FlexNN Namespace Reference

Namespace for the FlexNN neural network library. More...

Namespaces
namespace	Activations
namespace	Layers

Classes
struct	ExportOptions
	Options for export (currently version). More...
class	Layer
	Represents a single layer in a neural network. More...
struct	LayerHeader
struct	MainHeader
class	NeuralNetwork
	Linear stack of heterogeneous layers (vector<Layers::Layer>). More...
struct	Status
	Result of export/import. More...

Functions
Status	exportModel (const NeuralNetwork &net, const std::string &path, ExportOptions opts={})
	Export a trained NeuralNetwork to flat binary model.bin.
Status	importModel (NeuralNetwork &net, const std::string &path)
	Import a flat binary model.bin into a NeuralNetwork.
Eigen::MatrixXd	oneHotEncode (const Eigen::VectorXd &Y, int num_classes)
	One-hot encodes a vector of class labels.
void	readCSV_XY (const std::string &filename, Eigen::MatrixXd &X, Eigen::VectorXd &Y)
	Reads a CSV where first column is label, rest are features.
void	setRandomSeed (uint32_t seed)
	Set the global RNG seed for splitXY (deterministic splits).
uint32_t	getRandomSeed () noexcept
	Get the current global RNG seed.
std::vector< std::pair< Eigen::MatrixXd, Eigen::VectorXd > >	splitXY (const Eigen::MatrixXd &X, const Eigen::VectorXd &Y, const std::vector< double > &proportions)
	Splits dataset into multiple sets by proportions (uses global seed).
std::vector< std::pair< Eigen::MatrixXd, Eigen::VectorXd > >	splitXY (const Eigen::MatrixXd &X, const Eigen::VectorXd &Y, const std::vector< double > &proportions, uint32_t seed)
	Splits dataset with explicit seed (deterministic, no global state).

Detailed Description

Namespace for the FlexNN neural network library.

This namespace contains all the classes and functions related to the FlexNN library, including the NeuralNetwork class and Layer class. It provides a structured way to organize the library's components and avoid naming conflicts with other libraries.

Function Documentation

exportModel()

Status FlexNN::exportModel	(	const NeuralNetwork &	net,
		const std::string &	path,
		ExportOptions	opts = {}
	)

Export a trained NeuralNetwork to flat binary model.bin.

Validates each layer (e.g., Conv1D dilation==1, Softmax only on last layer, unknown DType, LeakyReLU alpha in (0,1)) and returns Status::Err on impossible export without writing. Otherwise writes main header (20B) + N*52B LayerHeaders + blobs (float32 LE, row-major) + file CRC32 and returns Ok. v2 writes ActivationParameters for LeakyReLU as aux.

Parameters

net	Network to export (weights are double, cast to float32)
path	Destination file (truncated if exists)
opts	Export options (formatVersion 1 or 2)

Returns

Status::Ok on success, Err with message on failure

getRandomSeed()

uint32_t FlexNN::getRandomSeed ( )

noexcept

Get the current global RNG seed.

Returns

Current seed (initially from random_device once)

importModel()

Status FlexNN::importModel	(	NeuralNetwork &	net,
		const std::string &	path
	)

Import a flat binary model.bin into a NeuralNetwork.

Validates magic, version, header_len, layer_count, header/file CRC32, offsets, and per-type counts. Reconstructs Layers::Dense / Conv1D with the same activation and float32 blobs cast back to double.

Parameters

net	Output network (overwritten on success, unchanged on failure)
path	Source file

Returns

Status::Ok on success, Err on failure

oneHotEncode()

Eigen::MatrixXd FlexNN::oneHotEncode	(	const Eigen::VectorXd &	Y,
		int	num_classes
	)

One-hot encodes a vector of class labels.

Parameters

Y	Input vector of class labels (size = batch)
num_classes	Number of unique classes

Returns

Matrix [num_classes × batch] where each column is one-hot

readCSV_XY()

void FlexNN::readCSV_XY	(	const std::string &	filename,
		Eigen::MatrixXd &	X,
		Eigen::VectorXd &	Y
	)

Reads a CSV where first column is label, rest are features.

Parameters

filename	Path to CSV
X	Output features [samples × features]
Y	Output labels [samples]

setRandomSeed()

void FlexNN::setRandomSeed

(

uint32_t

seed

)

Set the global RNG seed for splitXY (deterministic splits).

Thread-safe (atomic). The thread_local engine is reseeded on next splitXY call without explicit seed. For per-call determinism, use splitXY(X,Y,props,seed) instead.

Parameters

seed

New seed

splitXY() [1/2]

std::vector< std::pair< Eigen::MatrixXd, Eigen::VectorXd > > FlexNN::splitXY	(	const Eigen::MatrixXd &	X,
		const Eigen::VectorXd &	Y,
		const std::vector< double > &	proportions
	)

Splits dataset into multiple sets by proportions (uses global seed).

Shuffles indices via a thread_local mt19937 seeded from the global getRandomSeed() (or random_device on first call). For reproducibility, call setRandomSeed(42) before splitXY, or use the overload with explicit seed.

Parameters

X	Features [samples × features]
Y	Labels [samples]
proportions	Split fractions (sum ≤1, last is remainder)

Returns

Vector of (X_split, Y_split) pairs

splitXY() [2/2]

std::vector< std::pair< Eigen::MatrixXd, Eigen::VectorXd > > FlexNN::splitXY	(	const Eigen::MatrixXd &	X,
		const Eigen::VectorXd &	Y,
		const std::vector< double > &	proportions,
		uint32_t	seed
	)

Splits dataset with explicit seed (deterministic, no global state).

Uses a local mt19937(seed) so the global seed is unchanged. Prefer this in tests for hermetic determinism.

Parameters

X	Features [samples × features]
Y	Labels [samples]
proportions	Split fractions
seed	Explicit seed

Returns

Vector of splits