Inits

Source: nn/Inits.luau

Parameter initializers. Each returns a flat row-major data array for a shape. Independent of the autograd engine; operates on raw Lua arrays.

Methods

zeros(shape: { number }) -> { number }

ones(shape: { number }) -> { number }

normal(shape: { number }, rng: any, mean: number?, std: number?) -> { number }

uniform(shape: { number }, rng: any, lo: number?, hi: number?) -> { number }

glorotUniform(shape: { number }, rng: any) -> { number }

Xavier/Glorot uniform: +-sqrt(6 / (fanIn + fanOut)).

kaimingUniform(shape: { number }, rng: any) -> { number }

Kaiming uniform for ReLU: +-sqrt(6 / fanIn).

smallUniform(shape: { number }, rng: any, scale: number?) -> { number }

Small-initialized weights (pre-norm transformer convention; scale for our depths so activations stay ~unit magnitude under residual streams).