Module

skainet-lang-core/sk.ainet.lang.nn/Module

Constructors

constructor()

Properties

open override val children: List<ModuleNode>

Child nodes in the hierarchy.

open override val id: String

A stable identifier for the node. Not required to be globally unique.

abstract val modules: List<Module<T, V>>

abstract override val name: String

Human-readable name of the node, typically the module's name.

open override val params: List<ModuleParameter<*, *>>

Parameters owned by this node (weights, biases, etc.).

open override var path: String?

Optional path annotation for nicer logs, e.g., "model.encoder.layer1".

Functions

fun <T : DType, V> Module<T, V>.describe(input: Shape, ctx: ExecutionContext, dtypeClass: KClass<T>, batch_size: Int = -1): String

open fun forward(input: Tensor<T, V>, ctx: ExecutionContext): Tensor<T, V>

operator fun <T : DType, V> Module<T, V>.invoke(input: Tensor<T, V>, ctx: ExecutionContext): Tensor<T, V>

Convenience operator to call a Module with an explicit ExecutionContext. This does not re-introduce the legacy context-less invoke; callers must pass ctx.

trainableParameters

open fun trainableParameters(): List<Parameter>

Collect all trainable parameters for this module and its subtree. Trainable is defined by ModuleParameter.requiresGrad flag.

Zero-out accumulated gradients for all trainable parameters in this module subtree.