Policy

The policy defines the task a run is solving and the obligations that task places on each component. It fixes how model outputs are interpreted, how many output channels the model must produce, whether targets are normalized, and the structural checks applied to component inputs and outputs. It is the single place a run declares “this is a classification problem” or “this is a regression problem”.

Usage

The policy occupies the top-level policy slot.

policy:
  name: multiclass
  kwargs: {}

How it works

From the dataset, a policy builds a task specification: the output layout the model must produce, the data type of the targets, and whether targets are normalized. It then issues a contract to each component as that component is attached. A contract carries task-derived parameters together with validators that check a component conforms to the task. This is how one task definition keeps the model, normalizer, transformer, optimizer, and loss mutually consistent.

Variants

• Multiclass: classification over discrete labels.

• Regression: prediction of continuous targets.

Registering a new policy

A policy is a subclass of Policy that declares a name and version and builds the task specification:

_build_task_spec(self) -> TaskSpec

Return the output offsets, target data type, and target-normalization flag for the task.

The base supplies default contracts for each component kind; override the per-component contract methods only when a task needs stricter validation or extra parameters.

from typing import Any, ClassVar

from icegraph.engine.policy import Policy, PolicyFactory, TaskSpec

from .config import MyPolicyConfig

class MyPolicy(Policy[MyPolicyConfig]):
    name: ClassVar[str] = "my-policy"
    version: ClassVar[int] = 1

    @classmethod
    def validate_config(cls, config: dict[str, Any]) -> MyPolicyConfig:
        return MyPolicyConfig(**config)

    def _build_task_spec(self) -> TaskSpec:
        ...  # return a TaskSpec for the task

PolicyFactory.register(MyPolicy)