Function vikos::inert_gradient_descent_step
[−]
[src]
pub fn inert_gradient_descent_step<C, M>(cost: &C, model: &mut M, features: &M::Input, truth: M::Target, learning_rate: M::Target, inertia: M::Target, velocity: &mut Vec<M::Target>) where C: Cost, M: Model<Target=C::Error>
Changes all coefficents of model based on their derivation of the cost function at features
Will not get stuck on saddle points as easily as a plain SGD and will converge quicker in general.
A good default for inertia
is 0.9