Decode

The decode service turns raw dataset records into the tensors a run consumes. It is the bridge between the record format and the model’s input and output spaces.

Usage

Configured under services.decode. The decoders for attributes and records are themselves selectable plugins.

services:
  decode:
    features: []
    targets:  [ sim ]
    auxiliary: []
    keymap:
      truth: targets
    attrs:
      name: standard
      kwargs: {}
    records:
      name: standard
      kwargs: {}

How it works

Two decoders divide the work. The attribute decoder reads dataset-level metadata, such as column names, per-column statistics, and the set of observed label values. The record decoder reads a columnar block of records and decodes its columns (such as features, targets, etc) into flat tensors.

Configuration

Option	Description	Type	Default
`features`	Column names used as input features.	list[str]	`[]`
`targets`	Column names used as training targets.	list[str]	`[]`
`auxiliary`	Extra columns to decode but exclude from the loss (useful for plotting and analysis).	list[str]	`[]`
`keymap`	Overrides mapping logical roles to the keys used in the stored data.	mapping	identity
`attrs`	Attribute-decoder plugin selection (`name` / `kwargs`).	mapping	required
`records`	Record-decoder plugin selection (`name` / `kwargs`).	mapping	required

The keymap accepts the keys truth (the source of both targets and auxiliary), features, and weights. truth and features default to their own names, while weights defaults to generation, since the weights role reads whatever its decoder needs, and for simulation by default that is the per-event generation data.

Sub-slots