Gradien

Appearance

Debugging & Metrics Tools ​

Metrics Parallel ​

Located in Gradien.Metrics.

  • accuracy(logits, targets): Returns classification accuracy (0.0 - 1.0).
  • topk(logits, targets, k): Returns top-k accuracy.
  • mse(pred, target): Mean Squared Error.
  • prf1(pred, target, C): Returns Precision, Recall, and F1 Score.
  • confusion(pred, target, C): Returns a Confusion Matrix { {number} }.

Debug Tools ​

Located in Gradien.Debug and Gradien.Util.Anomaly.

Anomaly ​

Low-level checks for numerical stability.

  • hasNaN(t): Returns true if tensor contains NaN.
  • hasInf(t): Returns true if tensor contains Inf.
  • hasBadValues(t): Returns true if NaN or Inf.

GradStats ​

Analyzes gradients across an entire model. Useful for diagnosing vanishing/exploding gradients.

lua
GradStats.forModel(model: Module) -> { [Tensor]: StatTable }
lua
{
    min: number,
    max: number,
    mean: number,
    std: number,
    n: number
}

Debug.wrapOptimizer ​

Creates a wrapper around an optimizer that automatically performs gradient clipping and NaN checks before every step.

lua
(
    opt: Optimizer, 
    params: {Tensor}, 
    cfg: { 
        maxGradNorm: number?, 
        clipValue: number?, 
        warnOnNaN: boolean?,
        label: string? 
    }
) -> WrappedOptimizer