PC layer combining muPC scaling, optional residual connections, and DKP direct feedback – every extra OFF by default, so plain defaults reproduce DirectKPPCLayer exactly. See the file-level note above for the full design rationale. More...

#include <FullPCLayer.h>

Inheritance diagram for Deep::FullPCLayer:

Public Member Functions
	FullPCLayer (size_t size, size_t nextSize, size_t terminalSize, size_t batchSize, float learningRate, float inferenceRate, float feedback, float lmbda, ActivationType aType, ActivationType dType, IComputeBackend *backend=nullptr)
	~FullPCLayer () override=default
template<typename ArenaT>
void	BindMemory (ArenaT &arena)
	Binds this layer's buffers into a pre-allocated arena (MemoryArena for CPU, DeviceMemoryArena for GPU).
size_t	GetRequiredFloats () const noexcept
	Total floats this layer needs from its MemoryArena.
void	RandomizeWeights (std::mt19937 &seedGenerator) noexcept
	Randomizes W and Psi.
std::map< std::string, TensorDescriptor >	GetStateDict () const override
	Returns a dictionary of this.
void	SetLayerAbove (FullPCLayer *l) noexcept
	Sets the layer immediately above this one in the network.
void	SetLayerBelow (FullPCLayer *l) noexcept
	Sets the layer immediately below this one in the network.
void	SetTerminalLayer (FullPCLayer *l) noexcept
	Sets the network's terminal layer, for DirectFeedbackUpdate().
void	SetMuPCScale (float a) noexcept
	Set by the owning network at Compile() time, once the full layer_sizes list (and therefore d_in/N/L) is known. 1.0 (the default) reproduces plain, unscaled PC.
float	GetMuPCScale () const noexcept
	Returns the current muPC forward-scaling factor – see SetMuPCScale().
void	SetResidual (bool enabled) noexcept
	Set by the owning network at Compile() time. Requires size == nextSize (checked at the point it's actually used, not here, since nextSize may not be finalized yet when this is called). false (the default) reproduces plain PC.
bool	GetResidual () const noexcept
	Returns whether the residual connection is enabled – see SetResidual().
void	SetCrossEntropy (bool enabled) noexcept
	Enables softmax cross-entropy energy for THIS layer's error against layerBelow's mu (i.e. this only makes sense set on the terminal layer, against the second-to-last layer's logits – NOT looped over every layer the way SetMuPCScale/SetResidual/ SetMomentum's network-level setters are). OFF by default (plain Gaussian energy, matching DirectKPPCLayer exactly).
bool	GetCrossEntropy () const noexcept
	Returns whether softmax cross-entropy energy is enabled – see SetCrossEntropy().
float	CalculateState () noexcept override
	Matches Layer's virtual interface exactly (always computes real energy).
float	CalculateState (bool needEnergy) noexcept
	NOT a virtual override – see SimplePCLayer's identical pattern. Lets the settling loop skip the energy reduction (and its capture-time sync) when the caller doesn't need the value.
void	ComputeMuOnly () noexcept
	Computes only mu (forward prediction), skipping error/energy.
void	UpdateState () noexcept override
	Updates the states after calculation.
void	UpdateWeights () noexcept override
	Updates the weights after state updates.
void	DirectFeedbackUpdate () noexcept
	Perturbs W using the layer above's Psi and the terminal layer's error – the DFA phase, run once per batch before settling begins.
void	ClampState (const std::vector< float > &inputData) noexcept
	Clamps this layer's beliefs to inputData, fixing them against UpdateState() until UnclampState() is called.
void	UnclampState () noexcept
	Releases a previous ClampState() call.
void	ResetState () noexcept
	Resets this layer's beliefs (z) without touching learned weights.
bool	IsClamped () const noexcept
	Whether ClampState() is currently active on this layer.
void	InvalidateMuCache () noexcept
	Forces the next ComputeMuOnly() call to recompute mu from scratch, even if this layer is clamped. Needed under iPC: weights change every settling step, so a clamped layer's cached mu (valid under the standard, two-phase assumption that W is fixed throughout settling) goes stale the moment UpdateWeights() runs mid-loop.
void	SetMomentum (bool enabled, float beta=0.9f) noexcept
	Enables momentum (inertial) settling: the update direction is EMA-smoothed (decay beta) before being applied to z, instead of applied directly each step. OFF by default. v is allocated unconditionally in BindMemory() regardless of this flag's value at that time (same reasoning as biasGradScratch – cheap, and avoids an ordering hazard if this is called after Compile()). Reset to zero at the start of every TrainStep()/Predict() call, same as z.
bool	GetMomentum () const noexcept
	Returns whether momentum settling is enabled – see SetMomentum().
float	GetMomentumBeta () const noexcept
	Returns the momentum EMA decay rate – see SetMomentum().
void	SetOptimizer (OptimizerType o) noexcept
	Sets the optimizer for W. Call before Compile().
void	SetPsiOptimizer (OptimizerType o) noexcept
	Sets the optimizer for Psi. Call before Compile().
void	SetLearningRate (float learningRate) noexcept
	Sets the learning rate, propagating it to the optimizer state.
void	SetInferenceRate (float inferenceRate) noexcept
	Sets the inference rate.
void	SetFeedbackRate (float feedbackRate) noexcept
	Sets the feedback rate, propagating it to the optimizer state.
void	SetLambda (float lmbda) noexcept
	Sets the weight decay coefficient.
float *	GetBeliefs () noexcept override
	Should return belief state z.
const float *	GetErrors () const noexcept override
	Should return errors e.
const float *	GetMu () const noexcept
	Returns this layer's forward prediction buffer.
const float *	GetWeights () const noexcept
	Returns the weight matrix W, shape [nextSize, size].
const float *	GetDirectFeedbackWeights () const noexcept
	Returns the direct feedback (DFA) weight matrix Psi, shape [size, terminalSize].
const float *	GetBiases () const noexcept
	Returns the bias vector, length nextSize.
size_t	GetBatchSize () const noexcept override
	Should return batch size.
size_t	GetInputSize () const noexcept override
	Should return input size.
size_t	GetOutputSize () const noexcept override
	Should return output size.
size_t	GetTerminalSize () const noexcept
	Returns the network's final output layer size, as given at construction.
Public Member Functions inherited from Deep::Layer
virtual	~Layer ()=default
virtual void	Flush () noexcept
	Flushes remaining batches.

Protected Types
using	BackendDeleter = void ()(IComputeBackend)
	Custom deleter type for backend, needed since IComputeBackend is an abstract base – see the member below.

Protected Attributes
size_t	size
	This layer's own belief (z) size.
size_t	nextSize
	Size of the layer above (0 marks a terminal layer).
size_t	terminalSize
	Size of the network's final output layer.
size_t	batchSize
	Batch size.
float	lr
	Learning rate for W.
float	ir
	Inference rate (Euler integration step size).
float	fl
	Feedback rate for Psi.
float	lmbda
	Weight decay (L2 regularization) coefficient, shared between W and Psi.
float	a = 1.0f
	muPC forward-scaling factor – see SetMuPCScale().
bool	useResidual = false
	Whether the residual/skip connection is enabled – see SetResidual().
bool	isClamped = false
	Whether ClampState() is currently active.
bool	muCacheValid = false
	Whether cachedMu holds a valid, up-to-date value.
FullPCLayer *	layerAbove = nullptr
	The next layer up, or nullptr if this is terminal.
FullPCLayer *	layerBelow = nullptr
	The next layer down, or nullptr if this is the input layer.
FullPCLayer *	terminalLayer = nullptr
	The network's terminal layer, for DirectFeedbackUpdate().
ActivationType	activationType
	This layer's activation type.
OptimizerType	opt = OptimizerType::SGD
	Optimizer for W/b.
OptimizerType	optPsi = OptimizerType::SGD
	Optimizer for Psi.
std::unique_ptr< IComputeBackend, BackendDeleter >	backend
	This layer's compute backend.
float *	z = nullptr
	This layer's belief buffer, shape [batchSize, size].
float *	e = nullptr
	This layer's error buffer, shape [batchSize, size].
float *	W = nullptr
	Weight matrix, shape [nextSize, size].
float *	b = nullptr
	Bias vector, length nextSize.
float *	mu = nullptr
	Forward prediction buffer, shape [batchSize, nextSize].
float *	cachedMu = nullptr
	Cached mu, valid when muCacheValid is true.
float *	Psi = nullptr
	Direct feedback (DFA) weight matrix, shape [size, terminalSize].
float *	proj = nullptr
	Scratch buffer: terminal error projected back through layerAbove's Psi, for DirectFeedbackUpdate()'s W update.
float *	grad_W = nullptr
	W's gradient scratch buffer.
float *	grad_b = nullptr
	b's gradient scratch buffer.
float *	m_W = nullptr
	Adam/AdamW first-moment buffer for W.
float *	v_W = nullptr
	Adam/AdamW second-moment buffer for W.
float *	m_b = nullptr
	Adam/AdamW first-moment buffer for b.
float *	v_b = nullptr
	Adam/AdamW second-moment buffer for b.
float *	grad_Psi = nullptr
	Psi's gradient scratch buffer.
float *	m_Psi = nullptr
	Adam/AdamW first-moment buffer for Psi.
float *	v_Psi = nullptr
	Adam/AdamW second-moment buffer for Psi.
int *	t_device = nullptr
	Device-resident Adam/AdamW step count for W/b, for graph-captured steps.
float *	lr_device = nullptr
	Device-resident mirror of lr.
int *	tPsi_device = nullptr
	Device-resident Adam/AdamW step count for Psi.
float *	fl_device = nullptr
	Device-resident mirror of fl.
float *	zF = nullptr
	Activated belief, phi(z), used as the forward GEMM's input.
float *	zFDeriv = nullptr
	Activation derivative at z, for the feedback term.
float *	feedbackScratch = nullptr
	Scratch buffer for the feedback GEMM's output.
float *	v = nullptr
	Momentum buffer for the settling update – see SetMomentum().
bool	useMomentum = false
	Whether momentum settling is enabled – see SetMomentum().
float	momentumBeta = 0.9f
	Momentum EMA decay rate – see SetMomentum().
bool	useCrossEntropy = false
	Whether softmax cross-entropy energy is enabled – see SetCrossEntropy().
float *	rowEnergies = nullptr
	Per-row energy scratch for the cross-entropy path, length batchSize.
float *	biasGradScratch = nullptr
	nextSize-length scratch for SumRows' output in UpdateWeights()'s SGD branch (accumulate, which SumRows alone can't express). Same reasoning as DirectKPPCLayer's identical buffer.
std::unique_ptr< MemoryArena >	localArena
	This layer's own arena, used when it isn't bound into a network-wide one.
Protected Attributes inherited from Deep::Layer
size_t	size
	Size of input.
size_t	nextSize
	Size of output.

Detailed Description

PC layer combining muPC scaling, optional residual connections, and DKP direct feedback – every extra OFF by default, so plain defaults reproduce DirectKPPCLayer exactly. See the file-level note above for the full design rationale.

Member Typedef Documentation

◆ BackendDeleter

using Deep::FullPCLayer::BackendDeleter = void (*)(IComputeBackend*)

protected

Custom deleter type for backend, needed since IComputeBackend is an abstract base – see the member below.

Constructor & Destructor Documentation

◆ FullPCLayer()

Deep::FullPCLayer::FullPCLayer	(	size_t	size,
		size_t	nextSize,
		size_t	terminalSize,
		size_t	batchSize,
		float	learningRate,
		float	inferenceRate,
		float	feedback,
		float	lmbda,
		ActivationType	aType,
		ActivationType	dType,
		IComputeBackend *	backend = nullptr )

Parameters

size	input size
nextSize	output size (0 marks a terminal layer)
terminalSize	the size of the network's final output layer (e.g. 10 for MNIST) – required directly, not inferred, since the true terminal layer isn't known until the whole network has been assembled.
batchSize	batch size
learningRate	learning rate for W
inferenceRate	inference rate (Euler integration step size)
feedback	feedback rate for Psi
lmbda	weight decay (L2 regularization) coefficient, shared between W and Psi
aType	activation type
dType	activation derivative type
backend	Compute backend to run on; defaults to a new CPUBackend if nullptr.

◆ ~FullPCLayer()

Deep::FullPCLayer::~FullPCLayer ( )

overridedefault

Member Function Documentation

◆ BindMemory()

template<typename ArenaT>

template void Deep::FullPCLayer::BindMemory< DeviceMemoryArena > ( ArenaT & arena )

Binds this layer's buffers into a pre-allocated arena (MemoryArena for CPU, DeviceMemoryArena for GPU).

Template Parameters

ArenaT Either MemoryArena or DeviceMemoryArena.

◆ CalculateState() [1/2]

float Deep::FullPCLayer::CalculateState ( )

inlineoverridevirtualnoexcept

Matches Layer's virtual interface exactly (always computes real energy).

Implements Deep::Layer.

◆ CalculateState() [2/2]

float Deep::FullPCLayer::CalculateState ( bool needEnergy )

noexcept

NOT a virtual override – see SimplePCLayer's identical pattern. Lets the settling loop skip the energy reduction (and its capture-time sync) when the caller doesn't need the value.

◆ ClampState()

void Deep::FullPCLayer::ClampState ( const std::vector< float > & inputData )

noexcept

Clamps this layer's beliefs to inputData, fixing them against UpdateState() until UnclampState() is called.

◆ ComputeMuOnly()

void Deep::FullPCLayer::ComputeMuOnly ( )

noexcept

Computes only mu (forward prediction), skipping error/energy.

◆ DirectFeedbackUpdate()

void Deep::FullPCLayer::DirectFeedbackUpdate ( )

noexcept

Perturbs W using the layer above's Psi and the terminal layer's error – the DFA phase, run once per batch before settling begins.

◆ GetBatchSize()

size_t Deep::FullPCLayer::GetBatchSize ( ) const

inlineoverridevirtualnoexcept

Should return batch size.

Returns: ? = 0

Implements Deep::Layer.

◆ GetBeliefs()

float * Deep::FullPCLayer::GetBeliefs ( )

inlineoverridevirtualnoexcept

Should return belief state z.

Returns: ? = 0

Implements Deep::Layer.

◆ GetBiases()

const float * Deep::FullPCLayer::GetBiases ( ) const

inlinenoexcept

Returns the bias vector, length nextSize.

◆ GetCrossEntropy()

bool Deep::FullPCLayer::GetCrossEntropy ( ) const

inlinenoexcept

Returns whether softmax cross-entropy energy is enabled – see SetCrossEntropy().

◆ GetDirectFeedbackWeights()

const float * Deep::FullPCLayer::GetDirectFeedbackWeights ( ) const

inlinenoexcept

Returns the direct feedback (DFA) weight matrix Psi, shape [size, terminalSize].

◆ GetErrors()

const float * Deep::FullPCLayer::GetErrors ( ) const

inlineoverridevirtualnoexcept

Should return errors e.

Returns: ? = 0

Implements Deep::Layer.

◆ GetInputSize()

size_t Deep::FullPCLayer::GetInputSize ( ) const

inlineoverridevirtualnoexcept

Should return input size.

Returns: ? = 0

Implements Deep::Layer.

◆ GetMomentum()

bool Deep::FullPCLayer::GetMomentum ( ) const

inlinenoexcept

Returns whether momentum settling is enabled – see SetMomentum().

◆ GetMomentumBeta()

float Deep::FullPCLayer::GetMomentumBeta ( ) const

inlinenoexcept

Returns the momentum EMA decay rate – see SetMomentum().

◆ GetMu()

const float * Deep::FullPCLayer::GetMu ( ) const

inlinenoexcept

Returns this layer's forward prediction buffer.

◆ GetMuPCScale()

float Deep::FullPCLayer::GetMuPCScale ( ) const

inlinenoexcept

Returns the current muPC forward-scaling factor – see SetMuPCScale().

◆ GetOutputSize()

size_t Deep::FullPCLayer::GetOutputSize ( ) const

inlineoverridevirtualnoexcept

Should return output size.

Returns: ? = 0

Implements Deep::Layer.

◆ GetRequiredFloats()

size_t Deep::FullPCLayer::GetRequiredFloats ( ) const

noexcept

Total floats this layer needs from its MemoryArena.

◆ GetResidual()

bool Deep::FullPCLayer::GetResidual ( ) const

inlinenoexcept

Returns whether the residual connection is enabled – see SetResidual().

◆ GetStateDict()

std::map< std::string, Deep::TensorDescriptor > Deep::FullPCLayer::GetStateDict ( ) const

overridevirtual

Returns a dictionary of this.

Reimplemented from Deep::Layer.

◆ GetTerminalSize()

size_t Deep::FullPCLayer::GetTerminalSize ( ) const

inlinenoexcept

Returns the network's final output layer size, as given at construction.

◆ GetWeights()

const float * Deep::FullPCLayer::GetWeights ( ) const

inlinenoexcept

Returns the weight matrix W, shape [nextSize, size].

◆ InvalidateMuCache()

void Deep::FullPCLayer::InvalidateMuCache ( )

inlinenoexcept

Forces the next ComputeMuOnly() call to recompute mu from scratch, even if this layer is clamped. Needed under iPC: weights change every settling step, so a clamped layer's cached mu (valid under the standard, two-phase assumption that W is fixed throughout settling) goes stale the moment UpdateWeights() runs mid-loop.

◆ IsClamped()

bool Deep::FullPCLayer::IsClamped ( ) const

inlinenoexcept

Whether ClampState() is currently active on this layer.

◆ RandomizeWeights()

void Deep::FullPCLayer::RandomizeWeights ( std::mt19937 & seedGenerator )

noexcept

Randomizes W and Psi.

◆ ResetState()

void Deep::FullPCLayer::ResetState ( )

noexcept

Resets this layer's beliefs (z) without touching learned weights.

◆ SetCrossEntropy()

void Deep::FullPCLayer::SetCrossEntropy ( bool enabled )

inlinenoexcept

Enables softmax cross-entropy energy for THIS layer's error against layerBelow's mu (i.e. this only makes sense set on the terminal layer, against the second-to-last layer's logits – NOT looped over every layer the way SetMuPCScale/SetResidual/ SetMomentum's network-level setters are). OFF by default (plain Gaussian energy, matching DirectKPPCLayer exactly).

◆ SetFeedbackRate()

void Deep::FullPCLayer::SetFeedbackRate ( float feedbackRate )

noexcept

Sets the feedback rate, propagating it to the optimizer state.

◆ SetInferenceRate()

void Deep::FullPCLayer::SetInferenceRate ( float inferenceRate )

inlinenoexcept

Sets the inference rate.

◆ SetLambda()

void Deep::FullPCLayer::SetLambda ( float lmbda )

inlinenoexcept

Sets the weight decay coefficient.

◆ SetLayerAbove()

void Deep::FullPCLayer::SetLayerAbove ( FullPCLayer * l )

inlinenoexcept

Sets the layer immediately above this one in the network.

◆ SetLayerBelow()

void Deep::FullPCLayer::SetLayerBelow ( FullPCLayer * l )

inlinenoexcept

Sets the layer immediately below this one in the network.

◆ SetLearningRate()

void Deep::FullPCLayer::SetLearningRate ( float learningRate )

noexcept

Sets the learning rate, propagating it to the optimizer state.

◆ SetMomentum()

void Deep::FullPCLayer::SetMomentum	(	bool	enabled,
		float	beta = 0.9f )

inlinenoexcept

Enables momentum (inertial) settling: the update direction is EMA-smoothed (decay beta) before being applied to z, instead of applied directly each step. OFF by default. v is allocated unconditionally in BindMemory() regardless of this flag's value at that time (same reasoning as biasGradScratch – cheap, and avoids an ordering hazard if this is called after Compile()). Reset to zero at the start of every TrainStep()/Predict() call, same as z.

◆ SetMuPCScale()

void Deep::FullPCLayer::SetMuPCScale ( float a )

inlinenoexcept

Set by the owning network at Compile() time, once the full layer_sizes list (and therefore d_in/N/L) is known. 1.0 (the default) reproduces plain, unscaled PC.

◆ SetOptimizer()

void Deep::FullPCLayer::SetOptimizer ( OptimizerType o )

inlinenoexcept

Sets the optimizer for W. Call before Compile().

◆ SetPsiOptimizer()

void Deep::FullPCLayer::SetPsiOptimizer ( OptimizerType o )

inlinenoexcept

Sets the optimizer for Psi. Call before Compile().

◆ SetResidual()

void Deep::FullPCLayer::SetResidual ( bool enabled )

inlinenoexcept

Set by the owning network at Compile() time. Requires size == nextSize (checked at the point it's actually used, not here, since nextSize may not be finalized yet when this is called). false (the default) reproduces plain PC.

◆ SetTerminalLayer()

void Deep::FullPCLayer::SetTerminalLayer ( FullPCLayer * l )

inlinenoexcept

Sets the network's terminal layer, for DirectFeedbackUpdate().

◆ UnclampState()

void Deep::FullPCLayer::UnclampState ( )

noexcept

Releases a previous ClampState() call.

◆ UpdateState()

void Deep::FullPCLayer::UpdateState ( )

overridevirtualnoexcept

Updates the states after calculation.

Implements Deep::Layer.

◆ UpdateWeights()

void Deep::FullPCLayer::UpdateWeights ( )

overridevirtualnoexcept

Updates the weights after state updates.

Implements Deep::Layer.

Member Data Documentation

◆ a

float Deep::FullPCLayer::a = 1.0f

protected

muPC forward-scaling factor – see SetMuPCScale().

◆ activationType

ActivationType Deep::FullPCLayer::activationType

protected

This layer's activation type.

◆ b

float* Deep::FullPCLayer::b = nullptr

protected

Bias vector, length nextSize.

◆ backend

std::unique_ptr<IComputeBackend, BackendDeleter> Deep::FullPCLayer::backend

protected

This layer's compute backend.

◆ batchSize

size_t Deep::FullPCLayer::batchSize

protected

Batch size.

◆ biasGradScratch

float* Deep::FullPCLayer::biasGradScratch = nullptr

protected

nextSize-length scratch for SumRows' output in UpdateWeights()'s SGD branch (accumulate, which SumRows alone can't express). Same reasoning as DirectKPPCLayer's identical buffer.

◆ cachedMu

float* Deep::FullPCLayer::cachedMu = nullptr

protected

Cached mu, valid when muCacheValid is true.

◆ e

float* Deep::FullPCLayer::e = nullptr

protected

This layer's error buffer, shape [batchSize, size].

◆ feedbackScratch

float* Deep::FullPCLayer::feedbackScratch = nullptr

protected

Scratch buffer for the feedback GEMM's output.

◆ fl

float Deep::FullPCLayer::fl

protected

Feedback rate for Psi.

◆ fl_device

float* Deep::FullPCLayer::fl_device = nullptr

protected

Device-resident mirror of fl.

◆ grad_b

float* Deep::FullPCLayer::grad_b = nullptr

protected

b's gradient scratch buffer.

◆ grad_Psi

float* Deep::FullPCLayer::grad_Psi = nullptr

protected

Psi's gradient scratch buffer.

◆ grad_W

float* Deep::FullPCLayer::grad_W = nullptr

protected

W's gradient scratch buffer.

◆ ir

float Deep::FullPCLayer::ir

protected

Inference rate (Euler integration step size).

◆ isClamped

bool Deep::FullPCLayer::isClamped = false

protected

Whether ClampState() is currently active.

◆ layerAbove

FullPCLayer* Deep::FullPCLayer::layerAbove = nullptr

protected

The next layer up, or nullptr if this is terminal.

◆ layerBelow

FullPCLayer* Deep::FullPCLayer::layerBelow = nullptr

protected

The next layer down, or nullptr if this is the input layer.

◆ lmbda

float Deep::FullPCLayer::lmbda

protected

Weight decay (L2 regularization) coefficient, shared between W and Psi.

◆ localArena

std::unique_ptr<MemoryArena> Deep::FullPCLayer::localArena

protected

This layer's own arena, used when it isn't bound into a network-wide one.

◆ lr

float Deep::FullPCLayer::lr

protected

Learning rate for W.

◆ lr_device

float* Deep::FullPCLayer::lr_device = nullptr

protected

Device-resident mirror of lr.

◆ m_b

float* Deep::FullPCLayer::m_b = nullptr

protected

Adam/AdamW first-moment buffer for b.

◆ m_Psi

float* Deep::FullPCLayer::m_Psi = nullptr

protected

Adam/AdamW first-moment buffer for Psi.

◆ m_W

float* Deep::FullPCLayer::m_W = nullptr

protected

Adam/AdamW first-moment buffer for W.

◆ momentumBeta

float Deep::FullPCLayer::momentumBeta = 0.9f

protected

Momentum EMA decay rate – see SetMomentum().

◆ mu

float* Deep::FullPCLayer::mu = nullptr

protected

Forward prediction buffer, shape [batchSize, nextSize].

◆ muCacheValid

bool Deep::FullPCLayer::muCacheValid = false

protected

Whether cachedMu holds a valid, up-to-date value.

◆ nextSize

size_t Deep::FullPCLayer::nextSize

protected

Size of the layer above (0 marks a terminal layer).

◆ opt

OptimizerType Deep::FullPCLayer::opt = OptimizerType::SGD

protected

Optimizer for W/b.

◆ optPsi

OptimizerType Deep::FullPCLayer::optPsi = OptimizerType::SGD

protected

Optimizer for Psi.

◆ proj

float* Deep::FullPCLayer::proj = nullptr

protected

Scratch buffer: terminal error projected back through layerAbove's Psi, for DirectFeedbackUpdate()'s W update.

◆ Psi

float* Deep::FullPCLayer::Psi = nullptr

protected

Direct feedback (DFA) weight matrix, shape [size, terminalSize].

◆ rowEnergies

float* Deep::FullPCLayer::rowEnergies = nullptr

protected

Per-row energy scratch for the cross-entropy path, length batchSize.

◆ size

size_t Deep::FullPCLayer::size

protected

This layer's own belief (z) size.

◆ t_device

int* Deep::FullPCLayer::t_device = nullptr

protected

Device-resident Adam/AdamW step count for W/b, for graph-captured steps.

◆ terminalLayer

FullPCLayer* Deep::FullPCLayer::terminalLayer = nullptr

protected

The network's terminal layer, for DirectFeedbackUpdate().

◆ terminalSize

size_t Deep::FullPCLayer::terminalSize

protected

Size of the network's final output layer.

◆ tPsi_device

int* Deep::FullPCLayer::tPsi_device = nullptr

protected

Device-resident Adam/AdamW step count for Psi.

◆ useCrossEntropy

bool Deep::FullPCLayer::useCrossEntropy = false

protected

Whether softmax cross-entropy energy is enabled – see SetCrossEntropy().

◆ useMomentum

bool Deep::FullPCLayer::useMomentum = false

protected

Whether momentum settling is enabled – see SetMomentum().

◆ useResidual

bool Deep::FullPCLayer::useResidual = false

protected

Whether the residual/skip connection is enabled – see SetResidual().

◆ v

float* Deep::FullPCLayer::v = nullptr

protected

Momentum buffer for the settling update – see SetMomentum().

◆ v_b

float* Deep::FullPCLayer::v_b = nullptr

protected

Adam/AdamW second-moment buffer for b.

◆ v_Psi

float* Deep::FullPCLayer::v_Psi = nullptr

protected

Adam/AdamW second-moment buffer for Psi.

◆ v_W

float* Deep::FullPCLayer::v_W = nullptr

protected

Adam/AdamW second-moment buffer for W.

◆ W

float* Deep::FullPCLayer::W = nullptr

protected

Weight matrix, shape [nextSize, size].

◆ z

float* Deep::FullPCLayer::z = nullptr

protected

This layer's belief buffer, shape [batchSize, size].

◆ zF

float* Deep::FullPCLayer::zF = nullptr

protected

Activated belief, phi(z), used as the forward GEMM's input.

◆ zFDeriv

float* Deep::FullPCLayer::zFDeriv = nullptr

protected

Activation derivative at z, for the feedback term.

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

include/deepity/layers/FullPCLayer.h
src/layers/FullPCLayer.cpp

Public Member Functions

Protected Types

Protected Attributes

Detailed Description

Member Typedef Documentation

◆ BackendDeleter

Constructor & Destructor Documentation

◆ FullPCLayer()

◆ ~FullPCLayer()

Member Function Documentation

◆ BindMemory()

◆ CalculateState() [1/2]

◆ CalculateState() [2/2]

◆ ClampState()

◆ ComputeMuOnly()

◆ DirectFeedbackUpdate()

◆ GetBatchSize()

◆ GetBeliefs()

◆ GetBiases()

◆ GetCrossEntropy()

◆ GetDirectFeedbackWeights()

◆ GetErrors()

◆ GetInputSize()

◆ GetMomentum()

◆ GetMomentumBeta()

◆ GetMu()

◆ GetMuPCScale()

◆ GetOutputSize()

◆ GetRequiredFloats()

◆ GetResidual()

◆ GetStateDict()

◆ GetTerminalSize()

◆ GetWeights()

◆ InvalidateMuCache()

◆ IsClamped()

◆ RandomizeWeights()

◆ ResetState()

◆ SetCrossEntropy()

◆ SetFeedbackRate()

◆ SetInferenceRate()

◆ SetLambda()

◆ SetLayerAbove()

◆ SetLayerBelow()

◆ SetLearningRate()

◆ SetMomentum()

◆ SetMuPCScale()

◆ SetOptimizer()

◆ SetPsiOptimizer()

◆ SetResidual()

◆ SetTerminalLayer()

◆ UnclampState()

◆ UpdateState()

◆ UpdateWeights()

Member Data Documentation

◆ a

◆ activationType

◆ b

◆ backend

◆ batchSize

◆ biasGradScratch

◆ cachedMu

◆ e

◆ feedbackScratch

◆ fl

◆ fl_device

◆ grad_b

◆ grad_Psi

◆ grad_W

◆ ir

◆ isClamped

◆ layerAbove

◆ layerBelow

◆ lmbda

◆ localArena

◆ lr

◆ lr_device

◆ m_b

◆ m_Psi

◆ m_W

◆ momentumBeta