Sequential

classin packageklyn.math.nn

public class Sequential<T extends AbstractFloat = Float>

Sequential

Ordered composition of typed layers for classification and regression. Classification consumes Int labels; regression consumes targets in T. Input declarations build the network immediately; otherwise dimensions are inferred on the first fit, evaluate or predict call. Layers are not cloned: keep distinct instances per stage and per independent model. A model is not safe for concurrent training/prediction or parameter mutation. CPU training keeps batches ordered. Sufficiently large matrix products and independent optimizer ranges use shared native workers; small operations remain serial to avoid scheduling overhead. Reductions keep their order. Float16 keeps parameter/activation buffers in FP16, with FP32 matrix accumulation and Adam state. Training scales gradients dynamically and skips non-finite updates before reducing the scale. Optimizer state is not included in the parameter memory reported by summary().

Example

model = Sequential<Float>([Dense<Float>(128), Dense<Float>(10)])

Properties

Modifier and Type	Member	Description
public readonly property	device `device as Device`	Concrete device most recently resolved, or auto before resolution.

Constructors

Modifier and Type	Member	Description
public	Sequential `Sequential(layers as IList<Layer<T>>)`	Snapshots the order of a non-empty list of layers without copying parameters.

Methods

Modifier and Type	Member	Description
public	compile `compile(optimizer as Optimizer<T>, loss as ClassificationLoss<T>, metrics as IList<String> = ["accuracy"]) as Void`	Configures the optimizer and mean classification loss without resetting weights.
public	compile `compile(optimizer as Optimizer<T>, loss as RegressionLoss<T>, metrics as IList<String> = []) as Void`	Configures a typed floating-target regression objective.
public	compile `compile(optimizer as String = "adam", loss as String = "sparse_categorical_crossentropy", metrics as IList<String> = null) as Void`	Resolves standard optimizer/loss names to actual typed implementations.
public	evaluate `evaluate(notnull features as NDArray<T>, notnull labels as NDArray<Int>, batchSize as Long = 64l) as Evaluation`	Evaluates fixed parameters, without shuffling or optimizer updates.
public	evaluate `evaluate(notnull features as NDArray<T>, notnull targets as NDArray<T>, batchSize as Long = 64l) as Evaluation`	Computes sample-weighted regression loss with fixed parameters and no optimizer updates.
public	fit `fit(notnull features as NDArray<T>, notnull labels as NDArray<Int>, epochs as Long = 5l, batchSize as Long = 64l, shuffle as Boolean = true, seed as Long = 12345l, verbose as Int = 1, validationData as Tuple<NDArray<T>, NDArray<Int>> = null) as TrainingHistory`	Trains with shuffled minibatches and sample-weighted metrics.
public	fit `fit(notnull features as NDArray<T>, notnull targets as NDArray<T>, epochs as Long = 5l, batchSize as Long = 64l, shuffle as Boolean = true, seed as Long = 12345l, verbose as Int = 1, validationData as Tuple<NDArray<T>, NDArray<T>> = null) as TrainingHistory`	Trains on same-precision floating targets using the configured regression loss.
public static	load `load(filename as String, maxBytes as ULong = 2147483648ul) as Sequential<T>`	Restores using a filename; identical to the Path overload.
public static	load `load(path as Path, maxBytes as ULong = 2147483648ul) as Sequential<T>`	Restores an independent model, including all saved training state.
public	loadWeights `loadWeights(filename as String, maxBytes as ULong = 2147483648ul) as Void`	Loads weights using a filename.
public	loadWeights `loadWeights(path as Path, maxBytes as ULong = 2147483648ul) as Void`	Loads parameters transactionally into an identically configured built model.
public	predict `predict(notnull features as NDArray<T>) as NDArray<T>`	Computes model outputs without applying any implicit output activation.
public	save `save(filename as String, includeOptimizer as Boolean = true) as Void`	Saves using a filename; identical to the Path overload.
public	save `save(path as Path, includeOptimizer as Boolean = true) as Void`	Saves architecture, exact parameter bits and optional training state to KNN 1.0.
public	summary `summary() as Void`	Prints layer names, output shapes, trainable scalar counts and parameter bytes.
public	to `to(device as Device) as Sequential<T>`	Selects execution without resetting parameters, optimizer state or precision.