Migration from v1 to v2¶

v2 preserves the project goal but changes internals to a composable pipeline.

Architecture changes¶

v1	v2	Migration
`computeHiddenOutput` duplicated across models	`FeatureMap::transform`	Use `RandomAdditiveMap`, `RbfMap`, `ElmAutoEncoderLayer`, or `StackedFeatureMap`
Some models accepted `Backend`	Every model accepts `Backend`	Pass `Backend::kCpu` or `Backend::kGpu` explicitly
RBF activation enum	`RbfMap` feature map	Replace `ActivationFunction::kRbf` usage with `RbfMap`
H-OS-ELM fixed random stack	ELM-AE learned stack	Use `HierarchicalOsElm` with `hiddenNodesPerLayer`
Hard-coded `1e-8` ridge term	`ridgeAlpha`	Pass a positive `ridgeAlpha`
Demo fabricated training data	Bundled dataset and real metrics	Use dataset IO and demo endpoints

API notes¶

BatchElm keeps its constructor shape and adds ridgeAlpha.
OsElm and OsCelm accept RlsOptions.
OsCelm gains Backend for consistency.
HierarchicalOsElm now learns ELM-AE layers during initialization.

Code migration example¶

Before:

feature_elm::BatchElm<float> model(numInputs, hiddenNodes, feature_elm::ActivationFunction::kSigmoid);

After, with explicit backend and ridge alpha:

feature_elm::BatchElm<float> model(
    numInputs,
    hiddenNodes,
    feature_elm::ActivationFunction::kSigmoid,
    feature_elm::Backend::kCpu,
    1e-6f);

Documentation migration¶

See upgrade baseline and H-OS-ELM migration for regression notes.