window_grad {CNN} R Documentation

This is AdaGrad but with a moving window weighted average

Description

so the gradient is not accumulated over the entire history of the run.
it's also referred to as Idea #1 in Zeiler paper on AdaDelta.

Usage

window_grad(batch.size,
    l2.decay = 0.001,
    ro = 0.95);

Arguments

batch.size

[as integer]

l2.decay

[as double]

ro

[as double]

Details

Authors

MLkit

Value

this function returns data object of type TrainerAlgorithm.

clr value class

Examples

[Package CNN version 1.0.0.0 Index]