Network Class Reference

Represents a neural network. More...

#include <network.h>

Public Member Functions
	Network (size_t inputSize, bool useGPU=true, long long seed=NO_SEED)
	Construct a new network. More...
void	add (size_t numNeurons, const std::string &activation="linear")
	Add a new layer to the network. More...
sTensor	forward (const sTensor &batch)
	Forward-propagate a batch through the network. More...
void	train (sTensor &X, sTensor &y, int epochs, size_t batchSize, float learningRate, Loss *loss, std::vector< Metric * > &metrics)
	Train the network. More...

Private Member Functions
void	processEpoch (std::vector< sTensor > &batches, std::vector< sTensor > &targets, std::vector< sTensor > &targetsOnHost, float learningRate, Loss *loss, std::vector< Metric * > &metrics)
	Trains the model on a single epoch. More...

Private Attributes
DataLocation	location
	The location of the network. More...
std::vector< Layer >	layers
	List of network layers.
long long	seed
	Seed used for random initialization.
size_t	previousSize
	Keeps track of the size of the previous layer. More...

Detailed Description

Represents a neural network.

Examples

MNIST, and Titanic.

Constructor & Destructor Documentation

Network()

Network::Network ( size_t  inputSize, bool  useGPU = true, long long  seed = NO_SEED  )

explicit

Construct a new network.

The constructed network can use GPU acceleration if a GPU and CUDA are available, and the useGPU parameter is set to true.

Parameters

inputSize	The number of inputs to the neural network.
useGPU	Boolean to specify whether the network should use GPU acceleration.
seed	Seed that should be used for random initialization of the network.

Member Function Documentation

add()

void Network::add	(	size_t	numNeurons,
		const std::string &	activation = "linear"
	)

Add a new layer to the network.

Three activation functions can be used: Linear, ReLU and Sigmoid. The activation can be selected by specifying the activation parameter, using one of the strings: "linear", "relu" or "sigmoid". If any other string is specified, Linear activation will be used.

Parameters

numNeurons	The number of neurons the new layer should contain.
activation	The activation function to use. Can be "linear", "relu" or "sigmoid".

Examples

MNIST, and Titanic.

forward()

sTensor Network::forward

(

const sTensor &

batch

)

Forward-propagate a batch through the network.

The samples should be aligned along the first axis.

Parameters

batch

The batch to propagate.

Returns

The pointer to the output of the network. This returns Layer::aMatrix of the last layer.

processEpoch()

void Network::processEpoch ( std::vector< sTensor > &  batches, std::vector< sTensor > &  targets, std::vector< sTensor > &  targetsOnHost, float  learningRate, Loss *  loss, std::vector< Metric * > &  metrics  )

private

Trains the model on a single epoch.

Helper method used in Network::train(). The method makes use of targetsOnHost, which are the target batches stored on host. This is for performance reasons as some of the metrics require the input matrices to be located on host.

Parameters

batches	The list of batches to process. These have been split in Network::train() method.
targets	The list of targets to process. These have been split in Network::train() method.
targetsOnHost	The list of targets to processed but stored on host.
learningRate	The learning rate used during training.
loss	The loss function to use.
metrics	The list of metrics to compute aside from the loss function.

train()

void Network::train	(	sTensor &	X,
		sTensor &	y,
		int	epochs,
		size_t	batchSize,
		float	learningRate,
		Loss *	loss,
		std::vector< Metric * > &	metrics
	)

Train the network.

Both X and y should have the data samples aligned on the first axis. Each row in X should be aligned with the corresponding row in y.

Parameters

X	The data to train the network on.
y	The targets of the network.
epochs	The number of epochs to train the network for.
batchSize	The size of the batch.
learningRate	The learning rate of the algorithm.
loss	The loss function to use.
metrics	The list of metrics to compute aside from the loss function.

Examples

MNIST, and Titanic.

Member Data Documentation

location

DataLocation Network::location

private

The location of the network.

Specifies the location of all the data used by the network. See DataLocation for more info.

previousSize

size_t Network::previousSize

private

Keeps track of the size of the previous layer.

Layers need to know the size of the input passed to them. This is achieved using this variable, that keeps track of the size of the previous layer (or size of the input if in 1st layer). In this way, layers can pre-allocate required space during initialization.

---

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

• include/network.h
• src/ann/network.cpp