FlexNN::Layers::Layer Class Reference

Type-erased layer — holds one concrete layer type. More...

#include <Layer.hpp>

Public Member Functions
	Layer (Dense d)
	Layer (Conv1D c)
	Layer (BatchNorm1D b)
	Layer (MaxPool1D m)
	Layer (AvgPool1D a)
LayerType	type () const noexcept
	Layer kind (Dense, Conv1D, ...).
Activations::Activation	activation () const noexcept
	Activation for this layer.
Activations::ActivationParameters	activationParams () const noexcept
	Activation hyperparameters (e.g., LeakyReLU alpha).
std::pair< Eigen::MatrixXd, Eigen::MatrixXd >	forward (const Eigen::MatrixXd &input) const
	Forward pass for this layer (dispatch via visit).
Eigen::MatrixXd	backward (const Eigen::MatrixXd &nextW, const Eigen::MatrixXd &nextdZ, const Eigen::MatrixXd &currZ) const
	Backward pass for this layer (dispatch via visit).
Eigen::MatrixXd	backward (const Eigen::MatrixXd &upstream, const Eigen::MatrixXd &currZ) const
	Backward with already-propagated upstream (W_next^T*dZ_next).
void	updateWeights (const Eigen::MatrixXd &dW, const Eigen::VectorXd &db, double lr) noexcept
	SGD update — dispatches to concrete update.
void	update (const Eigen::MatrixXd &dW, const Eigen::VectorXd &db, double lr) noexcept
Eigen::MatrixXd	weightsMatrix () const
	Weight matrix for backward — Dense/Conv1D.
Eigen::MatrixXd	propagate (const Eigen::MatrixXd &dZ) const
	Propagate gradient to previous layer (W^T * dZ or col2im).
std::pair< Eigen::MatrixXd, Eigen::VectorXd >	grad (const Eigen::MatrixXd &dZ, const Eigen::MatrixXd &input) const
	Compute dW/db for this layer given dZ and input.
const auto &	variant () const noexcept
	Direct access to the underlying variant (for ModelIO).
auto &	variant () noexcept
const Dense *	asDense () const noexcept
Dense *	asDense () noexcept
const Conv1D *	asConv1D () const noexcept
Conv1D *	asConv1D () noexcept
const BatchNorm1D *	asBatchNorm1D () const noexcept
BatchNorm1D *	asBatchNorm1D () noexcept
const MaxPool1D *	asMaxPool1D () const noexcept
MaxPool1D *	asMaxPool1D () noexcept
const AvgPool1D *	asAvgPool1D () const noexcept
AvgPool1D *	asAvgPool1D () noexcept

Detailed Description

Type-erased layer — holds one concrete layer type.

In PR-05 the variant holds Dense and Conv1D. Later PRs expand it:

• PR-07: std::variant<Dense, Conv1D, BatchNorm1D>
• PR-08: std::variant<Dense, Conv1D, BatchNorm1D, MaxPool1D, AvgPool1D>

Each concrete type implements the same interface: LayerType type() const noexcept Activations::Activation activation() const noexcept pair<MatrixXd,MatrixXd> forward(const MatrixXd&) const MatrixXd backward(const MatrixXd& nextW, const MatrixXd& nextdZ, const MatrixXd& currZ) const void update(const MatrixXd& dW, const VectorXd& db, double lr) noexcept

The wrapper forwards to the active alternative via std::visit.

Constructor & Destructor Documentation

Layer() [1/5]

FlexNN::Layers::Layer::Layer ( Dense  d )

inline

Layer() [2/5]

FlexNN::Layers::Layer::Layer ( Conv1D  c )

inline

Layer() [3/5]

FlexNN::Layers::Layer::Layer ( BatchNorm1D  b )

inline

Layer() [4/5]

FlexNN::Layers::Layer::Layer ( MaxPool1D  m )

inline

Layer() [5/5]

FlexNN::Layers::Layer::Layer ( AvgPool1D  a )

inline

Member Function Documentation

activation()

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

inlinenoexcept

Activation for this layer.

activationParams()

Activations::ActivationParameters FlexNN::Layers::Layer::activationParams ( ) const

inlinenoexcept

Activation hyperparameters (e.g., LeakyReLU alpha).

Returns stored ActivationParameters for layers that carry it (Dense, Conv1D); otherwise default. Non-Leaky activations ignore the returned struct.

asAvgPool1D() [1/2]

const AvgPool1D * FlexNN::Layers::Layer::asAvgPool1D ( ) const

inlinenoexcept

asAvgPool1D() [2/2]

AvgPool1D * FlexNN::Layers::Layer::asAvgPool1D ( )

inlinenoexcept

asBatchNorm1D() [1/2]

const BatchNorm1D * FlexNN::Layers::Layer::asBatchNorm1D ( ) const

inlinenoexcept

asBatchNorm1D() [2/2]

BatchNorm1D * FlexNN::Layers::Layer::asBatchNorm1D ( )

inlinenoexcept

asConv1D() [1/2]

const Conv1D * FlexNN::Layers::Layer::asConv1D ( ) const

inlinenoexcept

asConv1D() [2/2]

Conv1D * FlexNN::Layers::Layer::asConv1D ( )

inlinenoexcept

asDense() [1/2]

const Dense * FlexNN::Layers::Layer::asDense ( ) const

inlinenoexcept

asDense() [2/2]

Dense * FlexNN::Layers::Layer::asDense ( )

inlinenoexcept

asMaxPool1D() [1/2]

const MaxPool1D * FlexNN::Layers::Layer::asMaxPool1D ( ) const

inlinenoexcept

asMaxPool1D() [2/2]

MaxPool1D * FlexNN::Layers::Layer::asMaxPool1D ( )

inlinenoexcept

backward() [1/2]

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

inline

Backward pass for this layer (dispatch via visit).

For the last layer, NeuralNetwork::backward fuses Softmax+CE and never calls this for the last layer. For hidden layers this applies ‘f’(Z) ⊙ (W_next^T * dZ_next)`.

Kept for dense_test compatibility (takes nextW/nextdZ).

backward() [2/2]

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

inline

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.

forward()

std::pair< Eigen::MatrixXd, Eigen::MatrixXd > FlexNN::Layers::Layer::forward ( const Eigen::MatrixXd &  input ) const

inline

Forward pass for this layer (dispatch via visit).

Parameters

input

Batch input [in_dim × batch]

Returns

Pair (Z, A) each [out_dim × batch]

grad()

std::pair< Eigen::MatrixXd, Eigen::VectorXd > FlexNN::Layers::Layer::grad ( const Eigen::MatrixXd &  dZ, const Eigen::MatrixXd &  input  ) const

inline

Compute dW/db for this layer given dZ and input.

For Dense: dW = dZ * X^T / batch, db = rowMean(dZ). For Conv1D: dW/db via Conv1D::grad. Returns pair (dW, db) as two matrices (db as column vector promoted to MatrixXd).

propagate()

Eigen::MatrixXd FlexNN::Layers::Layer::propagate ( const Eigen::MatrixXd &  dZ ) const

inline

Propagate gradient to previous layer (W^T * dZ or col2im).

Used by NeuralNetwork::backward to compute upstream for hidden layers: upstream = layers[i+1].propagate(dZ_next). For Dense this is W^T * dZ; for Conv1D this is col2im(W^T * dZ) via Conv1D::propagate.

Note

Not noexcept: allocates MatrixXd (host, bad_alloc terminates).

type()

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

inlinenoexcept

Layer kind (Dense, Conv1D, ...).

update()

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

inlinenoexcept

updateWeights()

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

inlinenoexcept

SGD update — dispatches to concrete update.

Kept as updateWeights for compatibility with legacy Layer::updateWeights signature used by NeuralNetwork::updateWeights. New code may call update.

variant() [1/2]

const auto & FlexNN::Layers::Layer::variant ( ) const

inlinenoexcept

Direct access to the underlying variant (for ModelIO).

Lets ModelIO do std::visit without public variant exposure of concrete types. Kept const — mutation goes via updateWeights.

variant() [2/2]

auto & FlexNN::Layers::Layer::variant ( )

inlinenoexcept

weightsMatrix()

Eigen::MatrixXd FlexNN::Layers::Layer::weightsMatrix ( ) const

inline

Weight matrix for backward — Dense/Conv1D.

For Dense this is W [out×in]; for Conv1D this is W [outCh×inCh*K]. Kept for ModelIO inspection; for backward prefer propagate() which handles Conv1D col2im correctly. Not noexcept (allocates on copy).

---

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

• include/layers/Layer.hpp