Skip to content

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 (RandomNormal alias)
  • 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. Includes normalize_shape() and compute_fans() helpers.

Helper Function

  • get_initializer(name_or_callable=None) — Resolve name ('glorot_uniform', 'he_normal', 'normal', etc.) or callable; defaults to GlorotUniform().

Compatibility Aliases

  • XavierUniform = GlorotUniform
  • XavierNormal = GlorotNormal
  • RandomNormal = Normal

Author: Nirmal Parmar