An abstracted class for an array of SimplePCLayer. More...

#include <SimplePCNetwork.h>

Public Member Functions
	SimplePCNetwork (int batchSize, DeviceType device=DeviceType::DEVICE_CPU) noexcept
	Constructs an empty network with a predetermined batch size.
	SimplePCNetwork (const SimplePCNetwork &)=delete
SimplePCNetwork &	operator= (const SimplePCNetwork &)=delete
	~SimplePCNetwork ()=default
void	AddLayer (int size, int nextSize, float lr, float ir, float lmbda, void(act)(float , size_t), void(dAct)(float , size_t, bool))
	Adds a layer to the network.
void	AddLayer (int size, int nextSize, float lr, float ir, float lmbda, ActivationType aType, ActivationType dType)
	Same as the function-pointer overload above, taking an ActivationType pair instead – the enum-based dispatch used by callers (e.g. the Python bindings) that name an activation by type rather than linking against its concrete function.
void	RandomizeWeights (std::mt19937 &rng, const char *distribution)
	Randomizes the weights of each layer.
void	RandomizeWeights (std::mt19937 &rng)
	Randomizes the weights of each layer.
void	ResetState () noexcept
	Resets each layer's state without touching learned weights.
void	Clamp (const std::vector< float > &input)
	Clamps the input to the first layer, necessary for prediction.
float	CalculateState (bool needEnergy=true)
	Calculates the state of each layer.
void	UpdateState ()
	Updates each layer's state.
void	UpdateWeights ()
	Updates each layer's weights.
SimplePCLayer *	GetTerminalLayer () noexcept
	Returns the network's terminal (final) layer.
std::vector< std::unique_ptr< SimplePCLayer > > &	GetLayers () noexcept
	Returns every layer in the network, in the order they were added.
const std::vector< std::unique_ptr< SimplePCLayer > > &	GetLayers () const noexcept
	Returns every layer in the network, in the order they were added.
int	GetBatchSize () const noexcept
	Returns the batch size for the network's layers.
DeviceType	GetDevice () const noexcept
	Returns which device this network's layers run on.
void	SetOptimizer (OptimizerType o) noexcept
	Sets the optimizer used for weight updates on every layer.
float	TrainStep (const std::vector< float > &x, const std::vector< float > &y, int inferenceSteps)
	Runs a complete training step (clamp, settle, update, unclamp).
std::vector< float >	Predict (const std::vector< float > &x, int inferenceSteps)
	Runs a forward prediction pass (clamp, settle, read).
void	ProjectForward () noexcept
	Runs a single, non-iterative forward pass through current weights, seeding each hidden layer's z from the PREVIOUS layer's mu – giving the settling loop a genuine, current-weights-based starting point instead of zero-init.
float	TrainStepWithProjection (const std::vector< float > &x, const std::vector< float > &y, int inferenceSteps, bool computeEnergy=true)
	Full train step WITH forward-projection initialization, all in ONE call – reset, clamp, project, settle, update weights, unclamp. Matches TrainStep()'s signature/return convention exactly, just with ProjectForward() inserted between clamping the input and clamping the target.
std::vector< float >	PredictWithProjection (const std::vector< float > &x, int inferenceSteps)
	Inference-only counterpart to TrainStepWithProjection(): reset, clamp input, project forward, settle, read out the terminal layer's beliefs. No target clamp, no weight update.
void	SetMuCacheThreshold (float threshold) noexcept
	Sets mu-cache threshold on every layer – see SimplePCLayer::SetMuCacheThreshold() for semantics. Safe to call any time after Compile().
void	Compile ()
	Loads all layers into one contiguous block of memory (MemoryArena for DEVICE_CPU, DeviceMemoryArena for DEVICE_GPU).

Detailed Description

An abstracted class for an array of SimplePCLayer.

Constructor & Destructor Documentation

◆ SimplePCNetwork() [1/2]

Deep::SimplePCNetwork::SimplePCNetwork	(	int	batchSize,
		DeviceType	device = DeviceType::DEVICE_CPU )

explicitnoexcept

Constructs an empty network with a predetermined batch size.

Parameters

batchSize	Batch size
device	Which device this network's layers should run on. Defaults to DEVICE_CPU, preserving existing behavior exactly for anyone not explicitly requesting DEVICE_GPU.

◆ SimplePCNetwork() [2/2]

Deep::SimplePCNetwork::SimplePCNetwork ( const SimplePCNetwork & )

delete

◆ ~SimplePCNetwork()

Deep::SimplePCNetwork::~SimplePCNetwork ( )

default

Member Function Documentation

◆ AddLayer() [1/2]

void Deep::SimplePCNetwork::AddLayer	(	int	size,
		int	nextSize,
		float	lr,
		float	ir,
		float	lmbda,
		ActivationType	aType,
		ActivationType	dType )

Same as the function-pointer overload above, taking an ActivationType pair instead – the enum-based dispatch used by callers (e.g. the Python bindings) that name an activation by type rather than linking against its concrete function.

Parameters

size	input size
nextSize	output size
lr	learning rate for beliefs
ir	learning rate for weights
lmbda	weight decay (L2 regularization) coefficient
aType	activation type
dType	activation derivative type

◆ AddLayer() [2/2]

void Deep::SimplePCNetwork::AddLayer	(	int	size,
		int	nextSize,
		float	lr,
		float	ir,
		float	lmbda,
		void(*	act )(float *, size_t),
		void(*	dAct )(float *, size_t, bool) )

Adds a layer to the network.

Parameters

size	input size
nextSize	output size
lr	learning rate for beliefs
ir	learning rate for weights
lmbda	weight decay (L2 regularization) coefficient
act	activation function
dAct	derivative of previous activation function

◆ CalculateState()

float Deep::SimplePCNetwork::CalculateState ( bool needEnergy = true )

Calculates the state of each layer.

Parameters

needEnergy ask for energy after

Returns: Returns total energy if asked for

◆ Clamp()

void Deep::SimplePCNetwork::Clamp ( const std::vector< float > & input )

Clamps the input to the first layer, necessary for prediction.

Parameters

input reference to input vector

◆ Compile()

void Deep::SimplePCNetwork::Compile ( )

Loads all layers into one contiguous block of memory (MemoryArena for DEVICE_CPU, DeviceMemoryArena for DEVICE_GPU).

◆ GetBatchSize()

int Deep::SimplePCNetwork::GetBatchSize ( ) const

inlinenoexcept

Returns the batch size for the network's layers.

Returns: size_t batchSize

◆ GetDevice()

DeviceType Deep::SimplePCNetwork::GetDevice ( ) const

inlinenoexcept

Returns which device this network's layers run on.

◆ GetLayers() [1/2]

const std::vector< std::unique_ptr< SimplePCLayer > > & Deep::SimplePCNetwork::GetLayers ( ) const

inlinenoexcept

Returns every layer in the network, in the order they were added.

Returns: A const reference to the internal layer list.

◆ GetLayers() [2/2]

std::vector< std::unique_ptr< SimplePCLayer > > & Deep::SimplePCNetwork::GetLayers ( )

inlinenoexcept

Returns every layer in the network, in the order they were added.

Returns: A reference to the internal layer list.

◆ GetTerminalLayer()

SimplePCLayer * Deep::SimplePCNetwork::GetTerminalLayer ( )

inlinenoexcept

Returns the network's terminal (final) layer.

Returns: Pointer to the last layer added via AddLayer().

◆ operator=()

SimplePCNetwork & Deep::SimplePCNetwork::operator= ( const SimplePCNetwork & )

delete

◆ Predict()

std::vector< float > Deep::SimplePCNetwork::Predict	(	const std::vector< float > &	x,
		int	inferenceSteps )

Runs a forward prediction pass (clamp, settle, read).

Parameters

x	The batched input data
inferenceSteps	The number of relaxation iterations

Returns: A vector containing the batched predictions

◆ PredictWithProjection()

std::vector< float > Deep::SimplePCNetwork::PredictWithProjection	(	const std::vector< float > &	x,
		int	inferenceSteps )

Inference-only counterpart to TrainStepWithProjection(): reset, clamp input, project forward, settle, read out the terminal layer's beliefs. No target clamp, no weight update.

Parameters

x	Flattened input batch.
inferenceSteps	Number of settling steps.

Returns: The terminal layer's beliefs after settling.

◆ ProjectForward()

void Deep::SimplePCNetwork::ProjectForward ( )

noexcept

Runs a single, non-iterative forward pass through current weights, seeding each hidden layer's z from the PREVIOUS layer's mu – giving the settling loop a genuine, current-weights-based starting point instead of zero-init.

Reuses each layer's EXISTING CalculateState() (already-verified forward computation, mu = f(W@z+b)) as a side effect – no new math, just a new sequence of existing calls. Assumes the input layer (layers[0]) is ALREADY clamped before this is called.

Warning: Layer 0's error (e) and this call's energy return value are meaningless here – CalculateState() computes both mu (what we want) and e/energy (a side effect we're discarding, since the NEXT layer's z hasn't been set to a meaningful value yet at the point each layer's CalculateState() runs). Only mu is used.

◆ RandomizeWeights() [1/2]

void Deep::SimplePCNetwork::RandomizeWeights ( std::mt19937 & rng )

Randomizes the weights of each layer.

Parameters

rng	The classic Mersenne Twister

◆ RandomizeWeights() [2/2]

void Deep::SimplePCNetwork::RandomizeWeights	(	std::mt19937 &	rng,
		const char *	distribution )

Randomizes the weights of each layer.

Parameters

rng	The classic Mersenne Twister
distribution	A string representation of (normal, uniform) distribution. For example: "normal(0, 1)" for a standard normal.

◆ ResetState()

void Deep::SimplePCNetwork::ResetState ( )

noexcept

Resets each layer's state without touching learned weights.

◆ SetMuCacheThreshold()

void Deep::SimplePCNetwork::SetMuCacheThreshold ( float threshold )

noexcept

Sets mu-cache threshold on every layer – see SimplePCLayer::SetMuCacheThreshold() for semantics. Safe to call any time after Compile().

◆ SetOptimizer()

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

inlinenoexcept

Sets the optimizer used for weight updates on every layer.

Parameters

o	The optimizer type to apply.

◆ TrainStep()

float Deep::SimplePCNetwork::TrainStep	(	const std::vector< float > &	x,
		const std::vector< float > &	y,
		int	inferenceSteps )

Runs a complete training step (clamp, settle, update, unclamp).

Parameters

x	The batched input data
y	The batched target data
inferenceSteps	The number of relaxation iterations

Returns: The final energy state of the network before weight updates

◆ TrainStepWithProjection()

float Deep::SimplePCNetwork::TrainStepWithProjection	(	const std::vector< float > &	x,
		const std::vector< float > &	y,
		int	inferenceSteps,
		bool	computeEnergy = true )

Full train step WITH forward-projection initialization, all in ONE call – reset, clamp, project, settle, update weights, unclamp. Matches TrainStep()'s signature/return convention exactly, just with ProjectForward() inserted between clamping the input and clamping the target.

Exists specifically to eliminate the Python/pybind boundary- crossing overhead of doing this same sequence via many separate calls from Python (reset_state, clamp_input, project_forward, clamp_state, then STEPS*2 individual calculate_state/update_state calls, update_weights, unclamp_state – over 40 individual crossings per batch at STEPS=20). This does the whole sequence in ONE crossing instead.

◆ UpdateState()

void Deep::SimplePCNetwork::UpdateState ( )

Updates each layer's state.

◆ UpdateWeights()

void Deep::SimplePCNetwork::UpdateWeights ( )

Updates each layer's weights.

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

include/deepity/networks/SimplePCNetwork.h
src/networks/SimplePCNetwork.cpp

Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ SimplePCNetwork() [1/2]

◆ SimplePCNetwork() [2/2]

◆ ~SimplePCNetwork()

Member Function Documentation

◆ AddLayer() [1/2]

◆ AddLayer() [2/2]

◆ CalculateState()

◆ Clamp()

◆ Compile()

◆ GetBatchSize()

◆ GetDevice()

◆ GetLayers() [1/2]

◆ GetLayers() [2/2]

◆ GetTerminalLayer()

◆ operator=()

◆ Predict()

◆ PredictWithProjection()

◆ ProjectForward()

◆ RandomizeWeights() [1/2]

◆ RandomizeWeights() [2/2]

◆ ResetState()

◆ SetMuCacheThreshold()

◆ SetOptimizer()

◆ TrainStep()

◆ TrainStepWithProjection()

◆ UpdateState()

◆ UpdateWeights()