FlexNN::Layers::Dense Class Reference

Fully-connected layer: Z = W*X + b, A = act(Z). More...

#include <Dense.hpp>

Public Member Functions
	Dense (int in, int out, Activations::Activation act=Activations::Activation::ReLU, const Activations::ActivationParameters &params=Activations::ActivationParameters{})
	Construct Dense with input/output sizes and activation.
	Dense (DenseParams p, Activations::Activation act=Activations::Activation::ReLU, const Activations::ActivationParameters &params=Activations::ActivationParameters{})
	Construct from DenseParams POD.
	Dense (int in, int out, const std::string &actStr)
	Deprecated string ctor — delegates to enum via try_parse.
LayerType	type () const noexcept
Activations::Activation	activation () const noexcept
const Activations::ActivationParameters &	activationParams () const noexcept
const DenseParams &	params () const noexcept
const Eigen::MatrixXd &	weights () const noexcept
const Eigen::VectorXd &	biases () const noexcept
void	setWeights (const Eigen::MatrixXd &W)
void	setBiases (const Eigen::VectorXd &b)
void	setActivationParameters (const Activations::ActivationParameters &p)
std::pair< Eigen::MatrixXd, Eigen::MatrixXd >	forward (const Eigen::MatrixXd &input) const
	Forward pass: Z = W*X + b, A = act(Z).
Eigen::MatrixXd	backward (const Eigen::MatrixXd &nextW, const Eigen::MatrixXd &nextdZ, const Eigen::MatrixXd &currZ) const
	Backward: ‘dZ = f’(Z) ⊙ (W_next^T * dZ_next)ordZ = W_next^T*dZ_next` if None.
Eigen::MatrixXd	backward (const Eigen::MatrixXd &upstream, const Eigen::MatrixXd &currZ) const
	Backward with already-propagated upstream (W_next^T*dZ_next).
Eigen::MatrixXd	propagate (const Eigen::MatrixXd &dZ) const
	Propagate gradient to previous layer: dX = W^T * dZ.
void	update (const Eigen::MatrixXd &dW, const Eigen::VectorXd &db, double lr) noexcept
	SGD update: W -= lr*dW, b -= lr*db.

Detailed Description

Fully-connected layer: Z = W*X + b, A = act(Z).

Layout: X is [in × batch] column-major, W is [out × in], b is [out] broadcast column-wise. Activation is applied elementwise via FlexNN::Activations::detail.

Constructor & Destructor Documentation

Dense() [1/3]

FlexNN::Layers::Dense::Dense	(	int	in,
		int	out,
		Activations::Activation	act = Activations::Activation::ReLU,
		const Activations::ActivationParameters &	params = Activations::ActivationParameters{}
	)

Construct Dense with input/output sizes and activation.

Initializes W and b with Random * 0.5 (same as legacy include/Layer.h:55). Kept for v0.1 compatibility; future versions may use Xavier/He initialization.

Parameters

in	Number of input features
out	Number of output neurons
act	Activation (default ReLU)

Dense() [2/3]

FlexNN::Layers::Dense::Dense ( DenseParams  p, Activations::Activation  act = Activations::Activation::ReLU, const Activations::ActivationParameters &  params = Activations::ActivationParameters{}  )

explicit

Construct from DenseParams POD.

Dense() [3/3]

FlexNN::Layers::Dense::Dense	(	int	in,
		int	out,
		const std::string &	actStr
	)

Deprecated string ctor — delegates to enum via try_parse.

Kept for one release so Dense(784,64,"relu") still compiles with a deprecation warning. On parse failure, activation is None.

Member Function Documentation

activation()

Activations::Activation FlexNN::Layers::Dense::activation ( ) const

inlinenoexcept

activationParams()

const Activations::ActivationParameters & FlexNN::Layers::Dense::activationParams ( ) const

inlinenoexcept

backward() [1/2]

Eigen::MatrixXd FlexNN::Layers::Dense::backward	(	const Eigen::MatrixXd &	nextW,
		const Eigen::MatrixXd &	nextdZ,
		const Eigen::MatrixXd &	currZ
	)		const

Backward: ‘dZ = f’(Z) ⊙ (W_next^T * dZ_next)ordZ = W_next^T*dZ_next` if None.

For Softmax hidden, this uses detail::softmax_backward_hidden via Activations::detail::backward. Last-layer Softmax is fused in NeuralNetwork::backward and never calls this for the last layer. activationParams threaded for LeakyReLU alpha.

Parameters

nextW	Weights of next layer (or empty for last layer)
nextdZ	Gradient of next layer's Z
currZ	This layer's pre-activation Z from forward

Returns

Gradient w.r.t. this layer's Z

backward() [2/2]

Eigen::MatrixXd FlexNN::Layers::Dense::backward	(	const Eigen::MatrixXd &	upstream,
		const Eigen::MatrixXd &	currZ
	)		const

Backward with already-propagated upstream (W_next^T*dZ_next).

Used by NeuralNetwork::backward after layers[i+1].propagate(dZ_next) so the caller handles Conv1D transpose via col2im. Equivalent to the three-arg version but skips the nextW multiply.

Parameters

upstream	Already W_next^T * dZ_next (or propagate result)
currZ	This layer's Z

Returns

Gradient w.r.t. Z

biases()

const Eigen::VectorXd & FlexNN::Layers::Dense::biases ( ) const

inlinenoexcept

forward()

std::pair< Eigen::MatrixXd, Eigen::MatrixXd > FlexNN::Layers::Dense::forward

(

const Eigen::MatrixXd &

input

)

const

Forward pass: Z = W*X + b, A = act(Z).

Parameters

input

Batch input [in × batch]

Returns

Pair (Z, A) each [out × batch]

params()

const DenseParams & FlexNN::Layers::Dense::params ( ) const

inlinenoexcept

propagate()

Eigen::MatrixXd FlexNN::Layers::Dense::propagate

(

const Eigen::MatrixXd &

dZ

)

const

Propagate gradient to previous layer: dX = W^T * dZ.

Parameters

dZ	Gradient w.r.t. Z of this layer [out × batch]

Returns

Gradient w.r.t. input [in × batch]

setActivationParameters()

void FlexNN::Layers::Dense::setActivationParameters ( const Activations::ActivationParameters &  p )

inline

setBiases()

void FlexNN::Layers::Dense::setBiases ( const Eigen::VectorXd &  b )

inline

setWeights()

void FlexNN::Layers::Dense::setWeights ( const Eigen::MatrixXd &  W )

inline

type()

LayerType FlexNN::Layers::Dense::type ( ) const

inlinenoexcept

update()

void FlexNN::Layers::Dense::update ( const Eigen::MatrixXd &  dW, const Eigen::VectorXd &  db, double  lr  )

noexcept

SGD update: W -= lr*dW, b -= lr*db.

weights()

const Eigen::MatrixXd & FlexNN::Layers::Dense::weights ( ) const

inlinenoexcept

---

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

• include/layers/Dense.hpp
• lib/layers/Dense.cpp