Initializers (Machine Gnostics MAGNET)¶
The machinegnostics.magnet module provides weight initialization strategies for MAGNET layers. Proper initialization is critical for gradient stability in networks using gnostic activations.
Key Initializer Classes¶
- GlorotUniform / GlorotNormal — Xavier/Glorot initialization (
limit = sqrt(6 / (fan_in + fan_out))) - HeUniform / HeNormal — He initialization for ReLU-family activations (
√(2/fan_in)or√(2/fan_in)stddev) - Normal — Random normal distribution (
RandomNormalalias) - Uniform — Uniform random distribution
- Ones — Constant ones (useful for biases)
- Zeros — Constant zeros (standard bias init)
Base Class¶
- Initializer — Abstract base (
Initializer(ABC)) with__init__(dtype=np.float64)and abstract__call__(shape) -> np.ndarray. Includesnormalize_shape()andcompute_fans()helpers.
Helper Function¶
get_initializer(name_or_callable=None)— Resolve name ('glorot_uniform','he_normal','normal', etc.) or callable; defaults toGlorotUniform().
Compatibility Aliases¶
XavierUniform = GlorotUniformXavierNormal = GlorotNormalRandomNormal = Normal
Author: Nirmal Parmar