---

name: custom-components description: "Implement custom PyTorch Forecasting metrics, models, package wrappers, data modules, and focused maintainer tests." disable-model-invocation: true metadata: disco-role: operating license: MIT

Custom Components

Use this sub-skill when you are adding or reviewing PyTorch Forecasting custom components for version 1.8.0: metrics/losses, v1 BaseModel implementations, v1 _pkg package containers, experimental v2 model/package/data-module components, and focused maintainer tests.

Do not use this sub-skill for ordinary package usage, release/changelog work, broad CI administration, documentation-site builds, or hyperparameter-tuning workflows.

Route by task

• For a custom metric or loss, start from the bundled template, then read the metric section in references/custom-components.md; it explains the MultiHorizonMetric contract, optional distribution-loss hooks, and package-wrapper tags when a metric must be discoverable.
• For a v1 custom model, read references/custom-components.md; it gives the BaseModel / BaseModelWithCovariates / _pkg pattern, TimeSeriesDataSet.from_dataset rules, tensor-dict shape expectations, and registry tags.
• For experimental v2 work, read references/custom-components.md; it summarizes the v2 BaseModel, Base_pkg, TimeSeries, and EncoderDecoderTimeSeriesDataModule concepts without relying on source templates.
• Before running checks, use references/maintainer-testing.md; it lists focused pytest selections, CPU smoke tests, ruff/mypy/pre-commit basics, and when to avoid broad optional extras.
• When a component fails import, .from_dataset, tensor-shape, package-linkage, optional-dependency, or estimator tests, use references/troubleshooting.md; it maps common failure signals to fixes.
• Copy scripts/custom_metric_template.py when implementing a new point-style MultiHorizonMetric; it is a syntactically valid, minimal class with a CPU self-check.

Operating stance

Prefer small CPU fixtures and package-relative commands. Keep optional dependencies behind lazy imports and explicit python_dependencies tags. Treat v2 APIs as beta/unstable; prefer stable v1 TimeSeriesDataSet + BaseModel.from_dataset() unless the task explicitly targets v2 package/data-module layers.