Module Kaun.MetricsSource

Layout-independent metric accumulator that produces host float values when computed.

accuracy ?threshold ?top_k () creates an accuracy metric.

• parameter threshold

  Threshold for binary classification (default: 0.5)

• parameter top_k

  For multi-class, count as correct if true label is in top-k predictions @note For per-class or aggregated variants, combine Metrics.confusion_matrix with custom post-processing.

Example

  let acc = Metrics.accuracy () in
  let top5_acc = Metrics.accuracy ~top_k:5 ()

precision ?threshold ?zero_division () creates a precision metric.

Precision = True Positives / (True Positives + False Positives)

• parameter threshold

  Binary classification threshold (default: 0.5)

• parameter zero_division

  Value to return when there are no positive predictions (default: 0.0)

Example

  let prec = Metrics.precision ()

recall ?threshold ?zero_division () creates a recall metric.

Recall = True Positives / (True Positives + False Negatives)

• parameter threshold

  Binary classification threshold (default: 0.5)

• parameter zero_division

  Value to return when there are no actual positives (default: 0.0)

f1_score ?threshold ?beta () creates an F-score metric.

F-score = (1 + β²) * (Precision * Recall) / (β² * Precision + Recall)

• parameter threshold

  Binary classification threshold (default: 0.5)

• parameter beta

  Weight of recall vs precision (default: 1.0 for F1)

auc_roc () creates an AUC-ROC (Area Under the Receiver Operating Characteristic) metric that integrates true/false positive rates observed across batches.

auc_pr () creates an AUC-PR (Area Under the Precision–Recall) metric. Computes the exact precision–recall integral by sorting predictions and accumulating precision/recall scores across all seen batches.

confusion_matrix ~num_classes ?normalize () accumulates a confusion matrix for classification tasks.

• parameter num_classes

  Number of classes

• parameter normalize

  Normalisation mode (default: `None)

mse ?reduction () creates a Mean Squared Error metric.

MSE = mean((predictions - targets)²)

rmse ?reduction () creates a Root Mean Squared Error metric.

RMSE = sqrt(mean((predictions - targets)²))

mae ?reduction () creates a Mean Absolute Error metric.

MAE = mean(|predictions - targets|)

loss () tracks the running mean of loss values. Pass batch losses through update ~loss to accumulate them.

mape ?eps () creates a Mean Absolute Percentage Error metric.

MAPE = mean(|predictions - targets| / (|targets| + eps)) * 100

• parameter eps

  Small value to avoid division by zero (default: 1e-7)

r2_score ?adjusted ?num_features () creates an R² coefficient of determination metric.

R² = 1 - (SS_res / SS_tot)

• parameter adjusted

  If true, compute adjusted R² (requires num_features)

• parameter num_features

  Number of features (needed for adjusted R²)

explained_variance () creates an explained variance metric.

EV = 1 - Var(targets - predictions) / Var(targets)

cross_entropy ?from_logits () creates a cross-entropy metric.

• parameter from_logits

  If true, apply softmax to predictions (default: true)

binary_cross_entropy ?from_logits () creates a binary cross-entropy metric.

• parameter from_logits

  If true, apply sigmoid to predictions (default: true)

kl_divergence ?eps () creates a Kullback–Leibler divergence metric.

KL(P||Q) = Σ P log(P / Q)

• parameter eps

  Small value for numerical stability (default: 1e-7)

perplexity ?base () creates a perplexity metric for language models.

Perplexity = base^(cross_entropy)

• parameter base

  Base for exponentiation (default: e)

ndcg ?k () creates a Normalised Discounted Cumulative Gain metric.

• parameter k

  Consider only the top-k ranked items (default: all)

map ?k () creates a Mean Average Precision metric for ranking.

mrr ?k () creates a Mean Reciprocal Rank metric.

MRR = mean(1 / rank_of_first_relevant_item)

• parameter k

  Consider only top-k items when computing the reciprocal rank (default: all)

bleu ?max_n ?weights ?smoothing () creates a BLEU score metric for pre-tokenized integer sequences.

• parameter max_n

  Maximum n-gram order (default: 4)

• parameter weights

  Weights for each n-gram order (default: uniform)

• parameter smoothing

  Apply smoothing for zero counts (default: true)

Predictions and targets must be shaped batch, seq_len with integer token identifiers. Zero values are treated as padding and ignored.

rouge ~variant ?use_stemmer () creates a ROUGE score metric for pre-tokenized integer sequences.

• parameter variant

  Which ROUGE variant to compute

• parameter use_stemmer

  Enable stemming (currently unsupported; raises when set)

Predictions and targets must be shaped batch, seq_len with integer token identifiers. Zero values are treated as padding and ignored.

meteor ?alpha ?beta ?gamma () creates a METEOR score metric for pre-tokenized integer sequences.

• parameter alpha

  Parameter controlling precision vs recall balance (default: 0.9)

• parameter beta

  Exponent for the chunk penalty (default: 3.0)

• parameter gamma

  Weight of the chunk penalty (default: 0.5)

Predictions and targets must be shaped batch, seq_len with integer token identifiers. Zero values are treated as padding and ignored.

psnr ?max_val () creates a Peak Signal-to-Noise Ratio metric.

PSNR = 10 * log10(max_val² / MSE)

• parameter max_val

  Maximum possible pixel value (default: 1.0)

ssim ?window_size ?k1 ?k2 () creates a Structural Similarity Index metric.

The implementation evaluates the global SSIM across the full prediction and target tensors using scalar statistics derived from window_size, k1, and k2.

iou ?threshold ?per_class ~num_classes () creates an Intersection over Union metric.

Inputs must contain integer class indices in 0, num_classes). When [num_classes = 2], [threshold] binarises predictions before computing IoU. When [per_class = true], the metric reports one IoU per class; otherwise it returns the mean over classes with non-zero support.

dice ?threshold ?per_class ~num_classes () creates a Sørenson Dice coefficient metric with the same input conventions as iou.

update metric ~predictions ~targets ?loss ?weights () updates the metric state. All tensors must share the same (hidden) layout. When supplied, the loss tensor is treated as an auxiliary scalar for metrics that track losses.

compute metric returns the aggregated metric value as a host float.

compute_tensor metric returns the aggregated metric value as a device tensor.

reset metric clears internal accumulators for a fresh run.

clone metric creates a new metric with the same configuration but fresh state.

name metric returns the metric's descriptive name.

create_custom ~dtype ~name ~init ~update ~compute ~reset constructs a custom metric from user-provided accumulator functions.

is_better metric ~higher_better ~old_val ~new_val determines whether the new metric value improves upon the previous one.

format metric value pretty-prints a metric value for logging.