opennn::Batch Struct Reference

Owns the host-side and (optional) device-side buffers for one mini-batch of dataset samples. More...

#include <batch.h>

Public Member Functions
	Batch (const Index samples_number=0, const Dataset *dataset=nullptr)
	Constructs a batch sized for a given dataset.
void	set (const Index samples_number=0, const Dataset *dataset=nullptr)
	(Re)allocates buffers for a given batch size and dataset.
void	fill (const vector< Index > &sample_indices, const vector< Index > &input_feature_indices, const vector< Index > &decoder_feature_indices, const vector< Index > &target_feature_indices, bool augment=false)
	Fills the host-side buffers from the bound dataset.
const vector< TensorView > &	get_inputs () const
	Returns the input TensorViews for the active device.
const TensorView &	get_targets () const
	Returns the target TensorView for the active device.
Index	get_samples_number () const
	Number of samples currently held in the batch.
void	print () const
	Prints a human-readable summary of the batch buffers to stdout.
bool	is_empty () const
	Whether the batch holds zero samples.
void	copy_device_async (const Index sample_count, cudaStream_t stream)
	Asynchronously copies the host buffers to the device on stream.

Public Attributes
Index	samples_number = 0
	Number of samples currently held in the batch.
const Dataset *	dataset = nullptr
	Dataset whose shapes determined the batch buffer sizes; not owned.
Buffer	input
	Owning storage for the input tensor (host or device).
Shape	input_shape
	Shape of input (per-sample dimensions).
Buffer	decoder
	Owning storage for the decoder tensor (encoder-decoder models).
Shape	decoder_shape
	Shape of decoder.
Buffer	target
	Owning storage for the target tensor.
Shape	target_shape
	Shape of target.
int	input_contiguous = -1
	Stride hint passed to Dataset::fill_inputs(); -1 to ignore.
int	decoder_contiguous = -1
	Stride hint passed to Dataset::fill_inputs() for the decoder; -1 to ignore.
int	target_contiguous = -1
	Stride hint passed to Dataset::fill_targets(); -1 to ignore.
vector< TensorView >	input_views_host_cache
	Host-side input TensorViews (one per input feature group).
TensorView	target_view_host_cache
	Host-side target TensorView.
vector< TensorView >	input_views_cache
	Device-side input TensorViews; populated only on CUDA mode.
TensorView	target_view_cache
	Device-side target TensorView; populated only on CUDA mode.
Index	num_input_features = 0
	Total number of input features across all input feature groups.
Index	num_decoder_features = 0
	Total number of decoder features.
Index	num_target_features = 0
	Total number of target features.
float *	inputs_host = nullptr
	Pinned host pointer used to stage inputs into the device buffer.
float *	decoder_host = nullptr
	Pinned host pointer used to stage decoder inputs into the device buffer.
float *	targets_host = nullptr
	Pinned host pointer used to stage targets into the device buffer.
Index	inputs_host_allocated_size = 0
	Allocation capacity of inputs_host, in floats.
Index	decoder_host_allocated_size = 0
	Allocation capacity of decoder_host, in floats.
Index	targets_host_allocated_size = 0
	Allocation capacity of targets_host, in floats.
Buffer	inputs_fp32_staging {Device::CUDA}
	Device-resident FP32 staging buffer used during mixed-precision uploads.

Detailed Description

Owns the host-side and (optional) device-side buffers for one mini-batch of dataset samples.

The Batch is filled once per training step from a Dataset using fill(). Its TensorViews (target_view, input_views) are then handed to the NeuralNetwork's forward / backward passes. On CUDA-enabled builds the batch maintains parallel host and device buffers; copy_device_async() uploads the host data on a user-supplied CUDA stream so it can overlap with compute.

Constructor & Destructor Documentation

Batch()

opennn::Batch::Batch	(	const Index	samples_number = 0,
		const Dataset *	dataset = nullptr )

Constructs a batch sized for a given dataset.

Parameters

samples_number	Number of samples per batch.
dataset	Dataset whose shapes determine the batch buffer sizes.

Member Function Documentation

copy_device_async()

void opennn::Batch::copy_device_async	(	const Index	sample_count,
		cudaStream_t	stream )

Asynchronously copies the host buffers to the device on stream.

Parameters

sample_count	Number of samples to copy from the host buffer.
stream	CUDA stream used for the asynchronous copy.

fill()

void opennn::Batch::fill	(	const vector< Index > &	sample_indices,
		const vector< Index > &	input_feature_indices,
		const vector< Index > &	decoder_feature_indices,
		const vector< Index > &	target_feature_indices,
		bool	augment = false )

Fills the host-side buffers from the bound dataset.

Parameters

sample_indices	Indices of samples copied into this batch.
input_feature_indices	Dataset columns used as inputs.
decoder_feature_indices	Dataset columns used as decoder inputs (empty when the model has no decoder).
target_feature_indices	Dataset columns used as targets.
augment	True to apply Dataset::augment_inputs() to the input buffer.

get_inputs()

const vector< TensorView > & opennn::Batch::get_inputs ( ) const

inline

Returns the input TensorViews for the active device.

Returns

Device-side views on CUDA builds, host-side views otherwise.

get_samples_number()

Index opennn::Batch::get_samples_number

(

)

const

Number of samples currently held in the batch.

get_targets()

const TensorView & opennn::Batch::get_targets ( ) const

inline

Returns the target TensorView for the active device.

Returns

Device-side view on CUDA builds, host-side view otherwise.

is_empty()

bool opennn::Batch::is_empty

(

)

const

Whether the batch holds zero samples.

print()

void opennn::Batch::print

(

)

const

Prints a human-readable summary of the batch buffers to stdout.

set()

void opennn::Batch::set	(	const Index	samples_number = 0,
		const Dataset *	dataset = nullptr )

(Re)allocates buffers for a given batch size and dataset.

Parameters

samples_number	Number of samples per batch.
dataset	Dataset whose shapes determine the batch buffer sizes.

Member Data Documentation

dataset

const Dataset* opennn::Batch::dataset = nullptr

Dataset whose shapes determined the batch buffer sizes; not owned.

decoder

Buffer opennn::Batch::decoder

Owning storage for the decoder tensor (encoder-decoder models).

decoder_contiguous

int opennn::Batch::decoder_contiguous = -1

Stride hint passed to Dataset::fill_inputs() for the decoder; -1 to ignore.

decoder_host

float* opennn::Batch::decoder_host = nullptr

Pinned host pointer used to stage decoder inputs into the device buffer.

decoder_host_allocated_size

Index opennn::Batch::decoder_host_allocated_size = 0

Allocation capacity of decoder_host, in floats.

decoder_shape

Shape opennn::Batch::decoder_shape

Shape of decoder.

input

Buffer opennn::Batch::input

Owning storage for the input tensor (host or device).

input_contiguous

int opennn::Batch::input_contiguous = -1

Stride hint passed to Dataset::fill_inputs(); -1 to ignore.

input_shape

Shape opennn::Batch::input_shape

Shape of input (per-sample dimensions).

input_views_cache

vector<TensorView> opennn::Batch::input_views_cache

Device-side input TensorViews; populated only on CUDA mode.

input_views_host_cache

vector<TensorView> opennn::Batch::input_views_host_cache

Host-side input TensorViews (one per input feature group).

inputs_fp32_staging

Buffer opennn::Batch::inputs_fp32_staging {Device::CUDA}

Device-resident FP32 staging buffer used during mixed-precision uploads.

inputs_host

float* opennn::Batch::inputs_host = nullptr

Pinned host pointer used to stage inputs into the device buffer.

inputs_host_allocated_size

Index opennn::Batch::inputs_host_allocated_size = 0

Allocation capacity of inputs_host, in floats.

num_decoder_features

Index opennn::Batch::num_decoder_features = 0

Total number of decoder features.

num_input_features

Index opennn::Batch::num_input_features = 0

Total number of input features across all input feature groups.

num_target_features

Index opennn::Batch::num_target_features = 0

Total number of target features.

samples_number

Index opennn::Batch::samples_number = 0

Number of samples currently held in the batch.

target

Buffer opennn::Batch::target

Owning storage for the target tensor.

target_contiguous

int opennn::Batch::target_contiguous = -1

Stride hint passed to Dataset::fill_targets(); -1 to ignore.

target_shape

Shape opennn::Batch::target_shape

Shape of target.

target_view_cache

TensorView opennn::Batch::target_view_cache

Device-side target TensorView; populated only on CUDA mode.

target_view_host_cache

TensorView opennn::Batch::target_view_host_cache

Host-side target TensorView.

targets_host

float* opennn::Batch::targets_host = nullptr

Pinned host pointer used to stage targets into the device buffer.

targets_host_allocated_size

Index opennn::Batch::targets_host_allocated_size = 0

Allocation capacity of targets_host, in floats.