RecipeProvenance

Enums · AiDotNet.Attributes

How a declared hyperparameter was arrived at in the paper that states it.

How It Works

Not every number in a paper is a value the authors chose. Some are searched over, some vary by dataset, and some are inherited from a cited work rather than stated. Treating all three as though the paper had simply said “use this” is how a reproduction quietly acquires a number nobody ever recommended.

Before this existed, the only way to be honest about a searched or per-dataset value was to omit it and explain in prose — which is why five of the first hundred declarations state an optimizer and no rate at all. iTransformer searches {1e-3, 5e-4, 1e-4}; BERT searches {5e-5, 3e-5, 2e-5}; TimesNet sets a rate per dataset in a table; DeepAR tunes it manually per dataset. All four are now recordable rather than lost.

Fields

Field Summary
RecipeProvenance DerivedFromCitedWork The paper does not state this value; it comes from a work the paper says it follows.
RecipeProvenance NamedReproduction The paper does not state this value and its authors released no code; it comes from a named public reproduction.
RecipeProvenance PerDataset The paper gives a different value per dataset, and the declared one is an example.
RecipeProvenance ReleasedCode The paper does not state this value; the authors’ released code and training configuration do.
RecipeProvenance Searched The paper searched over several values; the declared one is representative rather than prescribed.
RecipeProvenance Stated The paper states this value directly as the one it used.
int value__