eneural_net library

eNeural.net library.

Classes

ActivationFunction<N extends num, E>
Base class for Activation Functions.
ActivationFunctionFloat32x4
Base class for SIMD optimized functions using Float32x4.
ActivationFunctionLinear
Linear Activation Function (SIMD optimized).
ActivationFunctionSigmoid
Sigmoid Activation Function (SIMD optimized).
ActivationFunctionSigmoidBoundedFast
Fast Pseudo-Sigmoid Activation Function Bounded (SIMD optimized).
ActivationFunctionSigmoidFast
Fast Pseudo-Sigmoid Activation Function (SIMD optimized).
ActivationFunctionSigmoidFastInt
Experimental Integer Sigmoid Function.
ActivationFunctionSigmoidFastInt100
Experimental Integer Sigmoid Function (scale 100).
AdaDelta<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
AdaDelta optimizer (Zeiler, 2012). Needs no global learning rate.
AdaGrad<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
AdaGrad optimizer.
Adam<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Adam optimizer (Kingma & Ba, 2014), with optional AdamW decoupled weight decay, AMSGrad, and Nadam (Nesterov) variants.
ANN<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>>
Artificial Neural Network
Backpropagation<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Implementation of Backpropagation training algorithm.
Chronometer
A Chronometer useful for benchmarks.
ConjugateGradient<P extends VectorSample>
Nonlinear Conjugate Gradient (Fletcher–Reeves, with restarts + line search).
CosineAnnealingStrategy<N extends num, E, T extends Signal<N, E, T>>
Cosine annealing from base down to minValue over maxEpochs.
DataEntry
DataStatistics<N extends num>
DifferentialEvolution<P extends PopulationSample>
Differential Evolution (DE/rand/1/bin).
EvolutionStrategy<P extends PopulationSample>
(μ, λ) Evolution Strategy with a global step size (1/5-success adaptation).
ExponentialDecayStrategy<N extends num, E, T extends Signal<N, E, T>>
Exponential decay: base · gamma^epoch.
GeneticAlgorithm<P extends PopulationSample>
Genetic Algorithm: tournament selection, blend crossover, Gaussian mutation, elitism.
GradientOptimizer<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Base class for per-weight gradient optimizers (Adam, RMSProp, AdaGrad, AdaDelta, Quickprop, Lion, SGD/Momentum, ...).
HiddenLayerConfig<N extends num, E>
The configuration for the hidden layers.
Layer<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>>
Base class for ANN layers.
LayerFloat32x4
ANN Layer for Float32x4 types.
LayerHidden<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>>
Layer specialized for hidden neurons.
LayerInput<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>>
Layer specialized for input neurons.
LayerInt32x4
ANN Layer for Int32x4 types.
LayerOutput<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>>
Layer specialized for output neurons.
LBFGS<P extends VectorSample>
Limited-memory BFGS (two-loop recursion + line search).
LearningRateScheduleStrategy<N extends num, E, T extends Signal<N, E, T>>
Base for epoch-based learning-rate schedules. The value is recomputed each epoch from Training.trainedEpochs via computeValue.
LevenbergMarquardt<P extends VectorSample>
Levenberg–Marquardt (damped Gauss–Newton). Excellent for small networks.
Lion<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Lion optimizer (EvoLved Sign Momentum, Chen et al., 2023).
NativeBackpropagation<P extends NativeSample>
Backpropagation trainer accelerated by a native CPU/Metal backend.
NativeRProp<P extends NativeSample>
Resilient Backpropagation (iRProp+) trainer accelerated by a native CPU/Metal backend.
ParameterStrategy<N extends num, E, T extends Signal<N, E, T>>
Base class for training parameter strategy.
ParticleSwarm<P extends PopulationSample>
Particle Swarm Optimization.
PopulationTrainer<P extends PopulationSample>
Base for gradient-free trainers that optimize the flat weight vector (ANN.allWeights) using ANN.computeSamplesGlobalError as the fitness. Each learn call performs one generation/iteration and installs the best genome found into the ANN.
Quickprop<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Quickprop optimizer (Fahlman, 1988) — a second-order-ish method that fits a parabola through the current and previous gradient of each weight.
ResilientPropagation<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
RMSProp<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
RMSProp optimizer.
RProp<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Implementation of Resilient Backpropagation (version iRProp+).
Sample<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>>
Base class for ANN samples.
SampleFloat32x4
ANN sample based in Float32x4 data.
SampleInt32x4
SamplesGenerator
Samples Generator.
SamplesSet<P extends Sample<num, dynamic, dynamic, Scale<num>>>
Samples Set.
Scale<N extends num>
Base class for scales used for ANN, Signal and Sample.
ScaleDouble
A Scale<double>.
ScaleInt
A Scale<int>.
ScaleZoomable<N extends num>
ScaleZoomableDouble
ScaleZoomableInt
SeparableCMAES<P extends PopulationSample>
Separable (diagonal) CMA-ES: per-coordinate step sizes updated from the variance of the selected steps.
SGD<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Stochastic Gradient Descent with optional (classic or Nesterov) momentum.
Signal<N extends num, E, T extends Signal<N, E, T>>
SignalFloat32x4
SignalFloat32x4Mod4
SignalInt32x4
SimulatedAnnealing<P extends PopulationSample>
Simulated Annealing (single candidate, geometric cooling).
StaticParameterStrategy<N extends num, E, T extends Signal<N, E, T>>
A parameter strategy with a static/constant value.
StepDecayStrategy<N extends num, E, T extends Signal<N, E, T>>
Step decay: base · gamma^(epoch ~/ stepSize).
Training<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>>
Base class for training algorithms.
VectorTrainer<P extends VectorSample>
Base for trainers that operate on the flat weight vector (ANN.allWeights) and use finite-difference gradients/Jacobians — robust for the small-network regime these second-order methods target.
WarmupStrategy<N extends num, E, T extends Signal<N, E, T>>
Linear warmup to base over warmupEpochs, constant thereafter.
WebGpuBackpropagation<P extends WebGpuSample>
Backpropagation trainer accelerated by WebGPU (browser GPU).
WebGpuRProp<P extends WebGpuSample>
Resilient Backpropagation (iRProp+) trainer accelerated by WebGPU.

Enums

ActivationFunctionScope
Scope of the activation function.
NativeBackend
Native acceleration backend selection.
RPropVariant
Resilient Backpropagation variants (Riedmiller & Braun; Igel & Hüsken).

Mixins

WebGpuTrainerMixin<P extends WebGpuSample>
Shared WebGPU async-training logic for the WebGpuRProp/WebGpuBackpropagation trainers.

Functions

DefaultTrainingLogger(Training<num, dynamic, Signal<num, dynamic, dynamic>, Scale<num>, Sample<num, dynamic, Signal<num, dynamic, dynamic>, Scale<num>>> training, String type, String message, [dynamic error, StackTrace? stackTrace]) → dynamic
defaultTrainingLogger(Training<num, dynamic, Signal<num, dynamic, dynamic>, Scale<num>, Sample<num, dynamic, Signal<num, dynamic, dynamic>, Scale<num>>> training, String type, String message, [dynamic error, StackTrace? stackTrace]) → void
The default TrainingLogger.
registeredTrainings() → Iterable<String>
The registered algorithm names.
registerTraining(String name, TrainingBuilderFn builder) → void
Registers a trainer builder under name (case-insensitive).
restoreTrainingCheckpoint(TrainingD training, Map<String, dynamic> checkpoint) → void
Restores a checkpoint produced by saveTrainingCheckpoint into training (which must have the same topology/algorithm/hyperparameters).
saveTrainingCheckpoint(TrainingD training) → Map<String, dynamic>
Serializes a resumable checkpoint of training: the ANN weights plus, for gradient optimizers, the optimizer state buffers and step counter.
trainingByName(String name, AnnD ann, SamplesD samples, {Map<String, dynamic> params = const {}}) → TrainingD
Builds a trainer by name with optional params. Throws if unknown.

Typedefs

AnnD = ANN<double, Float32x4, SignalFloat32x4, Scale<double>>
ANN type used by the registry.
LearningRateScheduleBuilder<N extends num, E, T extends Signal<N, E, T>> = ParameterStrategy<N, E, T> Function(Propagation<N, E, T, dynamic, dynamic> propagation, double baseLearningRate)
Builds a learning-rate ParameterStrategy for an optimizer (used to plug in LR schedules). Receives the optimizer (a Propagation) and its base LR.
SamplesD = SamplesSet<SampleFloat32x4>
TrainingBuilder<N extends num, E, T extends Signal<N, E, T>, S extends Scale<N>, P extends Sample<N, E, T, S>> = Training<N, E, T, S, P> Function(ANN<N, E, T, S> ann, SamplesSet<P> samplesSet)
A builder for Training instances.
TrainingBuilderFn = TrainingD Function(AnnD ann, SamplesD samples, Map<String, dynamic> params)
Builds a trainer from a parameter map.
TrainingD = Training<double, Float32x4, SignalFloat32x4, Scale<double>, SampleFloat32x4>
Concrete trainer type used by the name-based registry (Float32x4 networks).
TrainingLogger = void Function(Training<num, dynamic, Signal<num, dynamic, dynamic>, Scale<num>, Sample<num, dynamic, Signal<num, dynamic, dynamic>, Scale<num>>> training, String type, String message, [dynamic error, StackTrace? stackTrace])
Training logger.