Dsxir.Optimizer.Search.Sampler behaviour (dsxir v0.3.0)

Copy Markdown

Behaviour for categorical hyperparameter samplers used by Dsxir.Optimizer.MIPROv2 and any other optimizer that searches a fixed categorical space.

The space is a map from dimension key to a {:categorical, [choices]} tuple. Configs are maps from dimension key to a chosen value (typically an integer index into the choices list). Observations pair a config with its scalar score.

suggest/3 returns n candidate configs and the (possibly mutated) sampler state. observe/2 takes a list of observations — batched, one call per suggest batch — and folds them into state. Random's observe/2 is a no-op; TPE's updates its good/bad partition.

Summary

Types

config()
observation()
space()

Callbacks

init(space, opts)
observe(state, list)
suggest(state, history, n)

Types

config()

@type config() :: %{required(term()) => term()}

observation()

@type observation() :: %{config: config(), score: float()}

space()

@type space() :: %{required(term()) => {:categorical, [term(), ...]}}

Callbacks

init(space, opts)

@callback init(space(), opts :: keyword()) :: state :: term()

observe(state, list)

@callback observe(state :: term(), [observation()]) :: state :: term()

suggest(state, history, n)

@callback suggest(state :: term(), history :: [observation()], n :: pos_integer()) ::
  {[config()], state :: term()}