initialisers

Initializers module for tensor initialization.

This module provides functions for initializing tensors with specific distributions or patterns.

init_xavier(shape, name='')[source]

Initialize a tensor with Xavier/Glorot initialization.

This function implements Xavier/Glorot initialization, which helps in setting initial random weights for neural networks. It’s particularly useful for maintaining the scale of gradients across layers.

Parameters:
  • shape (tuple[int, int]) – A tuple of two integers (f_in, f_out), where f_in is the number of input units and f_out is the number of output units.

  • name (str) – An optional string to name the created tensor. Defaults to an empty string.

Returns:

A Tensor object initialized with Xavier/Glorot initialization.

Raises:

ValueError – If the shape tuple does not contain exactly two integers.

Return type:

Tensor

Example

>>> weight = init_xavier((100, 50), name="layer1_weights")