opennn::BackPropagation Struct Reference

Workspace holding parameter gradients and per-layer deltas during a backward pass. More...

#include <back_propagation.h>

Public Member Functions
	BackPropagation (const Index=0, Loss *=nullptr)
	Constructs a workspace for the given batch size and loss.
virtual	~BackPropagation ()=default
void	set (const Index=0, Loss *=nullptr)
	Reconfigures the workspace for a new batch size or loss; reuses allocations when possible.
void	accumulate_output_deltas (size_t layer_index)
	Accumulates deltas from all consumer edges into the given layer's output delta.
TensorView &	get_output_delta ()
	Returns the output delta of the network (gradient w.r.t. the final outputs).
const TensorView &	get_output_delta () const
	Returns the output delta of the network (gradient w.r.t. the final outputs).
void	print () const
	Prints a human-readable summary of the workspace contents.

Public Attributes
const NeuralNetwork *	neural_network = nullptr
Buffer	gradient
vector< vector< TensorView > >	gradient_views
Buffer	delta_pool
vector< vector< TensorView > >	delta_views
vector< vector< pair< size_t, size_t > > >	consumer_edges
Index	batch_size = 0
Loss *	loss = nullptr
float	error = 0.0f
float	accuracy = 0.0f
float	loss_value = 0.0f
Index	active_tokens_count = 0

Detailed Description

Workspace holding parameter gradients and per-layer deltas during a backward pass.

Constructor & Destructor Documentation

BackPropagation()

opennn::BackPropagation::BackPropagation	(	const Index	= 0,
		Loss *	= nullptr )

Constructs a workspace for the given batch size and loss.

Parameters

batch_size	Maximum number of samples per backward pass.
loss	Loss object whose network drives buffer sizing (non-owning).

~BackPropagation()

virtual opennn::BackPropagation::~BackPropagation ( )

virtualdefault

Member Function Documentation

accumulate_output_deltas()

void opennn::BackPropagation::accumulate_output_deltas

(

size_t

layer_index

)

Accumulates deltas from all consumer edges into the given layer's output delta.

Parameters

layer_index

Index of the layer whose output delta is being collected.

get_output_delta() [1/2]

TensorView & opennn::BackPropagation::get_output_delta

(

)

Returns the output delta of the network (gradient w.r.t. the final outputs).

get_output_delta() [2/2]

const TensorView & opennn::BackPropagation::get_output_delta

(

)

const

Returns the output delta of the network (gradient w.r.t. the final outputs).

print()

void opennn::BackPropagation::print

(

)

const

Prints a human-readable summary of the workspace contents.

set()

void opennn::BackPropagation::set	(	const Index	= 0,
		Loss *	= nullptr )

Reconfigures the workspace for a new batch size or loss; reuses allocations when possible.

Parameters

batch_size	Maximum number of samples per backward pass.
loss	Loss object whose network drives buffer sizing (non-owning).

Member Data Documentation

accuracy

float opennn::BackPropagation::accuracy = 0.0f

active_tokens_count

Index opennn::BackPropagation::active_tokens_count = 0

batch_size

Index opennn::BackPropagation::batch_size = 0

consumer_edges

vector<vector<pair<size_t, size_t> > > opennn::BackPropagation::consumer_edges

delta_pool

Buffer opennn::BackPropagation::delta_pool

delta_views

vector<vector<TensorView> > opennn::BackPropagation::delta_views

error

float opennn::BackPropagation::error = 0.0f

gradient

Buffer opennn::BackPropagation::gradient

gradient_views

vector<vector<TensorView> > opennn::BackPropagation::gradient_views

loss

Loss* opennn::BackPropagation::loss = nullptr

loss_value

float opennn::BackPropagation::loss_value = 0.0f

neural_network

const NeuralNetwork* opennn::BackPropagation::neural_network = nullptr