opennn Namespace Reference

Classes
struct	ActivationOp
	Element-wise non-linear activation (Identity, Sigmoid, Tanh, ReLU, Softmax). More...
class	AdaptiveMomentEstimation
	Adam optimizer with first/second gradient moments for stochastic minibatch training. More...
class	Addition
	Element-wise addition layer that sums two input tensors of identical shape (residual connections). More...
struct	AddOp
	Element-wise sum of several input tensors (used by residual connections). More...
class	ApproximationNetwork
	Factory neural network preconfigured for regression / function approximation. More...
struct	AttentionOp
	Scaled dot-product attention with optional causal mask and dropout. More...
struct	AugmentationSettings
	Image augmentation parameters: reflections, rotations and translations applied at training. More...
class	AutoAssociationNetwork
	Factory neural network preconfigured for auto-association (anomaly detection). More...
class	Backend
	Process-wide singleton that owns the thread pool and the cuBLAS/cuDNN handles. More...
struct	BackPropagation
	Workspace holding parameter gradients and per-layer deltas during a backward pass. More...
struct	BackPropagationLM
	Backprop scratch state specific to Levenberg-Marquardt (per-sample errors, Jacobian, Hessian approx). More...
struct	Batch
	Minibatch container holding pinned host/device buffers and views into a Dataset. More...
struct	BatchNormOp
	Batch normalization with learnable scale/shift and running statistics for inference. More...
class	Bounding
	Output bounding layer that clips outputs to configured lower and upper limits. More...
struct	BoundOp
	Clamps each output channel to a configurable lower/upper interval. More...
struct	BoxPlot
	Five-number summary (minimum, Q1, median, Q3, maximum) used to draw a box plot. More...
struct	Buffer
	Owning raw byte buffer that lives on CPU or CUDA memory, with aligned (re)allocation. More...
class	ClassificationNetwork
	Factory neural network preconfigured for tabular classification. More...
struct	CombinationOp
	Affine combination output = input * weights + bias (the dense matmul building block). More...
struct	CombinationReluOp
	Fused affine + ReLU activation (uses cuBLASLt epilogue on GPU when available). More...
class	Configuration
	Global singleton holding the OpenNN device and precision configuration. More...
class	Convolutional
	2D convolutional layer with kernel, stride, padding, activation, and optional batch normalization. More...
class	ConvolutionalRelu
	Fused convolution + ReLU layer; runs as a single GPU op (cudnn) and is CUDA-Graph friendly. More...
struct	ConvolutionOp
	2D convolution operator (NHWC layout) backed by Eigen on CPU and cuDNN on GPU. More...
struct	ConvolutionReluOp
	Fused 2D convolution + ReLU activation (uses cuDNN fused epilogue on GPU). More...
struct	Correlation
	Result of a correlation analysis: model parameters, fit quality, and the method/form used. More...
class	CsvReader
	Tokenising CSV reader that returns string_views into a single backing buffer. More...
class	Dataset
	Abstract base class for OpenNN datasets, owning samples, variables, and metadata. More...
class	Dense
	Fully-connected layer with configurable activation, optional batch normalization and dropout. More...
class	DenseRelu
	Fused dense + ReLU layer; combines linear projection and ReLU activation in a single op for performance. More...
struct	Descriptives
	Summary statistics (minimum, maximum, mean, standard deviation) for one variable. More...
struct	DropoutOp
	Inverted dropout: at training time zeros activations with probability rate and rescales survivors. More...
class	Embedding
	Token-id to dense vector embedding layer with optional scaling and positional encoding. More...
struct	EmbeddingLookupOp
	Token embedding lookup with optional scaling and additive positional encoding. More...
struct	EnumMap
class	FileReader
	Thread-safe positional file reader (pread on POSIX, overlapped ReadFile on Windows). More...
class	FileWriter
	Streaming writer that finalises by atomic-renaming a .tmp file to its final path. More...
struct	FlatOp
	Flattens a multi-dimensional tensor into a 2D (batch, features) tensor. More...
class	Flatten
	Flatten layer that reshapes a multi-dimensional input into a single 1D feature vector. More...
class	ForecastingNetwork
	Factory neural network preconfigured for time-series forecasting. More...
struct	ForwardPropagation
	Workspace holding the activations of every layer during a forward pass. More...
class	GeneticAlgorithm
	Selects the optimal subset of input features using an evolutionary genetic algorithm. More...
class	GrowingInputs
	Selects the optimal subset of input features by greedily growing the input set. More...
class	GrowingNeurons
	Selects the optimal hidden neuron count by incrementally growing the number of neurons. More...
struct	Histogram
	Frequency histogram with per-bin minimums, maximums, centers, and counts. More...
class	ImageClassificationNetwork
	Factory convolutional neural network preconfigured for image classification. More...
class	ImageDataset
	Image dataset that streams BMP images from disk with optional augmentation. More...
class	InputsSelection
	Abstract base class for algorithms that search the optimal subset of input variables. More...
struct	InputsSelectionResults
	Aggregated results of an inputs selection run including optimal inputs and error histories. More...
class	Json
class	JsonDocument
class	JsonWriter
class	KMeans
	K-means clustering utility that partitions samples into the requested number of clusters. More...
class	LanguageDataset
	Token-based language dataset with input/target vocabularies and binary token cache. More...
class	Layer
	Abstract base class for all OpenNN layers; orchestrates operators and shape propagation. More...
struct	LayerNormOp
	Layer normalization with learnable scale/shift, applied across the embedding dimension. More...
class	LevenbergMarquardtAlgorithm
	Levenberg-Marquardt optimizer combining Gauss-Newton and gradient descent through an adaptive damping factor. More...
class	Loss
	Unified loss container supporting MSE, cross-entropy, Minkowski, weighted, and regularized variants. More...
struct	MergeOp
	Reshapes (batch, heads, seq, head_dim) tensors back into (batch, seq, embed); no parameters. More...
class	ModelExpression
	Emits a trained neural network as source code in C, Python, JavaScript, or PHP. More...
class	ModelSelection
	Orchestrates model selection by combining inputs selection and neurons selection over a TrainingStrategy. More...
class	MultiHeadAttention
	Multi-head scaled dot-product attention layer used in transformer architectures. More...
struct	MultiHeadProjectionOp
	Projects (input_features) into (heads * head_dim) and reshapes for multi-head attention. More...
class	NeuralNetwork
	Container of layers forming a feed-forward neural network, with parameter storage and I/O. More...
class	NeuronSelection
	Abstract base class for algorithms that select the optimal number of hidden neurons. More...
struct	NeuronsSelectionResults
	Aggregated results of a neurons selection run including the optimal neuron count and error histories. More...
class	Normalization3d
	Layer normalization over the embedding axis of a 3D (batch, sequence, embedding) tensor. More...
struct	Operator
	Base class for compute building blocks composed by layers (matmul, activation, dropout, etc.). More...
class	Optimizer
	Abstract base class for training optimizers (Adam, SGD, Quasi-Newton, Levenberg-Marquardt). More...
struct	OptimizerData
	Per-optimizer scratch state (moments, directions, iteration counter) backing the update step. More...
struct	Pool3dOp
	Sequence-wide 1D pooling over the embedding dimension (mean or max). More...
class	Pooling
	2D spatial pooling layer supporting max and average reduction. More...
class	Pooling3d
	Sequence pooling layer reducing the time axis of a (sequence, features) input. More...
struct	PoolOp
	2D pooling operator supporting max and average reductions. More...
class	QuasiNewtonMethod
	Quasi-Newton optimizer using the BFGS inverse-Hessian approximation with line search. More...
class	Recurrent
	Basic recurrent (RNN) layer that unrolls over time steps with a shared activation. More...
class	Registry
class	ResponseOptimization
	Optimizes input values so that a network's outputs satisfy user-defined conditions and objectives. More...
struct	ScaleOp
	Scales inputs to a target range using per-feature minimum/maximum or mean/std statistics. More...
class	Scaling
	Input scaling layer that normalizes features using per-variable descriptive statistics. More...
class	ScopedTimer
struct	Shape
	Fixed-capacity small-vector describing tensor dimensions (rank up to MaxRank). More...
class	SimpleResNet
	Factory residual neural network with a configurable number of blocks per stage. More...
struct	Stats
class	StochasticGradientDescent
	Stochastic gradient descent with optional momentum, Nesterov, and learning-rate decay. More...
class	TabularDataset
	Tabular dataset with CSV loading, scaling, descriptive statistics and correlation analysis. More...
struct	TensorSpec
	Lightweight description of a tensor's shape and data type (no storage attached). More...
struct	TensorView
	Non-owning view over a tensor: pointer, shape, and data type with rich reshape helpers. More...
class	TestingAnalysis
	Performs post-training analysis of a neural network: errors, confusion matrices, ROC, gain charts, etc. More...
class	TextClassificationNetwork
	Factory neural network preconfigured for text classification. More...
class	ThreadSafeQueue
class	TimeSeriesDataset
	Time series dataset with configurable past/future windows and autocorrelation analysis. More...
struct	TrainingResults
	History and final metrics produced by a training run. More...
class	TrainingStrategy
	High-level orchestrator pairing a Loss with an Optimizer for a network/dataset. More...
class	Transformer
	Factory encoder-decoder Transformer neural network for sequence-to-sequence tasks. More...
class	TransformerDecoder
	Drives token-by-token inference of a Transformer model with configurable sampling strategies. More...
struct	TypeInfo
	Compile-time traits mapping an opennn::Type to its underlying numeric type and library identifiers. More...
struct	TypeInfo< Type::BF16 >
	TypeInfo specialization for bfloat16 (BF16) tensors. More...
struct	TypeInfo< Type::FP32 >
	TypeInfo specialization for 32-bit floating point (FP32) tensors. More...
struct	TypeInfo< Type::INT8 >
	TypeInfo specialization for signed 8-bit integer (INT8) tensors. More...
struct	UnscaleOp
	Inverse of ScaleOp: maps normalized outputs back to the original feature range. More...
class	Unscaling
	Output unscaling layer that reverts normalization back to the original feature ranges. More...
struct	Variable
	Single dataset column descriptor: name, role, type, scaler, and optional categories. More...
class	VGG16
	Factory neural network reproducing the VGG-16 architecture. More...

Typedefs
using	bfloat16 = __nv_bfloat16
template<typename T, size_t N>
using	array = Eigen::array<T, N>

Enumerations
enum class	Device { Auto , CPU , CUDA }
	Execution device selection for OpenNN runtime (auto-detected, CPU or CUDA GPU). More...
enum class	Type { Auto , FP32 , BF16 , INT8 }
	Numeric precision used for training or inference tensors. More...
enum class	SampleRole { Training , Validation , Testing , None }
	Role of a sample in a dataset split. More...
enum class	LayerType {   Addition , Bounding , Convolutional , ConvolutionalRelu ,   Dense , DenseRelu , Embedding , Flatten ,   MultiHeadAttention , Normalization3d , Pooling , Pooling3d ,   Recurrent , Scaling , Unscaling }
	Identifier of every concrete layer subclass shipped with OpenNN. More...
enum class	PoolingMethod { MaxPooling , AveragePooling }
	Pooling reduction method used by Pooling and Pooling3d layers. More...
enum	DateFormat { AUTO , DMY , MDY , YMD }
	Order of the day, month, and year fields in a date string (AUTO probes the input). More...
enum class	VariableType {   None , Numeric , Binary , Categorical ,   DateTime , Constant }
	Data type of a dataset Variable. More...
enum class	ScalerMethod {   None , MinimumMaximum , MeanStandardDeviation , StandardDeviation ,   Logarithm , ImageMinMax }
	Feature scaling strategy applied to a Variable before training. More...
enum class	VariableRole {   None , Input , Target , Decoder ,   InputTarget , Time }
	Role a Variable plays in a dataset (input feature, target, decoder, time axis, etc.). More...

Functions
template<Type... Supported, typename F>
void	visit_type (Type t, F &&f)
	Dispatches f with the TypeInfo of the runtime Type t (must be in Supported).
template<Type... Supported, typename F>
void	visit_type_pair (Type t_in, Type t_out, F &&f)
	Dispatches f with the TypeInfo pair for an input and output runtime Type.
cudnnDataType_t	to_cudnn (Type type) noexcept
	Returns the cuDNN data type matching the given OpenNN Type (Auto resolves to FP32).
cudaDataType_t	to_cuda (Type type) noexcept
	Returns the CUDA data type matching the given OpenNN Type (Auto resolves to FP32).
Index	type_bytes (Type type) noexcept
	Returns the byte size of one element of the given OpenNN Type.
bool	is_gpu ()
	Returns true when the resolved configuration runs on a CUDA GPU.
bool	is_cpu ()
	Returns true when the resolved configuration runs on CPU.
bool	is_bf16_training ()
	Returns true when training is configured to use BF16 precision.
bool	is_bf16_inference ()
	Returns true when inference is configured to use BF16 precision.
Device	current_device ()
	Returns the active runtime device (CUDA if available, otherwise CPU).
Correlation	linear_correlation (const VectorR &, const VectorR &)
	Pearson linear correlation between two equal-length vectors.
Correlation	logarithmic_correlation (const VectorR &, const VectorR &)
	Logarithmic correlation: fits y = a + b * log(x) and returns the resulting fit.
Correlation	exponential_correlation (const VectorR &, const VectorR &)
	Exponential correlation: fits y = a * exp(b * x) and returns the resulting fit.
Correlation	power_correlation (const VectorR &, const VectorR &)
	Power correlation: fits y = a * x^b and returns the resulting fit.
Correlation	logistic_correlation (const VectorR &, const VectorR &)
	Logistic correlation between two vectors (binary or continuous targets).
Correlation	logistic_correlation (const VectorR &, const MatrixR &)
	Logistic correlation between a vector predictor and a one-hot matrix target.
Correlation	logistic_correlation (const MatrixR &, const VectorR &)
	Logistic correlation between a one-hot matrix predictor and a vector target.
Correlation	logistic_correlation (const MatrixR &, const MatrixR &)
	Logistic correlation between two one-hot matrices.
Correlation	point_biserial_correlation (const VectorR &, const VectorR &)
	Point-biserial correlation between a binary vector and a continuous vector.
Correlation	eta_squared_correlation (const VectorR &, const MatrixR &)
	Eta-squared (effect-size) correlation between a continuous vector and a categorical matrix.
Correlation	correlation (const MatrixR &, const MatrixR &)
	Generic correlation between two matrices, dispatching on column types (binary, categorical, continuous).
Correlation	linear_correlation_spearman (const VectorR &, const VectorR &)
	Spearman rank correlation between two vectors using a linear fit on ranks.
VectorR	calculate_spearman_ranks (const VectorR &)
	Computes Spearman ranks (average rank for ties) for the entries of a vector.
Correlation	logistic_correlation_spearman (const VectorR &, const VectorR &)
	Spearman-rank logistic correlation between two vectors.
Correlation	correlation_spearman (const MatrixR &, const MatrixR &)
	Generic Spearman correlation between two matrices, dispatching on column types.
float	r_correlation_to_z_correlation (const float)
	Fisher r-to-z transform of a correlation coefficient.
float	z_correlation_to_r_correlation (const float)
	Inverse Fisher z-to-r transform.
pair< float, float >	confidence_interval_z_correlation (const float, Index)
	Returns the [lower, upper] confidence interval for a correlation given its sample size.
VectorR	autocorrelations (const VectorR &, Index=10)
	Autocorrelations of a series for lags 0..max_lag.
VectorR	cross_correlations (const VectorR &, const VectorR &, Index)
	Cross-correlations between two series for lags 0..max_lag.
MatrixR	get_correlation_values (const Tensor< Correlation, 2 > &)
	Extracts the coefficient r from a 2D tensor of Correlation values.
const EnumMap< SampleRole > &	sample_role_map ()
	Returns the bidirectional string/enum map for SampleRole.
const string &	sample_role_to_string (SampleRole role)
	Returns the canonical string name for a SampleRole.
SampleRole	string_to_sample_role (const string &name)
	Parses a string (name or "0"/"1"/"2"/"3") into the matching SampleRole.
void	mean_squared_error (const TensorView &input, const TensorView &target, float &error, float *workspace_device)
	Computes the mean squared error between predictions and targets.
void	mean_squared_error_gradient (const TensorView &input, const TensorView &target, const TensorView &input_delta)
	Writes the MSE gradient with respect to the predictions into input_delta.
void	normalized_squared_error (const TensorView &input, const TensorView &target, float coefficient, float &error, float *workspace_device)
	Computes the squared error normalized by a dataset-level coefficient.
void	normalized_squared_error_gradient (const TensorView &input, const TensorView &target, float coefficient, const TensorView &input_delta)
	Writes the normalized-squared-error gradient with respect to the predictions into input_delta.
void	weighted_squared_error (const TensorView &input, const TensorView &target, float pos_w, float neg_w, float &error, float *workspace_device)
	Computes the binary squared error weighted asymmetrically for positive and negative classes.
void	weighted_squared_error_gradient (const TensorView &input, const TensorView &target, float pos_w, float neg_w, float coefficient, const TensorView &input_delta)
	Writes the gradient of the weighted squared error scaled by coefficient into input_delta.
void	binary_cross_entropy (const TensorView &input, const TensorView &target, float &error, float *workspace_device)
	Computes the binary cross-entropy between predicted probabilities and binary targets.
void	categorical_cross_entropy (const TensorView &input, const TensorView &target, float &error, float *workspace_device)
	Computes the multi-class (categorical) cross-entropy between softmax probabilities and one-hot targets.
void	cross_entropy_gradient (const TensorView &input, const TensorView &target, const TensorView &input_delta)
	Writes the cross-entropy gradient with respect to the (pre-softmax/logit) predictions into input_delta.
void	minkowski_error (const TensorView &input, const TensorView &target, float power, float &error, float *workspace_device)
	Computes the Minkowski error sum(|input - target|^power) for the given power exponent.
void	minkowski_error_gradient (const TensorView &input, const TensorView &target, float power, const TensorView &input_delta)
	Writes the Minkowski-error gradient with respect to the predictions into input_delta.
void	cross_entropy_3d (const TensorView &input, const TensorView &target, float &error, Index &active_tokens_out, Index &correct_tokens_out, float *errors_device=nullptr)
	Computes 3-D (sequence) cross-entropy used by transformer-style targets, ignoring padded positions.
void	cross_entropy_3d_gradient (const TensorView &input, const TensorView &target, const TensorView &input_delta, Index active_tokens_count)
	Writes the 3-D cross-entropy gradient into input_delta, normalizing by the host-side active-token count.
void	cross_entropy_3d_gradient_device_count (const TensorView &input, const TensorView &target, const TensorView &input_delta, const float *active_tokens_count_device)
	Variant of cross_entropy_3d_gradient that reads the active-token count from device memory.
void	l1_regularization (const TensorView &parameters, float lambda, float &penalty)
	Computes the L1 regularization penalty lambda * sum(|parameters|).
void	l1_regularization_gradient (const TensorView &parameters, float lambda, const TensorView &gradient)
	Adds the L1 regularization gradient lambda * sign(parameters) into the gradient tensor.
void	l2_regularization (const TensorView &parameters, float lambda, float &penalty)
	Computes the L2 regularization penalty lambda * sum(parameters^2).
void	l2_regularization_gradient (const TensorView &parameters, float lambda, const TensorView &gradient)
	Adds the L2 regularization gradient 2 * lambda * parameters into the gradient tensor.
Tensor3	load_image (const filesystem::path &)
	Loads an image from disk into a rank-3 (height, width, channels) tensor.
void	load_image (const filesystem::path &, float *dst, Index expected_height, Index expected_width, Index expected_channels, bool divide_by_255=false)
	Loads an image into a pre-allocated float buffer at the given shape.
Tensor3	resize_image (const Tensor3 &, Index, Index)
	Returns a resized copy of an image at the requested height and width.
void	reflect_image_horizontal (Tensor3 &)
	Mirrors the image horizontally (left-right) in place.
void	reflect_image_vertical (Tensor3 &)
	Mirrors the image vertically (top-bottom) in place.
void	rotate_image (const Tensor3 &, Tensor3 &, float)
	Rotates the image by the given angle (radians) into the destination tensor.
void	translate_image_x (const Tensor3 &, Tensor3 &, Index)
	Translates the image along the X axis by the given number of pixels.
void	translate_image_y (const Tensor3 &, Tensor3 &, Index)
	Translates the image along the Y axis by the given number of pixels.
void	atomic_rename (const filesystem::path &from, const filesystem::path &to)
	Atomically renames a file, replacing the destination if needed.
void	add_json_field (JsonWriter &writer, const string &name, const string &value)
void	write_json (JsonWriter &writer, initializer_list< pair< const char *, string > > props)
float	read_json_type (const Json *root, const string &field)
long	read_json_index (const Json *root, const string &field)
bool	read_json_bool (const Json *root, const string &field)
string	read_json_string (const Json *root, const string &field)
string	read_json_string_fallback (const Json *root, initializer_list< string > names)
const Json *	require_json_field (const Json *root, const string &field)
template<typename Func>
void	for_json_items (const Json *parent, const char *tag, long count, Func func)
JsonDocument	load_json_file (const filesystem::path &file_name)
const Json *	get_json_root (const JsonDocument &document, const string &tag)
const EnumMap< LayerType > &	layer_type_map ()
	Returns the bidirectional mapping between LayerType values and their string names.
const string &	layer_type_to_string (LayerType type)
	Returns the string name associated with the given LayerType.
LayerType	string_to_layer_type (const string &name)
	Returns the LayerType corresponding to the given string name.
void	check_rank (const Shape &shape, initializer_list< int > allowed, const char *layer, const char *what)
	Throws if shape rank is not one of allowed.
void	pad (const TensorView &input, TensorView &output)
	Pads the input tensor and writes the result into output.
void	bound (const TensorView &input, const TensorView &lower_bounds, const TensorView &upper_bounds, TensorView &output)
	Clamps each element of input to the [lower_bounds, upper_bounds] range.
void	bound_cpu (const TensorView &input, const TensorView &lower_bounds, const TensorView &upper_bounds, TensorView &output)
	CPU implementation of bound().
void	scale (const TensorView &input, const TensorView &minimums, const TensorView &maximums, const TensorView &means, const TensorView &standard_deviations, const TensorView &scalers, float min_range, float max_range, TensorView &output)
	Applies per-feature scaling (mean/std, min/max, or other scalers) to a tensor.
void	scale_cpu (const TensorView &input, const TensorView &minimums, const TensorView &maximums, const TensorView &means, const TensorView &standard_deviations, const TensorView &scalers, float min_range, float max_range, TensorView &output)
	CPU implementation of scale().
void	unscale (const TensorView &input, const TensorView &minimums, const TensorView &maximums, const TensorView &means, const TensorView &standard_deviations, const TensorView &scalers, float min_range, float max_range, TensorView &output)
	Inverse of scale(); reconstructs original values from a previously scaled tensor.
void	unscale_cpu (const TensorView &input, const TensorView &minimums, const TensorView &maximums, const TensorView &means, const TensorView &standard_deviations, const TensorView &scalers, float min_range, float max_range, TensorView &output)
	CPU implementation of unscale().
void	copy (const TensorView &source, TensorView &destination)
	Copies the contents of source into destination, dispatching to CPU or GPU as needed.
void	copy_cpu (const TensorView &source, TensorView &destination)
	CPU implementation of copy().
void	add (const TensorView &input_1, const TensorView &input_2, TensorView &output)
	Element-wise addition: output = input_1 + input_2.
void	add_cpu (const TensorView &input_1, const TensorView &input_2, TensorView &output)
	CPU implementation of add().
void	multiply (const TensorView &input_a, bool transpose_a, const TensorView &input_b, bool transpose_b, TensorView &output, float alpha=1.0f, float beta=0.0f)
	General matrix multiply: output = alpha * op(input_a) * op(input_b) + beta * output.
void	multiply_cpu (const TensorView &input_a, bool transpose_a, const TensorView &input_b, bool transpose_b, TensorView &output, float alpha=1.0f, float beta=0.0f)
	CPU implementation of multiply().
void	softmax (TensorView &output)
	Applies softmax in place along the trailing dimension of output.
void	softmax_cpu (TensorView &output)
	CPU implementation of softmax().
void	max_pooling_3d_forward (const TensorView &input, TensorView &output, TensorView &maximal_indices, bool is_training)
	Forward pass of 3D max pooling; records argmax positions when training.
void	max_pooling_3d_forward_cpu (const TensorView &input, TensorView &output, TensorView &maximal_indices, bool is_training)
	CPU implementation of max_pooling_3d_forward().
void	average_pooling_3d_forward (const TensorView &input, TensorView &output)
	Forward pass of 3D average pooling.
void	average_pooling_3d_forward_cpu (const TensorView &input, TensorView &output)
	CPU implementation of average_pooling_3d_forward().
void	max_pooling_3d_backward (const TensorView &maximal_indices, const TensorView &output_delta, TensorView &input_delta)
	Backward pass for 3D max pooling; routes gradients to argmax positions.
void	max_pooling_3d_backward_cpu (const TensorView &maximal_indices, const TensorView &output_delta, TensorView &input_delta)
	CPU implementation of max_pooling_3d_backward().
void	average_pooling_3d_backward (const TensorView &input, const TensorView &output_delta, TensorView &input_delta)
	Backward pass for 3D average pooling.
void	average_pooling_3d_backward_cpu (const TensorView &input, const TensorView &output_delta, TensorView &input_delta)
	CPU implementation of average_pooling_3d_backward().
void	split_heads (const TensorView &source, TensorView &destination)
	Reshapes a multi-head attention tensor by splitting the last axis into heads.
void	split_heads_cpu (const TensorView &source, TensorView &destination)
	CPU implementation of split_heads().
void	merge_heads (const TensorView &source, TensorView &destination)
	Inverse of split_heads(); merges per-head tensors back into a single representation.
void	merge_heads_cpu (const TensorView &source, TensorView &destination)
	CPU implementation of merge_heads().
void	sort_string_vector (vector< string > &)
	Sorts a vector of strings in place in ascending lexicographical order.
vector< string >	concatenate_string_vectors (const vector< string > &, const vector< string > &)
	Concatenates two string vectors and returns the resulting vector.
string	formatNumber (float, int)
	Formats a floating point number as a string using the given number of decimal digits.
float	round_to_precision (float, const int &)
	Rounds the given floating point value to the requested number of significant digits.
void	throw_if (bool condition, const string &message, const source_location &loc=source_location::current())
template<typename T>
ostream &	operator<< (ostream &os, const vector< T > &vec)
const string &	pooling_method_to_string (PoolingMethod method)
PoolingMethod	string_to_pooling_method (const string &name)
Stats &	global_stats ()
bool &	enabled ()
void	set_seed (unsigned seed)
	Seeds the library-wide pseudo-random number generator.
long long	get_seed ()
	Returns the seed currently used by the library RNG.
float	random_uniform (float=-1, float=1)
	Draws a float uniformly in [min, max].
Index	random_integer (Index, Index)
	Draws an integer uniformly in [min, max].
bool	random_bool (float=0.5)
	Draws a boolean that is true with the given probability.
void	set_random_uniform (MatrixR &, float=-0.1, float=0.1)
	Fills the matrix with uniform random values in [min, max].
void	set_random_uniform (VectorMap, float=-0.1, float=0.1)
	Fills the vector map with uniform random values in [min, max].
void	set_random_normal (MatrixMap, float=0, float=1)
	Fills the matrix with normal random values of the given mean and standard deviation.
void	set_random_integer (MatrixR &, Index, Index)
	Fills the matrix with uniform random integers in [min, max].
void	shuffle (VectorB &vector_to_shuffle)
	Randomly permutes the entries of a boolean vector in place.
template<typename T>
void	shuffle_vector (vector< T > &)
	Randomly permutes the entries of the given vector in place.
void	shuffle_vector_blocks (vector< Index > &, size_t=20)
	Shuffles contiguous blocks of the given length while preserving in-block order.
Index	get_random_element (const vector< Index > &)
	Picks one element uniformly at random from the given vector.
float	glorot_limit (Index fan_in, Index fan_out)
	Returns the Glorot/Xavier uniform initialization limit sqrt(6 / (fan_in + fan_out)).
void	register_classes ()
void	scale_mean_standard_deviation (MatrixMap, Index, const Descriptives &)
	Standardises a column of the matrix in place using its descriptives' mean and standard deviation.
void	scale_standard_deviation (MatrixMap, Index, const Descriptives &)
	Divides a column of the matrix by its standard deviation in place.
void	scale_minimum_maximum (MatrixMap, Index, const Descriptives &, float=-1.0f, float=1.0f)
	Rescales a column to the [min_range, max_range] interval using its descriptives.
void	scale_logarithmic (MatrixMap, Index)
	Applies an element-wise logarithm to the given column.
void	unscale_minimum_maximum (MatrixMap, Index, const Descriptives &, float=-1.0f, float=1.0f)
	Inverse of scale_minimum_maximum(): reconstructs original values for the given column.
void	unscale_mean_standard_deviation (MatrixMap, Index, const Descriptives &)
	Inverse of scale_mean_standard_deviation() for the given column.
void	unscale_standard_deviation (MatrixMap, Index, const Descriptives &)
	Inverse of scale_standard_deviation() for the given column.
void	unscale_logarithmic (MatrixMap, Index)
	Inverse of scale_logarithmic() for the given column.
void	unscale_image_minimum_maximum (MatrixMap, Index)
	Maps a column back from [-1, 1] to the [0, 255] image-pixel range.
float	minimum (const MatrixR &)
	Returns the smallest finite element of a matrix.
float	minimum (const VectorR &)
	Returns the smallest finite element of a vector.
float	minimum (const VectorR &, const vector< Index > &)
	Returns the smallest finite element of the selected rows of a vector.
VectorR	column_minimums (const Tensor2 &, const vector< Index > &={}, const vector< Index > &={})
	Per-column minimums of a 2D tensor.
float	maximum (const MatrixR &)
	Returns the largest finite element of a matrix.
float	maximum (const VectorR &)
	Returns the largest finite element of a vector.
float	maximum (const VectorR &, const vector< Index > &)
	Returns the largest finite element of the selected rows of a vector.
VectorR	column_maximums (const Tensor2 &, const vector< Index > &={}, const vector< Index > &={})
	Per-column maximums of a 2D tensor, optionally restricted to a row/column subset.
float	range (const VectorR &)
	Returns the maximum minus the minimum of a vector.
float	mean (const VectorR &)
	Arithmetic mean of a vector, ignoring NaNs.
float	mean (const MatrixR &, Index)
	Arithmetic mean of a matrix column, ignoring NaNs.
VectorR	mean (const MatrixR &)
	Column-wise arithmetic means of a matrix.
VectorR	mean (const MatrixR &, const vector< Index > &, const vector< Index > &)
	Column-wise means of a matrix restricted to the given rows and columns.
float	median (const VectorR &)
	Median of a vector.
float	median (const MatrixR &, Index)
	Median of a single matrix column.
VectorR	median (const MatrixR &)
	Column-wise medians of a matrix.
VectorR	median (const MatrixR &, const vector< Index > &)
	Column-wise medians of the selected columns.
VectorR	median (const MatrixR &, const vector< Index > &, const vector< Index > &)
	Column-wise medians restricted to the given rows and columns.
float	variance (const VectorR &)
	Sample variance of a vector.
float	variance (const VectorR &, const VectorI &)
	Sample variance of the selected entries of a vector.
float	standard_deviation (const VectorR &)
	Sample standard deviation of a vector.
VectorR	standard_deviation (const VectorR &, Index)
	Rolling standard deviation of a vector over a window of the given size.
VectorR	quartiles (const VectorR &)
	Returns [Q1, Q2, Q3] of a vector.
VectorR	quartiles (const VectorR &, const vector< Index > &)
	Returns [Q1, Q2, Q3] of the selected entries of a vector.
BoxPlot	box_plot (const VectorR &)
	Five-number summary (box plot) of a vector.
BoxPlot	box_plot (const VectorR &, const vector< Index > &)
	Five-number summary (box plot) of the selected entries of a vector.
Descriptives	vector_descriptives (const VectorR &)
	Returns the (min, max, mean, std) descriptives of a vector.
vector< Descriptives >	descriptives (const MatrixR &)
	Returns the per-column descriptives of a matrix.
vector< Descriptives >	descriptives (const MatrixR &, const vector< Index > &, const vector< Index > &)
	Returns per-column descriptives restricted to the given rows and columns.
Histogram	histogram (const VectorR &, Index=10)
	Builds an equal-width histogram of a vector.
Histogram	histogram_centered (const VectorR &, float=0.0f, Index=10)
	Builds a histogram with one bin centered on the given value.
Histogram	histogram (const VectorB &)
	Builds a two-bin histogram counting false/true entries.
Histogram	histogram (const VectorI &, Index=10)
	Builds an equal-width histogram of an integer vector.
vector< Histogram >	histograms (const MatrixR &, Index=10)
	Builds one histogram per matrix column.
VectorI	total_frequencies (const vector< Histogram > &)
	Sums the per-bin frequencies across a collection of histograms.
Index	minimal_index (const VectorR &)
	Index of the smallest element in a vector.
VectorI	minimal_indices (const VectorR &, Index)
	Indices of the n smallest elements of a vector.
VectorI	minimal_indices (const MatrixR &)
	Row/column coordinates of the smallest element of a matrix.
Index	maximal_index (const VectorR &)
	Index of the largest element in a vector.
VectorI	maximal_indices (const VectorR &, Index)
	Indices of the n largest elements of a vector.
VectorI	maximal_indices (const MatrixR &)
	Row/column coordinates of the largest element of a matrix.
bool	row_finite (const VectorR &values, Index i)
	Returns true if the i-th entry of the vector is finite.
bool	row_finite (const MatrixR &matrix, Index i)
	Returns true if every entry in row i of the matrix is finite.
VectorR	slice_rows (const VectorR &values, const vector< Index > &indices)
	Returns a copy of the vector containing only the entries at the given indices.
MatrixR	slice_rows (const MatrixR &matrix, const vector< Index > &indices)
	Returns a copy of the matrix containing only the rows at the given indices.
VectorR	filter_missing_values (const VectorR &)
	Returns a copy of the vector with NaN entries removed.
template<typename X, typename Y>
pair< X, Y >	filter_missing_values (const X &x, const Y &y)
	Returns x and y restricted to rows where both are finite (row counts must match).
bool	is_contiguous (const vector< Index > &indices)
	Returns true if the sorted indices form a contiguous run (each entry equals the previous plus one).
template<typename T>
bool	is_binary (const T &tensor)
	Returns true if every non-NaN entry of the tensor is exactly 0.0 or 1.0.
MatrixR	append_rows (const MatrixR &, const MatrixR &)
	Returns the row-wise concatenation of two matrices with matching column counts.
template<typename T>
vector< T >	gather_by_index (const vector< T > &data, const vector< Index > &indices)
	Returns the elements of data at the given indices.
vector< Index >	build_feasible_rows_mask (const MatrixR &outputs, const VectorR &minimums, const VectorR &maximums)
	Returns the indices of the rows of outputs that lie within the per-column bounds.
template<typename T>
bool	is_constant (const T &tensor)
	Returns true if every non-NaN entry of the tensor equals the first finite entry.
vector< Index >	get_true_indices (const VectorB &flags)
	Returns the positions of the true entries in a boolean vector.
VectorI	calculate_rank (const VectorR &, bool ascending=true)
	Returns the rank of each element (1-based), ascending by default.
vector< Index >	get_elements_greater_than (const vector< Index > &, Index)
	Returns the entries of indices that are strictly greater than the given threshold.
VectorI	get_nearest_points (const MatrixR &, const VectorR &, int=1)
	Finds the n rows of the matrix closest to the given point by Euclidean distance.
void	fill_tensor_data (const MatrixR &, const vector< Index > &, const vector< Index > &, float *, bool=true, int contiguous=-1)
	Copies the selected sub-matrix into a flat float buffer.
VectorR	perform_Householder_QR_decomposition (const MatrixR &, const VectorR &)
	Solves a linear least-squares problem via Householder QR decomposition.
VectorMap	vector_map (const MatrixR &, Index)
	Returns an Eigen VectorMap that views a single column of a matrix without copying.
vector< string >	get_tokens (const string &, const string &)
	Splits a string on every occurrence of any character in the separator set.
vector< string_view >	get_token_views (string_view, char)
	Splits a string view on the given separator, returning views into the original buffer.
string_view	trim_view (string_view)
	Returns a view onto the input with leading and trailing whitespace removed.
vector< string >	tokenize (const string &)
	Splits the input on whitespace into individual tokens.
vector< string_view >	tokenize_views (string_view)
	Whitespace-tokenises a string view, returning views into the source buffer.
vector< string >	convert_string_vector (const vector< vector< string > > &, const string &)
	Joins each inner vector with the separator, returning one flattened string per row.
bool	is_numeric_string (string_view)
	Returns true if the string can be parsed as a numeric literal.
bool	is_date_time_string (string_view)
	Returns true if the string matches one of the supported date/time formats.
time_t	date_to_timestamp (const string &, Index=0, const DateFormat &format=AUTO)
	Parses a date/time string into a Unix timestamp.
void	replace_all_appearances (string &, const string &, const string &)
	Replaces every occurrence of a substring with another, in place.
void	replace_all_word_appearances (string &, const string &, const string &)
	Replaces every whole-word occurrence of a token with another, in place.
string	get_trimmed (const string &)
	Returns a copy of the string with leading and trailing whitespace removed.
bool	has_numbers (const vector< string > &)
	Returns true if any element of the vector parses as a number.
bool	has_numbers (const vector< string_view > &)
	Returns true if any element of the vector parses as a number.
void	replace (string &, const string &, const string &)
	In-place replacement of every occurrence of a substring with another.
void	display_progress_bar (int, int)
	Prints a textual progress bar to stdout for an in-progress operation.
string	get_time (float)
	Formats a duration in seconds as a human-readable HH:MM:SS string.
string	get_first_word (const string &)
	Returns the first whitespace-delimited word of a string.
template<typename T>
string	vector_to_string (const vector< T > &values, const string &separator=" ")
	Serializes a vector to a string with the given element separator.
template<typename Derived>
string	vector_to_string (const Eigen::DenseBase< Derived > &values, const string &separator=" ")
	Serializes an Eigen dense expression to a string with the given separator.
void	string_to_vector (const string &input, VectorR &values)
	Parses a whitespace-separated string of floats into a VectorR.
template<typename T, size_t Rank>
string	tensor_to_string (const TensorR< Rank > &values, const string &separator=" ")
	Serializes a tensor's flat data to a string with the given separator.
template<typename T, size_t Rank>
void	string_to_tensor (const string &input, TensorR< Rank > &values)
	Parses a whitespace-separated string into the flat storage of a tensor.
bool	contains (const vector< string > &, const string &)
	Returns true if the vector contains the given element.
bool	contains (const vector< string > &, string_view)
	Returns true if the vector contains the given element.
int	to_int (Index value)
float	to_type (Index value)
Index	align_up (Index value, Index alignment)
Index	get_aligned_size (Index size)
Index	get_aligned_bytes (Index n_bytes)
Index	get_aligned_bytes (Index count, Type dtype)
bool	is_aligned (const void *ptr)
Index	get_aligned_size (const vector< TensorSpec > &specs)
Index	get_aligned_size (const vector< vector< TensorSpec > > &specs)
Index	get_aligned_bytes (const vector< TensorSpec > &specs)
Index	get_aligned_bytes (const vector< vector< TensorSpec > > &specs)
Index	get_aligned_bytes (const vector< Shape > &shapes, Type dtype)
Index	get_aligned_bytes (const vector< TensorSpec > &specs, Type dtype)
Index	get_aligned_bytes (const vector< vector< TensorSpec > > &specs, Type dtype)
TensorView &	view_at_slot_or (vector< TensorView > &views, const vector< size_t > &slots, size_t i, TensorView &fallback)
TensorView &	view_at_slot_or (vector< vector< TensorView > > &views, const vector< size_t > &slots, size_t i, TensorView &fallback)
string	shape_to_string (const Shape &, const string &=" ")
	Serializes a shape as a separator-joined string of dimensions.
Shape	string_to_shape (const string &, const string &=" ")
	Parses a separator-joined string of dimensions into a Shape.
template<typename... Vs>
size_t	hash_combine (const Vs &... values)
	Boost-style hash combine that mixes any number of hashable values into a single size_t.
ThreadPoolDevice &	get_device ()
	Convenience accessor for the global Eigen ThreadPoolDevice.
const EnumMap< VariableType > &	variable_type_map ()
	Returns the bidirectional string/enum map for VariableType.
const string &	variable_type_to_string (VariableType type)
	Returns the canonical string name for a VariableType.
VariableType	string_to_variable_type (const string &name)
	Parses a string into the matching VariableType enumerator.
const EnumMap< ScalerMethod > &	scaler_method_map ()
	Returns the bidirectional string/enum map for ScalerMethod.
const string &	scaler_method_to_string (ScalerMethod method)
	Returns the canonical string name for a ScalerMethod.
ScalerMethod	string_to_scaler_method (const string &name)
	Parses a string into the matching ScalerMethod enumerator.
const EnumMap< VariableRole > &	variable_role_map ()
	Returns the bidirectional string/enum map for VariableRole.
const string &	variable_role_to_string (VariableRole role)
	Returns the canonical string name for a VariableRole.
VariableRole	string_to_variable_role (const string &name)
	Parses a string into a VariableRole; "Id" is mapped to VariableRole::None.
bool	role_matches (VariableRole actual, VariableRole query)
	Returns true if actual satisfies query, treating InputTarget as Input or Target.

Variables
constexpr float	GRADIENT_NORM_EPS = 1e-6f
constexpr float	EPSILON = numeric_limits<float>::epsilon()
constexpr float	MAX = numeric_limits<float>::max()
constexpr float	NEG_INFINITY = -numeric_limits<float>::infinity()
constexpr float	QUIET_NAN = numeric_limits<float>::quiet_NaN()
constexpr float	SOFTMAX_MASK_VALUE = float(-1e9f)
static constexpr Index	ALIGN_BYTES = EIGEN_MAX_ALIGN_BYTES
static constexpr Index	ALIGN_ELEMENTS = ALIGN_BYTES / sizeof(float)
constexpr cudaDataType_t	CUDA_REDUCTION_DTYPE = CUDA_R_32F
constexpr cublasComputeType_t	CUBLAS_COMPUTE_DTYPE = CUBLAS_COMPUTE_32F_FAST_TF32

Typedef Documentation

array

template<typename T, size_t N>

using opennn::array = Eigen::array<T, N>

bfloat16

using opennn::bfloat16 = __nv_bfloat16

Enumeration Type Documentation

DateFormat

enum opennn::DateFormat

Order of the day, month, and year fields in a date string (AUTO probes the input).

Enumerator
AUTO
DMY
MDY
YMD

Device

enum class opennn::Device

strong

Execution device selection for OpenNN runtime (auto-detected, CPU or CUDA GPU).

Enumerator
Auto
CPU
CUDA

LayerType

enum class opennn::LayerType

strong

Identifier of every concrete layer subclass shipped with OpenNN.

Enumerator
Addition
Bounding
Convolutional
ConvolutionalRelu
Dense
DenseRelu
Embedding
Flatten
MultiHeadAttention
Normalization3d
Pooling
Pooling3d
Recurrent
Scaling
Unscaling

PoolingMethod

enum class opennn::PoolingMethod

strong

Pooling reduction method used by Pooling and Pooling3d layers.

Enumerator
MaxPooling
AveragePooling

SampleRole

enum class opennn::SampleRole

strong

Role of a sample in a dataset split.

Enumerator
Training
Validation
Testing
None

ScalerMethod

enum class opennn::ScalerMethod

strong

Feature scaling strategy applied to a Variable before training.

Enumerator
None
MinimumMaximum
MeanStandardDeviation
StandardDeviation
Logarithm
ImageMinMax

Type

enum class opennn::Type

strong

Numeric precision used for training or inference tensors.

Enumerator
Auto
FP32
BF16
INT8

VariableRole

enum class opennn::VariableRole

strong

Role a Variable plays in a dataset (input feature, target, decoder, time axis, etc.).

Enumerator
None
Input
Target
Decoder
InputTarget
Time

VariableType

enum class opennn::VariableType

strong

Data type of a dataset Variable.

Enumerator
None
Numeric
Binary
Categorical
DateTime
Constant

Function Documentation

add()

void opennn::add	(	const TensorView &	input_1,
		const TensorView &	input_2,
		TensorView &	output )

Element-wise addition: output = input_1 + input_2.

add_cpu()

void opennn::add_cpu	(	const TensorView &	input_1,
		const TensorView &	input_2,
		TensorView &	output )

CPU implementation of add().

add_json_field()

void opennn::add_json_field	(	JsonWriter &	writer,
		const string &	name,
		const string &	value )

align_up()

Index opennn::align_up ( Index value, Index alignment )

inlinenodiscard

append_rows()

MatrixR opennn::append_rows ( const MatrixR & , const MatrixR &  )

nodiscard

Returns the row-wise concatenation of two matrices with matching column counts.

atomic_rename()

void opennn::atomic_rename	(	const filesystem::path &	from,
		const filesystem::path &	to )

Atomically renames a file, replacing the destination if needed.

autocorrelations()

VectorR opennn::autocorrelations ( const VectorR & , Index = 10 )

nodiscard

Autocorrelations of a series for lags 0..max_lag.

Parameters

series	Input time series.
max_lag	Maximum lag to compute (default 10).

average_pooling_3d_backward()

void opennn::average_pooling_3d_backward	(	const TensorView &	input,
		const TensorView &	output_delta,
		TensorView &	input_delta )

Backward pass for 3D average pooling.

average_pooling_3d_backward_cpu()

void opennn::average_pooling_3d_backward_cpu	(	const TensorView &	input,
		const TensorView &	output_delta,
		TensorView &	input_delta )

CPU implementation of average_pooling_3d_backward().

average_pooling_3d_forward()

void opennn::average_pooling_3d_forward	(	const TensorView &	input,
		TensorView &	output )

Forward pass of 3D average pooling.

average_pooling_3d_forward_cpu()

void opennn::average_pooling_3d_forward_cpu	(	const TensorView &	input,
		TensorView &	output )

CPU implementation of average_pooling_3d_forward().

binary_cross_entropy()

void opennn::binary_cross_entropy	(	const TensorView &	input,
		const TensorView &	target,
		float &	error,
		float *	workspace_device )

Computes the binary cross-entropy between predicted probabilities and binary targets.

bound()

void opennn::bound	(	const TensorView &	input,
		const TensorView &	lower_bounds,
		const TensorView &	upper_bounds,
		TensorView &	output )

Clamps each element of input to the [lower_bounds, upper_bounds] range.

Parameters

input	Source tensor.
lower_bounds	Per-element (or broadcastable) lower limits.
upper_bounds	Per-element (or broadcastable) upper limits.
output	Destination tensor for the bounded values.

bound_cpu()

void opennn::bound_cpu	(	const TensorView &	input,
		const TensorView &	lower_bounds,
		const TensorView &	upper_bounds,
		TensorView &	output )

CPU implementation of bound().

box_plot() [1/2]

BoxPlot opennn::box_plot ( const VectorR & )

nodiscard

Five-number summary (box plot) of a vector.

box_plot() [2/2]

BoxPlot opennn::box_plot ( const VectorR & , const vector< Index > &  )

nodiscard

Five-number summary (box plot) of the selected entries of a vector.

build_feasible_rows_mask()

vector< Index > opennn::build_feasible_rows_mask ( const MatrixR & outputs, const VectorR & minimums, const VectorR & maximums )

nodiscard

Returns the indices of the rows of outputs that lie within the per-column bounds.

calculate_rank()

VectorI opennn::calculate_rank ( const VectorR & , bool ascending = true )

nodiscard

Returns the rank of each element (1-based), ascending by default.

calculate_spearman_ranks()

VectorR opennn::calculate_spearman_ranks ( const VectorR & )

nodiscard

Computes Spearman ranks (average rank for ties) for the entries of a vector.

categorical_cross_entropy()

void opennn::categorical_cross_entropy	(	const TensorView &	input,
		const TensorView &	target,
		float &	error,
		float *	workspace_device )

Computes the multi-class (categorical) cross-entropy between softmax probabilities and one-hot targets.

check_rank()

void opennn::check_rank ( const Shape & shape, initializer_list< int > allowed, const char * layer, const char * what )

inline

Throws if shape rank is not one of allowed.

Parameters

shape	Shape to validate (empty shapes are accepted).
allowed	Acceptable rank values.
layer	Layer name used in the error message.
what	Description of the tensor being checked (e.g. "input").

column_maximums()

VectorR opennn::column_maximums ( const Tensor2 & , const vector< Index > & = {}, const vector< Index > & = {} )

nodiscard

Per-column maximums of a 2D tensor, optionally restricted to a row/column subset.

column_minimums()

VectorR opennn::column_minimums ( const Tensor2 & , const vector< Index > & = {}, const vector< Index > & = {} )

nodiscard

Per-column minimums of a 2D tensor.

Parameters

matrix	Source matrix.
row_indices	Optional row subset (empty = all rows).
column_indices	Optional column subset (empty = all columns).

concatenate_string_vectors()

vector< string > opennn::concatenate_string_vectors	(	const vector< string > &	,
		const vector< string > &	)

Concatenates two string vectors and returns the resulting vector.

confidence_interval_z_correlation()

pair< float, float > opennn::confidence_interval_z_correlation ( const float , Index  )

nodiscard

Returns the [lower, upper] confidence interval for a correlation given its sample size.

contains() [1/2]

bool opennn::contains ( const vector< string > & , const string &  )

nodiscard

Returns true if the vector contains the given element.

contains() [2/2]

bool opennn::contains ( const vector< string > & , string_view  )

nodiscard

Returns true if the vector contains the given element.

convert_string_vector()

vector< string > opennn::convert_string_vector ( const vector< vector< string > > & , const string &  )

nodiscard

Joins each inner vector with the separator, returning one flattened string per row.

copy()

void opennn::copy	(	const TensorView &	source,
		TensorView &	destination )

Copies the contents of source into destination, dispatching to CPU or GPU as needed.

copy_cpu()

void opennn::copy_cpu	(	const TensorView &	source,
		TensorView &	destination )

CPU implementation of copy().

correlation()

Correlation opennn::correlation ( const MatrixR & , const MatrixR &  )

nodiscard

Generic correlation between two matrices, dispatching on column types (binary, categorical, continuous).

correlation_spearman()

Correlation opennn::correlation_spearman ( const MatrixR & , const MatrixR &  )

nodiscard

Generic Spearman correlation between two matrices, dispatching on column types.

cross_correlations()

VectorR opennn::cross_correlations ( const VectorR & , const VectorR & , Index  )

nodiscard

Cross-correlations between two series for lags 0..max_lag.

cross_entropy_3d()

void opennn::cross_entropy_3d	(	const TensorView &	input,
		const TensorView &	target,
		float &	error,
		Index &	active_tokens_out,
		Index &	correct_tokens_out,
		float *	errors_device = nullptr )

Computes 3-D (sequence) cross-entropy used by transformer-style targets, ignoring padded positions.

Parameters

input	Predicted logits/probabilities with shape (batch, sequence, vocab).
target	Token ids or one-hot targets with the matching sequence layout.
error	Output scalar holding the cumulative error across active tokens.
active_tokens_out	Output count of non-padding tokens that contributed to the error.
correct_tokens_out	Output count of correctly predicted tokens (for accuracy).
errors_device	Optional device-side reduction workspace.

cross_entropy_3d_gradient()

void opennn::cross_entropy_3d_gradient	(	const TensorView &	input,
		const TensorView &	target,
		const TensorView &	input_delta,
		Index	active_tokens_count )

Writes the 3-D cross-entropy gradient into input_delta, normalizing by the host-side active-token count.

cross_entropy_3d_gradient_device_count()

void opennn::cross_entropy_3d_gradient_device_count	(	const TensorView &	input,
		const TensorView &	target,
		const TensorView &	input_delta,
		const float *	active_tokens_count_device )

Variant of cross_entropy_3d_gradient that reads the active-token count from device memory.

cross_entropy_gradient()

void opennn::cross_entropy_gradient	(	const TensorView &	input,
		const TensorView &	target,
		const TensorView &	input_delta )

Writes the cross-entropy gradient with respect to the (pre-softmax/logit) predictions into input_delta.

current_device()

Device opennn::current_device ( )

inlinenodiscard

Returns the active runtime device (CUDA if available, otherwise CPU).

date_to_timestamp()

time_t opennn::date_to_timestamp ( const string & , Index = 0, const DateFormat & format = AUTO )

nodiscard

Parses a date/time string into a Unix timestamp.

Parameters

date_string	Date or datetime literal to parse.
gmt_offset	Number of hours to add when converting to UTC.
format	Day/month/year ordering, or AUTO to auto-detect.

descriptives() [1/2]

vector< Descriptives > opennn::descriptives ( const MatrixR & )

nodiscard

Returns the per-column descriptives of a matrix.

descriptives() [2/2]

vector< Descriptives > opennn::descriptives ( const MatrixR & , const vector< Index > & , const vector< Index > &  )

nodiscard

Returns per-column descriptives restricted to the given rows and columns.

display_progress_bar()

void opennn::display_progress_bar	(	int	,
		int	)

Prints a textual progress bar to stdout for an in-progress operation.

Parameters

progress_index	Current step.
total	Total number of steps.

enabled()

bool & opennn::enabled ( )

inline

eta_squared_correlation()

Correlation opennn::eta_squared_correlation ( const VectorR & , const MatrixR &  )

nodiscard

Eta-squared (effect-size) correlation between a continuous vector and a categorical matrix.

exponential_correlation()

Correlation opennn::exponential_correlation ( const VectorR & , const VectorR &  )

nodiscard

Exponential correlation: fits y = a * exp(b * x) and returns the resulting fit.

fill_tensor_data()

void opennn::fill_tensor_data	(	const MatrixR &	,
		const vector< Index > &	,
		const vector< Index > &	,
		float *	,
		bool	= true,
		int	contiguous = -1 )

Copies the selected sub-matrix into a flat float buffer.

Parameters

matrix	Source data matrix.
row_indices	Rows to copy.
column_indices	Columns to copy.
data	Destination buffer (caller-owned).
row_major	Lay rows contiguously when true, columns when false.
contiguous	Optional fast-path hint when -1 the layout is inferred.

filter_missing_values() [1/2]

VectorR opennn::filter_missing_values ( const VectorR & )

nodiscard

Returns a copy of the vector with NaN entries removed.

filter_missing_values() [2/2]

template<typename X, typename Y>

pair< X, Y > opennn::filter_missing_values ( const X & x, const Y & y )

nodiscard

Returns x and y restricted to rows where both are finite (row counts must match).

for_json_items()

template<typename Func>

void opennn::for_json_items	(	const Json *	parent,
		const char *	tag,
		long	count,
		Func	func )

formatNumber()

string opennn::formatNumber	(	float	,
		int	)

Formats a floating point number as a string using the given number of decimal digits.

gather_by_index()

template<typename T>

vector< T > opennn::gather_by_index ( const vector< T > & data, const vector< Index > & indices )

nodiscard

Returns the elements of data at the given indices.

get_aligned_bytes() [1/7]

Index opennn::get_aligned_bytes ( const vector< Shape > & shapes, Type dtype )

inlinenodiscard

get_aligned_bytes() [2/7]

Index opennn::get_aligned_bytes ( const vector< TensorSpec > & specs )

inlinenodiscard

get_aligned_bytes() [3/7]

Index opennn::get_aligned_bytes ( const vector< TensorSpec > & specs, Type dtype )

inlinenodiscard

get_aligned_bytes() [4/7]

Index opennn::get_aligned_bytes ( const vector< vector< TensorSpec > > & specs )

inlinenodiscard

get_aligned_bytes() [5/7]

Index opennn::get_aligned_bytes ( const vector< vector< TensorSpec > > & specs, Type dtype )

inlinenodiscard

get_aligned_bytes() [6/7]

Index opennn::get_aligned_bytes ( Index count, Type dtype )

inlinenodiscard

get_aligned_bytes() [7/7]

Index opennn::get_aligned_bytes ( Index n_bytes )

inlinenodiscard

get_aligned_size() [1/3]

Index opennn::get_aligned_size ( const vector< TensorSpec > & specs )

inlinenodiscard

get_aligned_size() [2/3]

Index opennn::get_aligned_size ( const vector< vector< TensorSpec > > & specs )

inlinenodiscard

get_aligned_size() [3/3]

Index opennn::get_aligned_size ( Index size )

inlinenodiscard

get_correlation_values()

MatrixR opennn::get_correlation_values ( const Tensor< Correlation, 2 > & )

nodiscard

Extracts the coefficient r from a 2D tensor of Correlation values.

get_device()

ThreadPoolDevice & opennn::get_device ( )

inline

Convenience accessor for the global Eigen ThreadPoolDevice.

get_elements_greater_than()

vector< Index > opennn::get_elements_greater_than ( const vector< Index > & , Index  )

nodiscard

Returns the entries of indices that are strictly greater than the given threshold.

get_first_word()

string opennn::get_first_word ( const string & )

nodiscard

Returns the first whitespace-delimited word of a string.

get_json_root()

const Json * opennn::get_json_root ( const JsonDocument & document, const string & tag )

nodiscard

get_nearest_points()

VectorI opennn::get_nearest_points ( const MatrixR & , const VectorR & , int = 1 )

nodiscard

Finds the n rows of the matrix closest to the given point by Euclidean distance.

get_random_element()

Index opennn::get_random_element ( const vector< Index > & )

nodiscard

Picks one element uniformly at random from the given vector.

get_seed()

long long opennn::get_seed ( )

nodiscard

Returns the seed currently used by the library RNG.

get_time()

string opennn::get_time ( float )

nodiscard

Formats a duration in seconds as a human-readable HH:MM:SS string.

get_token_views()

vector< string_view > opennn::get_token_views ( string_view , char  )

nodiscard

Splits a string view on the given separator, returning views into the original buffer.

get_tokens()

vector< string > opennn::get_tokens ( const string & , const string &  )

nodiscard

Splits a string on every occurrence of any character in the separator set.

Parameters

input	Source string.
separators	Set of single-character separators.

Returns

Tokens, with empty tokens dropped.

get_trimmed()

string opennn::get_trimmed ( const string & )

nodiscard

Returns a copy of the string with leading and trailing whitespace removed.

get_true_indices()

vector< Index > opennn::get_true_indices ( const VectorB & flags )

inlinenodiscard

Returns the positions of the true entries in a boolean vector.

global_stats()

Stats & opennn::global_stats ( )

inline

glorot_limit()

float opennn::glorot_limit ( Index fan_in, Index fan_out )

inlinenodiscard

Returns the Glorot/Xavier uniform initialization limit sqrt(6 / (fan_in + fan_out)).

has_numbers() [1/2]

bool opennn::has_numbers ( const vector< string > & )

nodiscard

Returns true if any element of the vector parses as a number.

has_numbers() [2/2]

bool opennn::has_numbers ( const vector< string_view > & )

nodiscard

Returns true if any element of the vector parses as a number.

hash_combine()

template<typename... Vs>

size_t opennn::hash_combine ( const Vs &... values )

nodiscard

Boost-style hash combine that mixes any number of hashable values into a single size_t.

histogram() [1/3]

Histogram opennn::histogram ( const VectorB & )

nodiscard

Builds a two-bin histogram counting false/true entries.

histogram() [2/3]

Histogram opennn::histogram ( const VectorI & , Index = 10 )

nodiscard

Builds an equal-width histogram of an integer vector.

histogram() [3/3]

Histogram opennn::histogram ( const VectorR & , Index = 10 )

nodiscard

Builds an equal-width histogram of a vector.

Parameters

values	Input data.
bins	Number of equal-width bins.

histogram_centered()

Histogram opennn::histogram_centered ( const VectorR & , float = 0.0f, Index = 10 )

nodiscard

Builds a histogram with one bin centered on the given value.

histograms()

vector< Histogram > opennn::histograms ( const MatrixR & , Index = 10 )

nodiscard

Builds one histogram per matrix column.

is_aligned()

bool opennn::is_aligned ( const void * ptr )

inlinenodiscard

is_bf16_inference()

bool opennn::is_bf16_inference ( )

inlinenodiscard

Returns true when inference is configured to use BF16 precision.

is_bf16_training()

bool opennn::is_bf16_training ( )

inlinenodiscard

Returns true when training is configured to use BF16 precision.

is_binary()

template<typename T>

bool opennn::is_binary ( const T & tensor )

inlinenodiscard

Returns true if every non-NaN entry of the tensor is exactly 0.0 or 1.0.

is_constant()

template<typename T>

bool opennn::is_constant ( const T & tensor )

inlinenodiscard

Returns true if every non-NaN entry of the tensor equals the first finite entry.

is_contiguous()

bool opennn::is_contiguous ( const vector< Index > & indices )

inlinenodiscard

Returns true if the sorted indices form a contiguous run (each entry equals the previous plus one).

is_cpu()

bool opennn::is_cpu ( )

inlinenodiscard

Returns true when the resolved configuration runs on CPU.

is_date_time_string()

bool opennn::is_date_time_string ( string_view )

nodiscard

Returns true if the string matches one of the supported date/time formats.

is_gpu()

bool opennn::is_gpu ( )

inlinenodiscard

Returns true when the resolved configuration runs on a CUDA GPU.

is_numeric_string()

bool opennn::is_numeric_string ( string_view )

nodiscard

Returns true if the string can be parsed as a numeric literal.

l1_regularization()

void opennn::l1_regularization	(	const TensorView &	parameters,
		float	lambda,
		float &	penalty )

Computes the L1 regularization penalty lambda * sum(|parameters|).

l1_regularization_gradient()

void opennn::l1_regularization_gradient	(	const TensorView &	parameters,
		float	lambda,
		const TensorView &	gradient )

Adds the L1 regularization gradient lambda * sign(parameters) into the gradient tensor.

l2_regularization()

void opennn::l2_regularization	(	const TensorView &	parameters,
		float	lambda,
		float &	penalty )

Computes the L2 regularization penalty lambda * sum(parameters^2).

l2_regularization_gradient()

void opennn::l2_regularization_gradient	(	const TensorView &	parameters,
		float	lambda,
		const TensorView &	gradient )

Adds the L2 regularization gradient 2 * lambda * parameters into the gradient tensor.

layer_type_map()

const EnumMap< LayerType > & opennn::layer_type_map ( )

inlinenodiscard

Returns the bidirectional mapping between LayerType values and their string names.

layer_type_to_string()

const string & opennn::layer_type_to_string ( LayerType type )

inlinenodiscard

Returns the string name associated with the given LayerType.

linear_correlation()

Correlation opennn::linear_correlation ( const VectorR & , const VectorR &  )

nodiscard

Pearson linear correlation between two equal-length vectors.

Parameters

x	First variable.
y	Second variable.

Returns

Correlation with intercept a, slope b, and coefficient r.

linear_correlation_spearman()

Correlation opennn::linear_correlation_spearman ( const VectorR & , const VectorR &  )

nodiscard

Spearman rank correlation between two vectors using a linear fit on ranks.

load_image() [1/2]

Tensor3 opennn::load_image

(

const filesystem::path &

)

Loads an image from disk into a rank-3 (height, width, channels) tensor.

load_image() [2/2]

void opennn::load_image	(	const filesystem::path &	,
		float *	dst,
		Index	expected_height,
		Index	expected_width,
		Index	expected_channels,
		bool	divide_by_255 = false )

Loads an image into a pre-allocated float buffer at the given shape.

Parameters

path	Image file on disk.
dst	Destination buffer with capacity expected_height * expected_width * expected_channels.
expected_height	Required output height (image is resized if it differs).
expected_width	Required output width.
expected_channels	Required number of channels.
divide_by_255	Divide pixel values by 255 to produce a [0, 1] range.

load_json_file()

JsonDocument opennn::load_json_file ( const filesystem::path & file_name )

nodiscard

logarithmic_correlation()

Correlation opennn::logarithmic_correlation ( const VectorR & , const VectorR &  )

nodiscard

Logarithmic correlation: fits y = a + b * log(x) and returns the resulting fit.

logistic_correlation() [1/4]

Correlation opennn::logistic_correlation ( const MatrixR & , const MatrixR &  )

nodiscard

Logistic correlation between two one-hot matrices.

logistic_correlation() [2/4]

Correlation opennn::logistic_correlation ( const MatrixR & , const VectorR &  )

nodiscard

Logistic correlation between a one-hot matrix predictor and a vector target.

logistic_correlation() [3/4]

Correlation opennn::logistic_correlation ( const VectorR & , const MatrixR &  )

nodiscard

Logistic correlation between a vector predictor and a one-hot matrix target.

logistic_correlation() [4/4]

Correlation opennn::logistic_correlation ( const VectorR & , const VectorR &  )

nodiscard

Logistic correlation between two vectors (binary or continuous targets).

logistic_correlation_spearman()

Correlation opennn::logistic_correlation_spearman ( const VectorR & , const VectorR &  )

nodiscard

Spearman-rank logistic correlation between two vectors.

max_pooling_3d_backward()

void opennn::max_pooling_3d_backward	(	const TensorView &	maximal_indices,
		const TensorView &	output_delta,
		TensorView &	input_delta )

Backward pass for 3D max pooling; routes gradients to argmax positions.

max_pooling_3d_backward_cpu()

void opennn::max_pooling_3d_backward_cpu	(	const TensorView &	maximal_indices,
		const TensorView &	output_delta,
		TensorView &	input_delta )

CPU implementation of max_pooling_3d_backward().

max_pooling_3d_forward()

void opennn::max_pooling_3d_forward	(	const TensorView &	input,
		TensorView &	output,
		TensorView &	maximal_indices,
		bool	is_training )

Forward pass of 3D max pooling; records argmax positions when training.

Parameters

input	Source activations.
output	Pooled output activations.
maximal_indices	Positions of the maxima within each pooling window.
is_training	If true, populates maximal_indices for use by the backward pass.

max_pooling_3d_forward_cpu()

void opennn::max_pooling_3d_forward_cpu	(	const TensorView &	input,
		TensorView &	output,
		TensorView &	maximal_indices,
		bool	is_training )

CPU implementation of max_pooling_3d_forward().

maximal_index()

Index opennn::maximal_index ( const VectorR & )

nodiscard

Index of the largest element in a vector.

maximal_indices() [1/2]

VectorI opennn::maximal_indices ( const MatrixR & )

nodiscard

Row/column coordinates of the largest element of a matrix.

maximal_indices() [2/2]

VectorI opennn::maximal_indices ( const VectorR & , Index  )

nodiscard

Indices of the n largest elements of a vector.

maximum() [1/3]

float opennn::maximum ( const MatrixR & )

nodiscard

Returns the largest finite element of a matrix.

maximum() [2/3]

float opennn::maximum ( const VectorR & )

nodiscard

Returns the largest finite element of a vector.

maximum() [3/3]

float opennn::maximum ( const VectorR & , const vector< Index > &  )

nodiscard

Returns the largest finite element of the selected rows of a vector.

mean() [1/4]

VectorR opennn::mean ( const MatrixR & )

nodiscard

Column-wise arithmetic means of a matrix.

mean() [2/4]

VectorR opennn::mean ( const MatrixR & , const vector< Index > & , const vector< Index > &  )

nodiscard

Column-wise means of a matrix restricted to the given rows and columns.

mean() [3/4]

float opennn::mean ( const MatrixR & , Index  )

nodiscard

Arithmetic mean of a matrix column, ignoring NaNs.

mean() [4/4]

float opennn::mean ( const VectorR & )

nodiscard

Arithmetic mean of a vector, ignoring NaNs.

mean_squared_error()

void opennn::mean_squared_error	(	const TensorView &	input,
		const TensorView &	target,
		float &	error,
		float *	workspace_device )

Computes the mean squared error between predictions and targets.

Parameters

input	Predicted values.
target	Ground-truth targets with the same shape as input.
error	Output scalar holding the resulting MSE.
workspace_device	Optional device-side reduction workspace (ignored on CPU).

mean_squared_error_gradient()

void opennn::mean_squared_error_gradient	(	const TensorView &	input,
		const TensorView &	target,
		const TensorView &	input_delta )

Writes the MSE gradient with respect to the predictions into input_delta.

Parameters

input	Predicted values.
target	Ground-truth targets.
input_delta	Output gradient tensor with the same shape as input.

median() [1/5]

VectorR opennn::median ( const MatrixR & )

nodiscard

Column-wise medians of a matrix.

median() [2/5]

VectorR opennn::median ( const MatrixR & , const vector< Index > &  )

nodiscard

Column-wise medians of the selected columns.

median() [3/5]

VectorR opennn::median ( const MatrixR & , const vector< Index > & , const vector< Index > &  )

nodiscard

Column-wise medians restricted to the given rows and columns.

median() [4/5]

float opennn::median ( const MatrixR & , Index  )

nodiscard

Median of a single matrix column.

median() [5/5]

float opennn::median ( const VectorR & )

nodiscard

Median of a vector.

merge_heads()

void opennn::merge_heads	(	const TensorView &	source,
		TensorView &	destination )

Inverse of split_heads(); merges per-head tensors back into a single representation.

merge_heads_cpu()

void opennn::merge_heads_cpu	(	const TensorView &	source,
		TensorView &	destination )

CPU implementation of merge_heads().

minimal_index()

Index opennn::minimal_index ( const VectorR & )

nodiscard

Index of the smallest element in a vector.

minimal_indices() [1/2]

VectorI opennn::minimal_indices ( const MatrixR & )

nodiscard

Row/column coordinates of the smallest element of a matrix.

minimal_indices() [2/2]

VectorI opennn::minimal_indices ( const VectorR & , Index  )

nodiscard

Indices of the n smallest elements of a vector.

minimum() [1/3]

float opennn::minimum ( const MatrixR & )

nodiscard

Returns the smallest finite element of a matrix.

minimum() [2/3]

float opennn::minimum ( const VectorR & )

nodiscard

Returns the smallest finite element of a vector.

minimum() [3/3]

float opennn::minimum ( const VectorR & , const vector< Index > &  )

nodiscard

Returns the smallest finite element of the selected rows of a vector.

minkowski_error()

void opennn::minkowski_error	(	const TensorView &	input,
		const TensorView &	target,
		float	power,
		float &	error,
		float *	workspace_device )

Computes the Minkowski error sum(|input - target|^power) for the given power exponent.

minkowski_error_gradient()

void opennn::minkowski_error_gradient	(	const TensorView &	input,
		const TensorView &	target,
		float	power,
		const TensorView &	input_delta )

Writes the Minkowski-error gradient with respect to the predictions into input_delta.

multiply()

void opennn::multiply	(	const TensorView &	input_a,
		bool	transpose_a,
		const TensorView &	input_b,
		bool	transpose_b,
		TensorView &	output,
		float	alpha = 1.0f,
		float	beta = 0.0f )

General matrix multiply: output = alpha * op(input_a) * op(input_b) + beta * output.

Parameters

input_a	Left operand.
transpose_a	Transpose input_a before multiplying.
input_b	Right operand.
transpose_b	Transpose input_b before multiplying.
output	Destination tensor (also read when beta != 0).
alpha	Scale factor applied to the matrix product.
beta	Scale factor applied to the existing output.

multiply_cpu()

void opennn::multiply_cpu	(	const TensorView &	input_a,
		bool	transpose_a,
		const TensorView &	input_b,
		bool	transpose_b,
		TensorView &	output,
		float	alpha = 1.0f,
		float	beta = 0.0f )

CPU implementation of multiply().

normalized_squared_error()

void opennn::normalized_squared_error	(	const TensorView &	input,
		const TensorView &	target,
		float	coefficient,
		float &	error,
		float *	workspace_device )

Computes the squared error normalized by a dataset-level coefficient.

Parameters

input	Predicted values.
target	Ground-truth targets.
coefficient	Normalization coefficient (typically the target variance).
error	Output scalar holding the normalized squared error.
workspace_device	Optional device-side reduction workspace.

normalized_squared_error_gradient()

void opennn::normalized_squared_error_gradient	(	const TensorView &	input,
		const TensorView &	target,
		float	coefficient,
		const TensorView &	input_delta )

Writes the normalized-squared-error gradient with respect to the predictions into input_delta.

operator<<()

template<typename T>

ostream & opennn::operator<<	(	ostream &	os,
		const vector< T > &	vec )

pad()

void opennn::pad	(	const TensorView &	input,
		TensorView &	output )

Pads the input tensor and writes the result into output.

perform_Householder_QR_decomposition()

VectorR opennn::perform_Householder_QR_decomposition ( const MatrixR & , const VectorR &  )

nodiscard

Solves a linear least-squares problem via Householder QR decomposition.

point_biserial_correlation()

Correlation opennn::point_biserial_correlation ( const VectorR & , const VectorR &  )

nodiscard

Point-biserial correlation between a binary vector and a continuous vector.

pooling_method_to_string()

const string & opennn::pooling_method_to_string ( PoolingMethod method )

inline

power_correlation()

Correlation opennn::power_correlation ( const VectorR & , const VectorR &  )

nodiscard

Power correlation: fits y = a * x^b and returns the resulting fit.

quartiles() [1/2]

VectorR opennn::quartiles ( const VectorR & )

nodiscard

Returns [Q1, Q2, Q3] of a vector.

quartiles() [2/2]

VectorR opennn::quartiles ( const VectorR & , const vector< Index > &  )

nodiscard

Returns [Q1, Q2, Q3] of the selected entries of a vector.

r_correlation_to_z_correlation()

float opennn::r_correlation_to_z_correlation ( const float )

nodiscard

Fisher r-to-z transform of a correlation coefficient.

random_bool()

bool opennn::random_bool ( float = 0.5 )

nodiscard

Draws a boolean that is true with the given probability.

random_integer()

Index opennn::random_integer ( Index , Index  )

nodiscard

Draws an integer uniformly in [min, max].

random_uniform()

float opennn::random_uniform ( float = -1, float = 1 )

nodiscard

Draws a float uniformly in [min, max].

range()

float opennn::range ( const VectorR & )

nodiscard

Returns the maximum minus the minimum of a vector.

read_json_bool()

bool opennn::read_json_bool ( const Json * root, const string & field )

nodiscard

read_json_index()

long opennn::read_json_index ( const Json * root, const string & field )

nodiscard

read_json_string()

string opennn::read_json_string ( const Json * root, const string & field )

nodiscard

read_json_string_fallback()

string opennn::read_json_string_fallback ( const Json * root, initializer_list< string > names )

nodiscard

read_json_type()

float opennn::read_json_type ( const Json * root, const string & field )

nodiscard

reflect_image_horizontal()

void opennn::reflect_image_horizontal

(

Tensor3 &

)

Mirrors the image horizontally (left-right) in place.

reflect_image_vertical()

void opennn::reflect_image_vertical

(

Tensor3 &

)

Mirrors the image vertically (top-bottom) in place.

register_classes()

void opennn::register_classes

(

)

replace()

void opennn::replace	(	string &	,
		const string &	,
		const string &	)

In-place replacement of every occurrence of a substring with another.

replace_all_appearances()

void opennn::replace_all_appearances	(	string &	,
		const string &	,
		const string &	)

Replaces every occurrence of a substring with another, in place.

replace_all_word_appearances()

void opennn::replace_all_word_appearances	(	string &	,
		const string &	,
		const string &	)

Replaces every whole-word occurrence of a token with another, in place.

require_json_field()

const Json * opennn::require_json_field ( const Json * root, const string & field )

nodiscard

resize_image()

Tensor3 opennn::resize_image	(	const Tensor3 &	,
		Index	,
		Index	)

Returns a resized copy of an image at the requested height and width.

role_matches()

bool opennn::role_matches ( VariableRole actual, VariableRole query )

inline

Returns true if actual satisfies query, treating InputTarget as Input or Target.

rotate_image()

void opennn::rotate_image	(	const Tensor3 &	,
		Tensor3 &	,
		float	)

Rotates the image by the given angle (radians) into the destination tensor.

round_to_precision()

float opennn::round_to_precision	(	float	,
		const int &	)

Rounds the given floating point value to the requested number of significant digits.

row_finite() [1/2]

bool opennn::row_finite ( const MatrixR & matrix, Index i )

inlinenodiscard

Returns true if every entry in row i of the matrix is finite.

row_finite() [2/2]

bool opennn::row_finite ( const VectorR & values, Index i )

inlinenodiscard

Returns true if the i-th entry of the vector is finite.

sample_role_map()

const EnumMap< SampleRole > & opennn::sample_role_map ( )

inlinenodiscard

Returns the bidirectional string/enum map for SampleRole.

sample_role_to_string()

const string & opennn::sample_role_to_string ( SampleRole role )

inlinenodiscard

Returns the canonical string name for a SampleRole.

scale()

void opennn::scale	(	const TensorView &	input,
		const TensorView &	minimums,
		const TensorView &	maximums,
		const TensorView &	means,
		const TensorView &	standard_deviations,
		const TensorView &	scalers,
		float	min_range,
		float	max_range,
		TensorView &	output )

Applies per-feature scaling (mean/std, min/max, or other scalers) to a tensor.

Parameters

input	Source tensor.
minimums	Column-wise minimums used for min-max scaling.
maximums	Column-wise maximums used for min-max scaling.
means	Column-wise means used for standardisation.
standard_deviations	Column-wise standard deviations.
scalers	Per-column scaler kind selector.
min_range	Lower bound of the target range for min-max scaling.
max_range	Upper bound of the target range for min-max scaling.
output	Destination tensor for the scaled values.

scale_cpu()

void opennn::scale_cpu	(	const TensorView &	input,
		const TensorView &	minimums,
		const TensorView &	maximums,
		const TensorView &	means,
		const TensorView &	standard_deviations,
		const TensorView &	scalers,
		float	min_range,
		float	max_range,
		TensorView &	output )

CPU implementation of scale().

scale_logarithmic()

void opennn::scale_logarithmic	(	MatrixMap	,
		Index	)

Applies an element-wise logarithm to the given column.

scale_mean_standard_deviation()

void opennn::scale_mean_standard_deviation	(	MatrixMap	,
		Index	,
		const Descriptives &	)

Standardises a column of the matrix in place using its descriptives' mean and standard deviation.

Parameters

matrix	Data matrix to modify in place.
column	Column index to scale.
descriptives	Pre-computed descriptives of the column.

scale_minimum_maximum()

void opennn::scale_minimum_maximum	(	MatrixMap	,
		Index	,
		const Descriptives &	,
		float	= -1.0f,
		float	= 1.0f )

Rescales a column to the [min_range, max_range] interval using its descriptives.

scale_standard_deviation()

void opennn::scale_standard_deviation	(	MatrixMap	,
		Index	,
		const Descriptives &	)

Divides a column of the matrix by its standard deviation in place.

scaler_method_map()

const EnumMap< ScalerMethod > & opennn::scaler_method_map ( )

inline

Returns the bidirectional string/enum map for ScalerMethod.

scaler_method_to_string()

const string & opennn::scaler_method_to_string ( ScalerMethod method )

inline

Returns the canonical string name for a ScalerMethod.

set_random_integer()

void opennn::set_random_integer	(	MatrixR &	,
		Index	,
		Index	)

Fills the matrix with uniform random integers in [min, max].

set_random_normal()

void opennn::set_random_normal	(	MatrixMap	,
		float	= 0,
		float	= 1 )

Fills the matrix with normal random values of the given mean and standard deviation.

set_random_uniform() [1/2]

void opennn::set_random_uniform	(	MatrixR &	,
		float	= -0.1,
		float	= 0.1 )

Fills the matrix with uniform random values in [min, max].

set_random_uniform() [2/2]

void opennn::set_random_uniform	(	VectorMap	,
		float	= -0.1,
		float	= 0.1 )

Fills the vector map with uniform random values in [min, max].

set_seed()

void opennn::set_seed

(

unsigned

seed

)

Seeds the library-wide pseudo-random number generator.

shape_to_string()

string opennn::shape_to_string ( const Shape & , const string & = " " )

nodiscard

Serializes a shape as a separator-joined string of dimensions.

shuffle()

void opennn::shuffle

(

VectorB &

vector_to_shuffle

)

Randomly permutes the entries of a boolean vector in place.

shuffle_vector()

template<typename T>

void opennn::shuffle_vector

(

vector< T > &

)

Randomly permutes the entries of the given vector in place.

shuffle_vector_blocks()

void opennn::shuffle_vector_blocks	(	vector< Index > &	,
		size_t	= 20 )

Shuffles contiguous blocks of the given length while preserving in-block order.

slice_rows() [1/2]

MatrixR opennn::slice_rows ( const MatrixR & matrix, const vector< Index > & indices )

inlinenodiscard

Returns a copy of the matrix containing only the rows at the given indices.

slice_rows() [2/2]

VectorR opennn::slice_rows ( const VectorR & values, const vector< Index > & indices )

inlinenodiscard

Returns a copy of the vector containing only the entries at the given indices.

softmax()

void opennn::softmax

(

TensorView &

output

)

Applies softmax in place along the trailing dimension of output.

softmax_cpu()

void opennn::softmax_cpu

(

TensorView &

output

)

CPU implementation of softmax().

sort_string_vector()

void opennn::sort_string_vector

(

vector< string > &

)

Sorts a vector of strings in place in ascending lexicographical order.

split_heads()

void opennn::split_heads	(	const TensorView &	source,
		TensorView &	destination )

Reshapes a multi-head attention tensor by splitting the last axis into heads.

split_heads_cpu()

void opennn::split_heads_cpu	(	const TensorView &	source,
		TensorView &	destination )

CPU implementation of split_heads().

standard_deviation() [1/2]

float opennn::standard_deviation ( const VectorR & )

nodiscard

Sample standard deviation of a vector.

standard_deviation() [2/2]

VectorR opennn::standard_deviation ( const VectorR & , Index  )

nodiscard

Rolling standard deviation of a vector over a window of the given size.

string_to_layer_type()

LayerType opennn::string_to_layer_type ( const string & name )

inlinenodiscard

Returns the LayerType corresponding to the given string name.

string_to_pooling_method()

PoolingMethod opennn::string_to_pooling_method ( const string & name )

inline

string_to_sample_role()

SampleRole opennn::string_to_sample_role ( const string & name )

inlinenodiscard

Parses a string (name or "0"/"1"/"2"/"3") into the matching SampleRole.

string_to_scaler_method()

ScalerMethod opennn::string_to_scaler_method ( const string & name )

inline

Parses a string into the matching ScalerMethod enumerator.

string_to_shape()

Shape opennn::string_to_shape ( const string & , const string & = " " )

nodiscard

Parses a separator-joined string of dimensions into a Shape.

string_to_tensor()

template<typename T, size_t Rank>

void opennn::string_to_tensor	(	const string &	input,
		TensorR< Rank > &	values )

Parses a whitespace-separated string into the flat storage of a tensor.

string_to_variable_role()

VariableRole opennn::string_to_variable_role ( const string & name )

inline

Parses a string into a VariableRole; "Id" is mapped to VariableRole::None.

string_to_variable_type()

VariableType opennn::string_to_variable_type ( const string & name )

inline

Parses a string into the matching VariableType enumerator.

string_to_vector()

void opennn::string_to_vector	(	const string &	input,
		VectorR &	values )

Parses a whitespace-separated string of floats into a VectorR.

tensor_to_string()

template<typename T, size_t Rank>

string opennn::tensor_to_string ( const TensorR< Rank > & values, const string & separator = " " )

nodiscard

Serializes a tensor's flat data to a string with the given separator.

throw_if()

void opennn::throw_if ( bool condition, const string & message, const source_location & loc = source_location::current() )

inline

to_cuda()

cudaDataType_t opennn::to_cuda ( Type type )

inlinenodiscardnoexcept

Returns the CUDA data type matching the given OpenNN Type (Auto resolves to FP32).

to_cudnn()

cudnnDataType_t opennn::to_cudnn ( Type type )

inlinenodiscardnoexcept

Returns the cuDNN data type matching the given OpenNN Type (Auto resolves to FP32).

to_int()

int opennn::to_int ( Index value )

inlinenodiscard

to_type()

float opennn::to_type ( Index value )

inlinenodiscard

tokenize()

vector< string > opennn::tokenize ( const string & )

nodiscard

Splits the input on whitespace into individual tokens.

tokenize_views()

vector< string_view > opennn::tokenize_views ( string_view )

nodiscard

Whitespace-tokenises a string view, returning views into the source buffer.

total_frequencies()

VectorI opennn::total_frequencies ( const vector< Histogram > & )

nodiscard

Sums the per-bin frequencies across a collection of histograms.

translate_image_x()

void opennn::translate_image_x	(	const Tensor3 &	,
		Tensor3 &	,
		Index	)

Translates the image along the X axis by the given number of pixels.

translate_image_y()

void opennn::translate_image_y	(	const Tensor3 &	,
		Tensor3 &	,
		Index	)

Translates the image along the Y axis by the given number of pixels.

trim_view()

string_view opennn::trim_view ( string_view )

nodiscard

Returns a view onto the input with leading and trailing whitespace removed.

type_bytes()

Index opennn::type_bytes ( Type type )

inlinenodiscardnoexcept

Returns the byte size of one element of the given OpenNN Type.

unscale()

void opennn::unscale	(	const TensorView &	input,
		const TensorView &	minimums,
		const TensorView &	maximums,
		const TensorView &	means,
		const TensorView &	standard_deviations,
		const TensorView &	scalers,
		float	min_range,
		float	max_range,
		TensorView &	output )

Inverse of scale(); reconstructs original values from a previously scaled tensor.

unscale_cpu()

void opennn::unscale_cpu	(	const TensorView &	input,
		const TensorView &	minimums,
		const TensorView &	maximums,
		const TensorView &	means,
		const TensorView &	standard_deviations,
		const TensorView &	scalers,
		float	min_range,
		float	max_range,
		TensorView &	output )

CPU implementation of unscale().

unscale_image_minimum_maximum()

void opennn::unscale_image_minimum_maximum	(	MatrixMap	,
		Index	)

Maps a column back from [-1, 1] to the [0, 255] image-pixel range.

unscale_logarithmic()

void opennn::unscale_logarithmic	(	MatrixMap	,
		Index	)

Inverse of scale_logarithmic() for the given column.

unscale_mean_standard_deviation()

void opennn::unscale_mean_standard_deviation	(	MatrixMap	,
		Index	,
		const Descriptives &	)

Inverse of scale_mean_standard_deviation() for the given column.

unscale_minimum_maximum()

void opennn::unscale_minimum_maximum	(	MatrixMap	,
		Index	,
		const Descriptives &	,
		float	= -1.0f,
		float	= 1.0f )

Inverse of scale_minimum_maximum(): reconstructs original values for the given column.

unscale_standard_deviation()

void opennn::unscale_standard_deviation	(	MatrixMap	,
		Index	,
		const Descriptives &	)

Inverse of scale_standard_deviation() for the given column.

variable_role_map()

const EnumMap< VariableRole > & opennn::variable_role_map ( )

inline

Returns the bidirectional string/enum map for VariableRole.

variable_role_to_string()

const string & opennn::variable_role_to_string ( VariableRole role )

inline

Returns the canonical string name for a VariableRole.

variable_type_map()

const EnumMap< VariableType > & opennn::variable_type_map ( )

inline

Returns the bidirectional string/enum map for VariableType.

variable_type_to_string()

const string & opennn::variable_type_to_string ( VariableType type )

inline

Returns the canonical string name for a VariableType.

variance() [1/2]

float opennn::variance ( const VectorR & )

nodiscard

Sample variance of a vector.

variance() [2/2]

float opennn::variance ( const VectorR & , const VectorI &  )

nodiscard

Sample variance of the selected entries of a vector.

vector_descriptives()

Descriptives opennn::vector_descriptives ( const VectorR & )

nodiscard

Returns the (min, max, mean, std) descriptives of a vector.

vector_map()

VectorMap opennn::vector_map ( const MatrixR & , Index  )

nodiscard

Returns an Eigen VectorMap that views a single column of a matrix without copying.

vector_to_string() [1/2]

template<typename Derived>

string opennn::vector_to_string ( const Eigen::DenseBase< Derived > & values, const string & separator = " " )

inlinenodiscard

Serializes an Eigen dense expression to a string with the given separator.

vector_to_string() [2/2]

template<typename T>

string opennn::vector_to_string ( const vector< T > & values, const string & separator = " " )

nodiscard

Serializes a vector to a string with the given element separator.

view_at_slot_or() [1/2]

TensorView & opennn::view_at_slot_or ( vector< TensorView > & views, const vector< size_t > & slots, size_t i, TensorView & fallback )

inline

view_at_slot_or() [2/2]

TensorView & opennn::view_at_slot_or ( vector< vector< TensorView > > & views, const vector< size_t > & slots, size_t i, TensorView & fallback )

inline

visit_type()

template<Type... Supported, typename F>

void opennn::visit_type	(	Type	t,
		F &&	f )

Dispatches f with the TypeInfo of the runtime Type t (must be in Supported).

Parameters

t	Runtime numeric type to resolve.
f	Callable invoked with the matching TypeInfo as its only argument.

visit_type_pair()

template<Type... Supported, typename F>

void opennn::visit_type_pair	(	Type	t_in,
		Type	t_out,
		F &&	f )

Dispatches f with the TypeInfo pair for an input and output runtime Type.

Parameters

t_in	Runtime numeric type for inputs.
t_out	Runtime numeric type for outputs.
f	Callable invoked with (in_info, out_info).

weighted_squared_error()

void opennn::weighted_squared_error	(	const TensorView &	input,
		const TensorView &	target,
		float	pos_w,
		float	neg_w,
		float &	error,
		float *	workspace_device )

Computes the binary squared error weighted asymmetrically for positive and negative classes.

Parameters

input	Predicted probabilities in [0,1].
target	Binary targets (0 or 1).
pos_w	Weight applied to positive samples.
neg_w	Weight applied to negative samples.
error	Output scalar holding the weighted squared error.
workspace_device	Optional device-side reduction workspace.

weighted_squared_error_gradient()

void opennn::weighted_squared_error_gradient	(	const TensorView &	input,
		const TensorView &	target,
		float	pos_w,
		float	neg_w,
		float	coefficient,
		const TensorView &	input_delta )

Writes the gradient of the weighted squared error scaled by coefficient into input_delta.

write_json()

void opennn::write_json	(	JsonWriter &	writer,
		initializer_list< pair< const char *, string > >	props )

z_correlation_to_r_correlation()

float opennn::z_correlation_to_r_correlation ( const float )

nodiscard

Inverse Fisher z-to-r transform.

Variable Documentation

ALIGN_BYTES

Index opennn::ALIGN_BYTES = EIGEN_MAX_ALIGN_BYTES

staticconstexpr

ALIGN_ELEMENTS

Index opennn::ALIGN_ELEMENTS = ALIGN_BYTES / sizeof(float)

staticconstexpr

CUBLAS_COMPUTE_DTYPE

cublasComputeType_t opennn::CUBLAS_COMPUTE_DTYPE = CUBLAS_COMPUTE_32F_FAST_TF32

constexpr

CUDA_REDUCTION_DTYPE

cudaDataType_t opennn::CUDA_REDUCTION_DTYPE = CUDA_R_32F

constexpr

EPSILON

float opennn::EPSILON = numeric_limits<float>::epsilon()

constexpr

GRADIENT_NORM_EPS

float opennn::GRADIENT_NORM_EPS = 1e-6f

inlineconstexpr

MAX

float opennn::MAX = numeric_limits<float>::max()

constexpr

NEG_INFINITY

float opennn::NEG_INFINITY = -numeric_limits<float>::infinity()

constexpr

QUIET_NAN

float opennn::QUIET_NAN = numeric_limits<float>::quiet_NaN()

constexpr

SOFTMAX_MASK_VALUE

float opennn::SOFTMAX_MASK_VALUE = float(-1e9f)

constexpr