Skip to content

Initializer — Abstract Base Initializer Class

The foundational abstract base class for all MAGNET weight initializers.


Overview

Initializer is an abstract base class (ABC) that defines the interface for all weight initialization strategies in MAGNET. It provides common utilities for shape validation and fan-in/fan-out computation, and requires subclasses to implement the __call__ method for generating weight arrays.


Key Features

  • Abstract Interface: Enforces consistent __call__(shape) signature across all initializers
  • Shape Normalization: Helper method normalize_shape() for dimension validation
  • Fan Computation: compute_fans() utility for calculating input/output dimensions
  • Dtype Control: Configurable output data type (default np.float64)

Parameters

Parameter Type Default Description
dtype np.dtype np.float64 Output data type for generated arrays

Methods

__call__(shape) (abstract)

Generate and return a weight array of the given shape. Must be implemented by all subclasses.

normalize_shape(shape)

Validate and normalize the shape tuple, raising errors for invalid inputs.

compute_fans(shape)

Compute fan-in and fan-out dimensions from a shape tuple, useful for Glorot/He initializers.


Example Usage

from machinegnostics.magnet import Initializer

class MyInitializer(Initializer):
    def __call__(self, shape):
        shape = self.normalize_shape(shape)
        return np.ones(shape, dtype=self.dtype)

init = MyInitializer()
W = init((10, 20))

Author: Nirmal Parmar