#include <CscNeuralNetwork.h>

Inheritance diagram for conscience_core::ai::nn::CscNeuralNetwork:

Classes
struct	DataSet

struct	NeuralNetworkHyperparameters

struct	NeuralNetworkShape

struct	ResultTraining

struct	TrainingHyperparameters
	Structure representing the hyperparameters for model training. More...

Public Member Functions
	CscNeuralNetwork (CscNeuralNetworkType neuralNetworkType, NeuralNetworkHyperparameters neuralNetworkHyperparametersInit, TrainingHyperparameters trainingHyperparametersInit)

	CscNeuralNetwork (CscNeuralNetworkType neuralNetworkType, NeuralNetworkHyperparameters neuralNetworkHyperparametersInit, TrainingHyperparameters trainingHyperparametersInit, float weightsLoadedFromModel, float biasesLoadedFromModel)

virtual	~CscNeuralNetwork ()

CscNeuralNetworkType	getNeuralNetworkType () const

CscNeuralNetwork::NeuralNetworkHyperparameters *	getNeuralNetworkHyperparameters () const

CscNeuralNetwork::TrainingHyperparameters *	getTrainingHyperparameters () const

virtual ResultTraining	trainNeuralNetwork (const int nbEpochs, const int nbMiniBatchs, DataSet trainingDataSet, DataSet validationDataSet, const bool debugMode=false)=0
	Trains the neural network using the provided input and target data. More...

Static Public Member Functions
static float ***	adaptDataForMiniBatches (float **dataToTransform, const int minibatchSize, const int dataSetSize)
	Prepares the dataset for minibatch processing by dividing it into smaller batches. More...

static void	normalizeDataAutoRange (float **data, size_t size, int nbInputsOrOutputs, CscNormalizationType normalizationType)

static void	normalizeDataCustomRange0to1 (float *data, size_t size, int nbInputsOrOutputs, float minValue, float *maxValue)

static void	normalizeDataCustomRangeMinus1to1 (float *data, size_t size, int nbInputsOrOutputs, float minValue, float *maxValue)

static void	normalizeData (float **data, size_t size, int rowData, float newMin, float newMax, float minValue, float maxValue)

static void	deleteAllDataMiniBatches (float ***dataMiniBatches, const int nbMiniBatches, const int nbDataPerMinibatch)
	Frees the memory allocated for the mini-batches of data. More...

static void	initializeWeights (float weightsToInitialize, int size, CscInitializationWeightsStrategy strategy, int fanIn, int fanOut=0, float weightsLoaded=nullptr)
	Initializes the weights of the neural network according to a specified initialization strategy. More...

static void	initializeBias (float biasToInitialize, int size, CscInitializationBiasStrategy strategy, float biasLoaded=nullptr)

Protected Member Functions
virtual void	createNeuralNetwork (CscActivationFunction hiddenLayersFunctionActivation, CscActivationFunction outputLayerFunctionActivation, CscLossFunction lossFunction, CscInitializationWeightsStrategy initilizationStrategy, CscInitializationBiasStrategy initializationBiasStrategy, bool withCostFunction)=0

void	shuffleMinibatchData (float *minibatchDataInputs, float *minibatchDataTargets, const int numberOfBatches, const int minibatchSize)
	Shuffles the data within each minibatch as well as between minibatches. More...

Static Protected Member Functions
static float	calculateStddevHe (int fanIn)

static float	calculateLimitGlorot (int fanIn, int fanOut)

static float	calculateStddevGlorot (int fanIn, int fanOut)

static float	calculateStddevLeCun (int fanIn)

static void	initNormal (float *weights, int size, mt19937 &gen, double stddev)

static void	initUniform (float *weights, int size, mt19937 &gen, float limit)

Protected Attributes
CscNeuralNetworkType	neuralNetworkType

TrainingHyperparameters *	trainingHyperparameters = nullptr

NeuralNetworkHyperparameters *	neuralNetworkHyperparameters = nullptr

unique_ptr< CscLogger >	logger

Constructor & Destructor Documentation

◆ CscNeuralNetwork() [1/2]

conscience_core::ai::nn::CscNeuralNetwork::CscNeuralNetwork	(	CscNeuralNetworkType	neuralNetworkType,
		NeuralNetworkHyperparameters *	neuralNetworkHyperparametersInit,
		TrainingHyperparameters *	trainingHyperparametersInit
	)

◆ CscNeuralNetwork() [2/2]

conscience_core::ai::nn::CscNeuralNetwork::CscNeuralNetwork	(	CscNeuralNetworkType	neuralNetworkType,
		NeuralNetworkHyperparameters *	neuralNetworkHyperparametersInit,
		TrainingHyperparameters *	trainingHyperparametersInit,
		float **	weightsLoadedFromModel,
		float **	biasesLoadedFromModel
	)

◆ ~CscNeuralNetwork()

conscience_core::ai::nn::CscNeuralNetwork::~CscNeuralNetwork ( )

virtual

Member Function Documentation

◆ adaptDataForMiniBatches()

float *** conscience_core::ai::nn::CscNeuralNetwork::adaptDataForMiniBatches	(	float **	dataToTransform,
		const int	minibatchSize,
		const int	dataSetSize
	)

static

Prepares the dataset for minibatch processing by dividing it into smaller batches.

This method organizes input data into a structure suitable for minibatch training. Each minibatch consists of a subset of the input data, allowing for more efficient memory usage and potentially faster convergence of the neural network due to more frequent updates.

Parameters

dataToTransform	Pointer to the original dataset to be divided into minibatches.
minibatchSize	The size of each minibatch.
dataSetSize	The total number of samples in the dataset.

Returns: A 3D array where the first dimension corresponds to the minibatch index, the second dimension indexes into a specific sample within a minibatch, and the third dimension represents the data of that sample.

Warning: The new tabs created must be deleted after they have been used for training.

◆ calculateLimitGlorot()

float conscience_core::ai::nn::CscNeuralNetwork::calculateLimitGlorot	(	int	fanIn,
		int	fanOut
	)

staticprotected

◆ calculateStddevGlorot()

float conscience_core::ai::nn::CscNeuralNetwork::calculateStddevGlorot	(	int	fanIn,
		int	fanOut
	)

staticprotected

◆ calculateStddevHe()

float conscience_core::ai::nn::CscNeuralNetwork::calculateStddevHe ( int fanIn )

staticprotected

◆ calculateStddevLeCun()

float conscience_core::ai::nn::CscNeuralNetwork::calculateStddevLeCun ( int fanIn )

staticprotected

◆ createNeuralNetwork()

virtual void conscience_core::ai::nn::CscNeuralNetwork::createNeuralNetwork	(	CscActivationFunction	hiddenLayersFunctionActivation,
		CscActivationFunction	outputLayerFunctionActivation,
		CscLossFunction	lossFunction,
		CscInitializationWeightsStrategy	initilizationStrategy,
		CscInitializationBiasStrategy	initializationBiasStrategy,
		bool	withCostFunction
	)

protectedpure virtual

◆ deleteAllDataMiniBatches()

void conscience_core::ai::nn::CscNeuralNetwork::deleteAllDataMiniBatches	(	float ***	dataMiniBatches,
		const int	nbMiniBatches,
		const int	nbDataPerMinibatch
	)

static

Frees the memory allocated for the mini-batches of data.

This function iterates through each mini-batch of data, and for each mini-batch, it iterates through each data point. It frees the memory allocated for each data point and then frees the memory allocated for the mini-batch itself. Finally, it frees the memory allocated for the array of pointers to the mini-batches.

Parameters

dataMiniBatches	A pointer to an array of pointers to arrays of pointers. Each element in the first array represents a mini-batch. Each element in a mini-batch represents a single data point. The memory for each data point, each mini-batch, and the array of mini-batches itself is dynamically allocated and needs to be freed.
nbMiniBatches	The number of mini-batches in `dataMiniBatches`. This is the size of the first array.
nbDataPerMinibatch	The number of data points in each mini-batch. This is the size of each second-level array.

◆ getNeuralNetworkHyperparameters()

CscNeuralNetwork::NeuralNetworkHyperparameters * conscience_core::ai::nn::CscNeuralNetwork::getNeuralNetworkHyperparameters ( ) const

◆ getNeuralNetworkType()

CscNeuralNetworkType conscience_core::ai::nn::CscNeuralNetwork::getNeuralNetworkType ( ) const

◆ getTrainingHyperparameters()

CscNeuralNetwork::TrainingHyperparameters * conscience_core::ai::nn::CscNeuralNetwork::getTrainingHyperparameters ( ) const

◆ initializeBias()

void conscience_core::ai::nn::CscNeuralNetwork::initializeBias	(	float *	biasToInitialize,
		int	size,
		CscInitializationBiasStrategy	strategy,
		float *	biasLoaded = `nullptr`
	)

static

◆ initializeWeights()

void conscience_core::ai::nn::CscNeuralNetwork::initializeWeights	(	float *	weightsToInitialize,
		int	size,
		CscInitializationWeightsStrategy	strategy,
		int	fanIn,
		int	fanOut = `0`,
		float *	weightsLoaded = `nullptr`
	)

static

Initializes the weights of the neural network according to a specified initialization strategy.

This method initializes the weights of the neural network using one of four supported initialization strategies: He, GlorotUniform, GlorotNormal, or LeCunNormal. Each strategy has its characteristics and is chosen based on the specific needs of the network's architecture. Proper weight initialization is crucial for the successful learning and convergence of the neural network.

Parameters

weightsToInitialize	Pointer to the array of weights to be initialized. This array must be pre-allocated with enough size to hold all the weights.
size	The number of weights to initialize in the array.
strategy	The `CscInitializationWeightsStrategy` enumeration specifying the initialization strategy to use for the weights.
fanIn	The number of incoming units to the node for which the weights are being initialized, used in the calculation for some strategies.
fanOut	The number of outgoing units from the node for which the weights are being initialized, used in the calculation for some strategies.

Note: The function uses specific distributions based on the chosen strategy to initialize the weights. For He, a normal distribution is used with a standard deviation calculated based on fanIn. For GlorotUniform, a uniform distribution is used with limits calculated based on fanIn and fanOut. GlorotNormal uses a normal distribution similar to GlorotUniform but with a different standard deviation. LeCunNormal uses a normal distribution with a standard deviation calculated solely based on fanIn.

◆ initNormal()

void conscience_core::ai::nn::CscNeuralNetwork::initNormal	(	float *	weights,
		int	size,
		mt19937 &	gen,
		double	stddev
	)

staticprotected

◆ initUniform()

void conscience_core::ai::nn::CscNeuralNetwork::initUniform	(	float *	weights,
		int	size,
		mt19937 &	gen,
		float	limit
	)

staticprotected

◆ normalizeData()

void conscience_core::ai::nn::CscNeuralNetwork::normalizeData	(	float **	data,
		size_t	size,
		int	rowData,
		float	newMin,
		float	newMax,
		float	minValue,
		float	maxValue
	)

static

Normalizes the elements of a given float array to a specified new range.

This method adjusts the values within the array so that they fit within a new range, defined by newMin and newMax. The normalization is done based on a given range of values in the array, specified by the parameters minValue and maxValue. The method scales each element of the array from its original range (minValue to maxValue) to the new range (newMin to newMax). If the array's size is zero, or if newMin is equal to or greater than newMax, the method logs an error and returns without modifying the array.

Parameters

data	Pointer to the first element of the array of float values to be normalized.
size	The number of elements in the array.
newMin	The minimum value of the new range to which the data should be normalized.
newMax	The maximum value of the new range to which the data should be normalized.
minValue	The minimum value in the original range of the data. This parameter is used directly in the normalization calculation and is not modified by the method.
maxValue	The maximum value in the original range of the data. This parameter is used directly in the normalization calculation and is not modified by the method.

◆ normalizeDataAutoRange()

void conscience_core::ai::nn::CscNeuralNetwork::normalizeDataAutoRange	(	float **	data,
		size_t	size,
		int	nbInputsOrOutputs,
		CscNormalizationType	normalizationType
	)

static

◆ normalizeDataCustomRange0to1()

void conscience_core::ai::nn::CscNeuralNetwork::normalizeDataCustomRange0to1	(	float **	data,
		size_t	size,
		int	nbInputsOrOutputs,
		float *	minValue,
		float *	maxValue
	)

static

◆ normalizeDataCustomRangeMinus1to1()

void conscience_core::ai::nn::CscNeuralNetwork::normalizeDataCustomRangeMinus1to1	(	float **	data,
		size_t	size,
		int	nbInputsOrOutputs,
		float *	minValue,
		float *	maxValue
	)

static

◆ shuffleMinibatchData()

void conscience_core::ai::nn::CscNeuralNetwork::shuffleMinibatchData	(	float ***	minibatchDataInputs,
		float ***	minibatchDataTargets,
		const int	numberOfBatches,
		const int	minibatchSize
	)

protected

Shuffles the data within each minibatch as well as between minibatches.

This method performs an in-place shuffle of the minibatch data, both the inputs and the targets. The shuffling is performed within each minibatch and across minibatches to ensure that the model does not learn any unintended patterns from the order of the data.

Parameters

minibatchDataInputs	A 3D array containing the input data divided into minibatches.
minibatchDataTargets	A 3D array containing the target data divided into minibatches.
numberOfBatches	The total number of minibatches.
minibatchSize	The size of each minibatch.

◆ trainNeuralNetwork()

virtual ResultTraining conscience_core::ai::nn::CscNeuralNetwork::trainNeuralNetwork	(	const int	nbEpochs,
		const int	nbMiniBatchs,
		DataSet *	trainingDataSet,
		DataSet *	validationDataSet,
		const bool	debugMode = `false`
	)

pure virtual

Trains the neural network using the provided input and target data.

This method trains the neural network over a specified number of epochs and mini-batches, using the TensorFlow C API. It initializes the session and model parameters, shuffles the data for each epoch (except the first one), processes each mini-batch by executing the graph, and updates the weights and biases according to the cost function. It also provides optional debugging information, including logging the progress of training and the change in weights and biases.

Parameters

nbEpochs	The number of epochs to train the neural network for.
nbMiniBatchs	The number of mini-batches to divide the input data into.
debugMode	A boolean flag that, when true, enables logging of debugging information during training.

Returns: The mean of the loss/cost function over the mini-batches in the final epoch of training.

Note: This function requires TensorFlow C API setup and assumes that the session, graph, and status are correctly initialized and available globally or within the class.

Warning: The method does not handle the creation and destruction of the TensorFlow session, graph, and status objects. These should be appropriately managed outside this method.

Attention: This method assumes that the neural network, its variables (weights and biases), and the cost function are already defined in the TensorFlow graph.

Implemented in conscience_core::ai::nn::CscNeuralNetworkTensorFlow.

Member Data Documentation

◆ logger

unique_ptr<CscLogger> conscience_core::ai::nn::CscNeuralNetwork::logger

protected

◆ neuralNetworkHyperparameters

NeuralNetworkHyperparameters* conscience_core::ai::nn::CscNeuralNetwork::neuralNetworkHyperparameters = nullptr

protected

◆ neuralNetworkType

CscNeuralNetworkType conscience_core::ai::nn::CscNeuralNetwork::neuralNetworkType

protected

◆ trainingHyperparameters

TrainingHyperparameters* conscience_core::ai::nn::CscNeuralNetwork::trainingHyperparameters = nullptr

protected

The documentation for this class was generated from the following files:

ArtificialIntelligence/CscNeuralNetwork.h
ArtificialIntelligence/CscNeuralNetwork.cpp

Conscience Core

Classes

Public Member Functions

Static Public Member Functions

Protected Member Functions

Static Protected Member Functions

Protected Attributes

Constructor & Destructor Documentation

◆ CscNeuralNetwork() [1/2]

◆ CscNeuralNetwork() [2/2]

◆ ~CscNeuralNetwork()

Member Function Documentation

◆ adaptDataForMiniBatches()

◆ calculateLimitGlorot()

◆ calculateStddevGlorot()

◆ calculateStddevHe()

◆ calculateStddevLeCun()

◆ createNeuralNetwork()

◆ deleteAllDataMiniBatches()

◆ getNeuralNetworkHyperparameters()

◆ getNeuralNetworkType()

◆ getTrainingHyperparameters()

◆ initializeBias()

◆ initializeWeights()

◆ initNormal()

◆ initUniform()

◆ normalizeData()

◆ normalizeDataAutoRange()

◆ normalizeDataCustomRange0to1()

◆ normalizeDataCustomRangeMinus1to1()

◆ shuffleMinibatchData()

◆ trainNeuralNetwork()

Member Data Documentation

◆ logger

◆ neuralNetworkHyperparameters

◆ neuralNetworkType

◆ trainingHyperparameters