Module weight_init 

Weight initialization strategies for neural network parameters.

Provides common initialization methods including Xavier/Glorot, Kaiming/He, LeCun, orthogonal, and basic constant/normal/uniform initializations. Uses a deterministic LCG-based RNG (no rand crate dependency).

Structs

InitRng

A deterministic linear congruential generator (LCG) for reproducible weight initialization without depending on the rand crate.

InitStats

Statistics about an initialized weight tensor.

Enums

FanMode

Selects whether to use fan_in or fan_out for Kaiming initialization.

InitError

Errors that can occur during weight initialization.

Functions

compute_fans

Compute (fan_in, fan_out) from a weight tensor shape.

constant_init

Initialize all elements to a constant value.

gain_for_activation

Return the recommended gain for a given activation function name.

kaiming_normal

Kaiming (He) normal initialization.

kaiming_uniform

Kaiming (He) uniform initialization.

lecun_normal

LeCun normal initialization: N(0, 1/sqrt(fan_in)).

lecun_uniform

LeCun uniform initialization: U(-limit, limit) where limit = sqrt(3/fan_in).

normal_init

Normal initialization with specified mean and standard deviation.

ones_init

Initialize all elements to one.

orthogonal_init

Orthogonal initialization via QR-like Gram-Schmidt on a random matrix.

uniform_init

Uniform initialization with specified bounds [low, high).

xavier_normal

Xavier (Glorot) normal initialization.

xavier_uniform

Xavier (Glorot) uniform initialization.

zeros_init

Initialize all elements to zero.