Package com.jml.neural_network.layers.initilizers


package com.jml.neural_network.layers.initilizers
A package containing an interface for specifying a layer parameter initializer. Also includes various built-in initializers. These initializers can be used to set the staring values of the weights / bias terms (or any other trainable parameters in the layer) before training.
  • Interface Summary
    Interface
    Description
    Interface for bias/weight initialization.
  • Class Summary
    Class
    Description
    layer parameter initializer to produce a constant value.
    {layer parameter initializer to produce random values from a truncated normal distribution with standard deviation std = sqrt(2 / (fanIn + fanOut)) and mean of 0 where fanIn is the input dimension for the layer and fanOut is the output dimension for the layer.
    layer parameter initializer to produce random values from a clipped uniform distribution in [-lim, lim] where lim = sqrt(6 / (fanIn + fanOut)), fanIn is the input dimension for the layer, and fanOut is the output dimension for the layer.
    layer parameter initializer to produce random values from a normal distribution with standard deviation std = sqrt(2 / fanIn) and mean of 0 where fanIn is the input dimension for the layer.
    layer parameter initializer to produce random values from a clipped uniform distribution in [-lim, lim] where lim = sqrt(6 / fanIn), fanIn is the input dimension for the layer.
    layer parameter initializer to produce ones.
    layer parameter initializer to produce random orthogonal matrix.
    layer parameter initializer to produce random values from a normal distribution.
    layer parameter initializer to produce random values from a uniform distribution.
    layer parameter initializer to produce zeros.