Algorithms

ProxGrad(model, step = 1.0)

Proximal gradient method with step size step. Works for any loss and any penalty with a prox method.

Example

x, y, β = SparseRegression.fakedata(L2DistLoss(), 1000, 10)
s = SModel(x, y, L2DistLoss())
strat = strategy(MaxIter(50), ProxGrad(s))
learn!(s, strat)

Fista(model, step = 1.0)

Accelerated proximal gradient method. Works for any loss and any penalty with a prox method.

AdaptiveProxGrad(s, divisor = 1.5)

Proximal gradient method with adaptive step sizes. AdaptiveProxGrad uses element-wise learning rates. Every time the sign of a coefficient switches, the step size for that coefficient is divided by divisor.

GradientDescent(model, step = 1.0)

Gradient Descent. Works for any loss and any penalty.

Example

x, y, β = SparseRegression.fakedata(L2DistLoss(), 1000, 10)
s = SModel(x, y, L2DistLoss())
strat = strategy(MaxIter(50), GradientDescent(s))
learn!(s, strat)

Sweep(model)

Linear/ridge regression via sweep operator. Works for (scaled) L2DistLoss with NoPenalty or L2Penalty. The Sweep algorithm has a closed form solution and is complete after one iteration. It therefore doesn't need additional learning strategies such as MaxIter, Converged, etc.

Example

x, y, β = SparseRegression.fakedata(L2DistLoss(), 1000, 10)
s = SModel(x, y, L2DistLoss())
learn!(s, Sweep(s))

LinRegCholesky(model)

Linear/ridge regression via cholesky decomposition. Works for (scaled) L2DistLoss with NoPenalty or L2Penalty. The LinRegCholesky algorithm has a closed form solution and is complete after one iteration. It therefore doesn't need additional learning strategies such as MaxIter, Converged, etc.

Example

x, y, β = SparseRegression.fakedata(L2DistLoss(), 1000, 10)
s = SModel(x, y, L2DistLoss())
learn!(s, Sweep(s))