InverseScaling

Reduces the learning rate using a power schedule.

Assuming an initial learning rate \(\eta\), the learning rate at step \(t\) is:

\[\\frac{eta}{(t + 1) ^ p}\]

where \(p\) is a user-defined parameter.

Parameters

• learning_rate ('float')

• power – defaults to 0.5

Methods

get

Returns the learning rate at a given iteration.

Parameters

• t ('int')