Tinkex.Types.RegularizerOutput (Tinkex v0.3.4)

View Source

Output metrics from a single regularizer computation.

This struct captures both the loss contribution and optional gradient tracking information for monitoring regularizer dynamics.

Fields

  • :name - Regularizer name (matches RegularizerSpec.name)
  • :value - Raw loss value before weighting
  • :weight - Weight applied to the loss
  • :contribution - Weighted contribution: weight * value
  • :grad_norm - L2 norm of gradients (when tracking enabled)
  • :grad_norm_weighted - Weighted gradient norm: weight * grad_norm
  • :custom - Custom metrics returned by the regularizer function

Examples

%RegularizerOutput{
  name: "l1_sparsity",
  value: 22.4,
  weight: 0.01,
  contribution: 0.224,
  grad_norm: 7.48,
  grad_norm_weighted: 0.0748,
  custom: %{"l1_total" => 44.8, "l1_mean" => 22.4}
}

Summary

Types

t()

Functions

from_computation(name, loss_value, weight, custom_metrics, grad_norm \\ nil)

Create a RegularizerOutput from computation results.

Types

t()

@type t() :: %Tinkex.Types.RegularizerOutput{
  contribution: float(),
  custom: %{required(String.t()) => number()},
  grad_norm: float() | nil,
  grad_norm_weighted: float() | nil,
  name: String.t(),
  value: float(),
  weight: float()
}

Functions

from_computation(name, loss_value, weight, custom_metrics, grad_norm \\ nil)

@spec from_computation(
  name :: String.t(),
  loss_value :: float(),
  weight :: float(),
  custom_metrics :: map() | nil,
  grad_norm :: float() | nil
) :: t()

Create a RegularizerOutput from computation results.

Parameters

  • name - Regularizer identifier
  • loss_value - Raw loss value (before weighting)
  • weight - Weight multiplier
  • custom_metrics - Map of custom metrics (or nil)
  • grad_norm - L2 norm of gradients (optional)

Examples

RegularizerOutput.from_computation("l1", 22.4, 0.01, %{"l1_mean" => 22.4}, 7.48)