mapi_steps¶
This package has the built-in core.* steps. For the settings of each step, read the
generated steps catalog.
Built-in preprocessing steps for modeling-api (core.*).
Every step implements the :class:mapi_sdk.steps.Step contract and exports static JSON
(settings schema, templated :class:~mapi_spec.effects.SchemaEffect, context need) so the
control plane can validate specs without importing this package. Steps are discovered
through the mapi.steps entry point (:func:builtin_steps); :func:registry returns a
ready :class:~mapi_sdk.registry.Registry, and python -m mapi_steps.catalog prints the
catalog JSON.
========== ===========================================================================
Reshape resample, count_events, filter_coords, to_long, to_padded, stack, unstack
Value fx_convert, mean_scale, max_abs_scale, standardize, log1p, logit, impute,
replace_sentinel, clip
Feature fourier, trend, calendar, holidays, lag, lag_tensor, one_hot, encode_levels,
missing_indicator, formula
Precompute adstock, saturation (point models)
Split time_holdout, as_of, leave_group_out, last_per_group, mask_block
(:mod:mapi_steps.splits; chosen in data.split, applied by the harness)
========== ===========================================================================
Catalog export¶
Static catalog export: python -m mapi_steps.catalog [--splits] [--out FILE].
Prints the :class:mapi_spec.catalog.Catalog JSON of every built-in step (settings JSON
Schema, templated SchemaEffect, context need, kind, flags). Runner images generate it at
build time; the control plane loads it to validate specs without importing step code.
--splits adds the split catalog under a separate splits key (not part of
Catalog). The same JSON ships with the package as mapi_steps/catalog.json.
catalog_json
¶
Deterministic JSON text of the catalog (optionally with the split catalog).
Splits¶
The seven built-in train/holdout splits (plan §4.1 "Split" group; v0.14 adds column
and stratified).
Splits are not pipeline steps: a spec chooses one in data.split and the harness applies
it at the split stage (:func:mapi_sdk.split.split_dataset), after the stateless prefix.
This module gives each split kind a catalog identity (core.<kind>@1.0.0) with its
settings schema and plain-language summaries for pickers and docs, plus a direct
:meth:SplitStrategy.split for notebooks and tests.
======================== ====================================================================
core.time_holdout hold out the last last periods / a fraction / after a label
core.as_of train up to and including as_of; optional gap periods
core.leave_group_out hold out whole groups (explicit groups or n random ones)
core.last_per_group last last rows of each group with at least min_len rows
core.mask_block random blocks over dims covering fraction of the cells
core.column the rows whose column value is in holdout (a fixed split)
core.stratified a random fraction of the rows of each level of by
======================== ====================================================================