FlexNN::NeuralNetwork Class Reference

Linear stack of heterogeneous layers (vector<Layers::Layer>). More...

#include <FlexNN.h>

Public Member Functions
	NeuralNetwork (const std::vector< Layers::Layer > &layers)
	Construct from new modular layers (preferred).
	NeuralNetwork (std::vector< Layers::Layer > &&layers)
	NeuralNetwork (const std::vector< Layer > &oldLayers)
	Deprecated: construct from legacy FlexNN::Layer (Dense-only).
void	train (const Eigen::MatrixXd &input, const Eigen::VectorXd &target, double learningRate, int epochs)
	Train the network on (X, Y) with SGD.
double	accuracy (const Eigen::MatrixXd &X, const Eigen::VectorXd &Y)
	Accuracy on (X, Y) — argmax per column.
Eigen::MatrixXd	predict (const Eigen::MatrixXd &input)
	Predict — forward pass, returns activation of last layer.
const std::vector< Layers::Layer > &	layers () const noexcept
	Direct access to layers (for tests / ModelIO).
std::vector< Layers::Layer > &	layers () noexcept

Detailed Description

Linear stack of heterogeneous layers (vector<Layers::Layer>).

Currently holds only Layers::Dense (PR-04). Later PRs expand the variant to Conv1D, BatchNorm1D, MaxPool1D, AvgPool1D without changing this header — only include/layers/Layer.hpp's variant list grows.

Old code NeuralNetwork({Layer(784,64,"relu"), Layer(64,10,"softmax")}) still compiles via the deprecated converting constructor that maps each legacy Layer to Layers::Dense via Activations::try_parse.

Constructor & Destructor Documentation

NeuralNetwork() [1/3]

FlexNN::NeuralNetwork::NeuralNetwork ( const std::vector< Layers::Layer > &  layers )

inline

Construct from new modular layers (preferred).

Parameters

layers

Heterogeneous stack — each element is a Layers::Layer wrapping a concrete Dense/Conv1D/... Example: NeuralNetwork({Layers::Dense(784,64, Activations::ReLU)})

NeuralNetwork() [2/3]

FlexNN::NeuralNetwork::NeuralNetwork ( std::vector< Layers::Layer > &&  layers )

inline

NeuralNetwork() [3/3]

FlexNN::NeuralNetwork::NeuralNetwork

(

const std::vector< Layer > &

oldLayers

)

Deprecated: construct from legacy FlexNN::Layer (Dense-only).

Converts each legacy Layer (which stores std::string activationFunction and MatrixXd W/VectorXd b) to Layers::Dense via Activations::try_parse. On unknown activation string, uses Activation::None.

Kept so src/main.cpp (which still uses Layer(..., "relu")) builds until PR-09 migrates it to Layers::Dense. New code should not use this.

Member Function Documentation

accuracy()

double FlexNN::NeuralNetwork::accuracy	(	const Eigen::MatrixXd &	X,
		const Eigen::VectorXd &	Y
	)

Accuracy on (X, Y) — argmax per column.

layers() [1/2]

const std::vector< Layers::Layer > & FlexNN::NeuralNetwork::layers ( ) const

inlinenoexcept

Direct access to layers (for tests / ModelIO).

layers() [2/2]

std::vector< Layers::Layer > & FlexNN::NeuralNetwork::layers ( )

inlinenoexcept

predict()

Eigen::MatrixXd FlexNN::NeuralNetwork::predict ( const Eigen::MatrixXd &  input )

inline

Predict — forward pass, returns activation of last layer.

train()

void FlexNN::NeuralNetwork::train	(	const Eigen::MatrixXd &	input,
		const Eigen::VectorXd &	target,
		double	learningRate,
		int	epochs
	)

Train the network on (X, Y) with SGD.

X is [features × batch] (see Utility.h), Y is label vector (converted to one-hot inside). Uses the interleaved forward layout [input, Z0,A0, Z1,A1, ...] and backward that fuses Softmax+CE as dZ_last = (A - Y)/m.

---

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

• include/FlexNN.h
• lib/FlexNN.cpp