

Parameters:

layerIdx - Layer index.

Returns:

Biases of layer with given index.

hasBiases

public boolean hasBiases(int layerIdx)

Checks if layer with given index has biases.

Parameters:

layerIdx - Layer index.

Returns:

Value of predicate 'layer with layerIdx has biases'.

setBiases

public MultilayerPerceptron setBiases(int layerIdx,
                                      Vector bias)

Sets the biases of layer with a given index.

Parameters:

layerIdx - Layer index.

bias - New values for biases.

Returns:

This MLP with updated biases.

setBias

public MultilayerPerceptron setBias(int layerIdx,
                                    int neuronIdx,
                                    double val)

Set the bias of given neuron in given layer.

Parameters:

layerIdx - Layer index.

neuronIdx - Neuron index.

val - New value of bias.

Returns:

This MLP with updated biases.

bias

public double bias(int layerIdx,
                   int neuronIdx)

Get the bias of given neuron in given layer.

Parameters:

layerIdx - Layer index.

neuronIdx - Neuron index.

Returns:

Bias with specified coordinates.

setWeights

public MultilayerPerceptron setWeights(int layerIdx,
                                       Matrix weights)

Sets the weighs of layer with a given index.

Parameters:

layerIdx - Layer index.

weights - New values for weights.

Returns:

This MLP with updated weights.

setWeight

public MultilayerPerceptron setWeight(int layerIdx,
                                      int fromNeuron,
                                      int toNeuron,
                                      double val)

Set the weight of neuron with given index in previous layer to neuron with given index in given layer.

Parameters:

layerIdx - Layer index.

fromNeuron - Neuron index in previous layer.

toNeuron - Neuron index in current layer.

val - New value of weight.

Returns:

This MLP with updated weights.

weight

public double weight(int layerIdx,
                     int fromNeuron,
                     int toNeuron)

Get the weight of neuron with given index in previous layer to neuron with given index in given layer.

Parameters:

layerIdx - Layer index.

fromNeuron - Neuron index in previous layer.

toNeuron - Neuron index in current layer.

Returns:

Weight with specified coordinates.

layersCount

public int layersCount()

Get count of layers in this MLP.

Returns:

Count of layers in this MLP.

architecture

public MLPArchitecture architecture()

Get architecture of this MLP.

Returns:

Architecture of this MLP.

differentiateByParameters

public Vector differentiateByParameters(IgniteFunction<Vector,IgniteDifferentiableVectorToDoubleFunction> loss,
                                        Matrix inputsBatch,
                                        Matrix truthBatch)

Compose function in the following way: feed output of this model as input to second argument to loss function. After that we have a function g of three arguments: input, ground truth, parameters. If we consider function h(w) = 1 / M sum_{i=1}^{M} g(w, input_i, groundTruth_i), where M is number of entries in batch, we get function of one argument: parameters vector w. This function is being differentiated.

Specified by:

differentiateByParameters in interface SmoothParametrized<MultilayerPerceptron>

Parameters:

loss - Loss function.

inputsBatch - Batch of inputs.

truthBatch - Batch of ground truths.

Returns:

Gradient of h at current point in parameters space.

parameters

public Vector parameters()

Get parameters vector.

Returns:

Parameters vector.

setParameters

public MultilayerPerceptron setParameters(Vector vector)

Set parameters.

Parameters:

vector - Parameters vector.

parametersCount

public int parametersCount()

Get count of parameters of this model.

Returns:

Count of parameters of this model.

toString

public String toString()

Overrides:

toString in class Object

toString

public String toString(boolean pretty)

Specified by:

toString in interface IgniteModel<Matrix,Matrix>

Parameters:

pretty - Use pretty mode.

withParameters

public M withParameters(Vector vector)

Return new model with given parameters vector.

Parameters:

vector - Parameters vector.