---

name: metrics description: "Routes Ignite metric attachment, arithmetic, family-specific evaluation, and compute/reset workflows." disable-model-invocation: true metadata: disco-role: operating license: BSD 3-Clause

Ignite metrics workflows

Use this sub-skill when the task is about computing, validating, combining, or troubleshooting Ignite metrics rather than the training loop itself.

Include here

• Metric attachment to an evaluator or custom engine, including output_transform, skip_unrolling, and distributed metric reduction behavior.
• Direct reset / update / compute usage for custom or ad hoc evaluation loops.
• Metric arithmetic with MetricsLambda, operator chaining, indexing, and derived metrics such as F-beta.
• Core classification metrics: Accuracy, Precision, Recall, Fbeta, ConfusionMatrix, TopKCategoricalAccuracy, ClassificationReport, AveragePrecision, PrecisionRecallCurve, ROC_AUC, and RocCurve.
• Regression, clustering, and distance metrics that live under ignite.metrics.regression and ignite.metrics.clustering.
• NLP, vision, GAN, fairness, and recommender-system metrics, including Bleu, Rouge, SSIM, FID, InceptionScore, GpuInfo, SubgroupAccuracyDifference, DemographicParityDifference, HitRate, and NDCG.
• Metric grouping and reuse patterns with MetricGroup.

Exclude or route elsewhere

• Engine creation, event wiring, resume logic, and deterministic loop control belong in sub-skills/engine/.
• Checkpointing, schedulers, progress bars, profilers, and logger integrations belong in sub-skills/handlers/.
• Distributed launchers and backend selection belong in sub-skills/distributed/.
• Legacy ignite.contrib compatibility notes live in references/legacy-contrib.md.

Start here

• Read references/api-reference.md for the metric families, dependencies, and shape expectations.
• Read references/workflows.md for end-to-end evaluator recipes, derived-metric patterns, and family-specific examples.
• Read references/troubleshooting.md when a metric raises NotComputableError, shape errors, missing optional dependency errors, or distributed reduction surprises.
• Run scripts/metric_smoke.py for a tiny synthetic check that exercises classification, binary ranking, fairness, image, and recommender-style metrics.

Common triggers

• "How do I attach Accuracy, F-beta, ROC AUC, or SSIM to Ignite?"
• "Why does NotComputableError happen for this metric?"
• "How do I build a derived metric from Precision and Recall?"
• "Which optional dependency do I need for FID, ROC AUC, or GPU info?"
• "How do I evaluate subgroup fairness or recommender-system metrics?"

Useful boundary notes

This route owns metric math and metric-specific validation, but it does not own the trainer/evaluator structure itself. If the metric is part of a larger loop or a checkpoint/logging problem, keep the loop in sub-skills/engine/ and the side effects in sub-skills/handlers/.