Edit on GitHub

The Solver State

Given an AbstractManoptProblem, that is a certain optimisation task, the state specifies the solver to use. It contains the parameters of a solver and all fields necessary during the algorithm, e.g. the current iterate, a StoppingCriterion or a Stepsize.

Manopt.AbstractManoptSolverState — Type
AbstractManoptSolverState

A general super type for all solver states.

Fields

The following fields are assumed to be default. If you use different ones, provide the access functions accordingly

source

Since every subtype of an AbstractManoptSolverState directly relate to a solver, the concrete states are documented together with the corresponding solvers. This page documents the general functionality available for every state.

A first example is to access, i.e. obtain or set, the current iterate. This might be useful to continue investigation at the current iterate, or to set up a solver for a next experiment, respectively.

Manopt.get_iterate — Function
get_iterate(O::AbstractManoptSolverState)

return the (last stored) iterate within AbstractManoptSolverStates`. By default also undecorates the state beforehand.

source
get_iterate(agst::AbstractGradientSolverState)

return the iterate stored within gradient options. THe default resturns agst.p.

source
Manopt.get_gradient — Method
get_gradient(M::AbstractManifold, sgo::ManifoldStochasticGradientObjective, p, k)
get_gradient!(M::AbstractManifold, sgo::ManifoldStochasticGradientObjective, Y, p, k)

Evaluate one of the summands gradients $\operatorname{grad}f_k$, $k∈\{1,…,n\}$, at x (in place of Y).

If you use a single function for the stochastic gradient, that works inplace, then get_gradient is not available, since the length (or number of elements of the gradient required for allocation) can not be determined.

source
get_gradient(M::AbstractManifold, sgo::ManifoldStochasticGradientObjective, p)
get_gradient!(M::AbstractManifold, sgo::ManifoldStochasticGradientObjective, X, p)

Evaluate the complete gradient $\operatorname{grad} f = \displaystyle\sum_{i=1}^n \operatorname{grad} f_i(p)$ at p (in place of X).

If you use a single function for the stochastic gradient, that works inplace, then get_gradient is not available, since the length (or number of elements of the gradient required for allocation) can not be determined.

source
X = get_gradient(M::AbstractManifold, p::ManifoldAlternatingGradientObjective, p, k)
get_gradient!(M::AbstractManifold, p::ManifoldAlternatingGradientObjective, X, p, k)

Evaluate one of the component gradients $\operatorname{grad}f_k$, $k∈\{1,…,n\}$, at x (in place of Y).

source

Decorators for AbstractManoptSolverState

A solver state can be decorated using the following trait and function to initialize

Manopt.dispatch_state_decorator — Function
dispatch_state_decorator(s::AbstractManoptSolverState)

Indicate internally, whether an AbstractManoptSolverState s to be of decorating type, i.e. it stores (encapsulates) state in itself, by default in the field s.state.

Decorators indicate this by returning Val{true} for further dispatch.

The default is Val{false}, i.e. by default an state is not decorated.

source
Manopt.decorate_state! — Function
decorate_state!(s::AbstractManoptSolverState)

decorate the AbstractManoptSolverStates with specific decorators.

Optional Arguments

optional arguments provide necessary details on the decorators. A specific one is used to activate certain decorators.

  • debug – (Array{Union{Symbol,DebugAction,String,Int},1}()) a set of symbols representing DebugActions, Strings used as dividers and a subsampling integer. These are passed as a DebugGroup within :All to the DebugSolverState decorator dictionary. Only excention is :Stop that is passed to :Stop.
  • record – (Array{Union{Symbol,RecordAction,Int},1}()) specify recordings by using Symbols or RecordActions directly. The integer can again be used for only recording every $i$th iteration.
  • return_state - (false) indicate whether to wrap the options in a ReturnSolverState, indicating that the solver should return options and not (only) the minimizer.

other keywords are ignored.

See also

DebugSolverState, RecordSolverState, ReturnSolverState

source

A simple example is the

as well as DebugSolverState and RecordSolverState.

State Actions

A state action is a struct for callback functions that can be attached within for example the just mentioned debug decorator or the record decorator.

Several state decorators or actions might store intermediate values like the (last) iterate to compute some change or the last gradient. In order to minimise the storage of these, there is a generic StoreStateAction that acts as generic common storage that can be shared among different actions.

Manopt.StoreStateAction — Type
StoreStateAction <: AbstractStateAction

internal storage for AbstractStateActions to store a tuple of fields from an AbstractManoptSolverStates

This functor posesses the usual interface of functions called during an iteration, i.e. acts on (p,o,i), where p is a AbstractManoptProblem, o is an AbstractManoptSolverState and i is the current iteration.

Fields

  • values – a dictionary to store interims values based on certain Symbols
  • keys – an NTuple of Symbols to refer to fields of AbstractManoptSolverState
  • once – whether to update the internal values only once per iteration
  • lastStored – last iterate, where this AbstractStateAction was called (to determine once

Constructiors

AbstractStateAction([keys=(), once=true])

Initialize the Functor to an (empty) set of keys, where once determines whether more that one update per iteration are effective

AbstractStateAction(keys, once=true])

Initialize the Functor to a set of keys, where the dictionary is initialized to be empty. Further, once determines whether more that one update per iteration are effective, otherwise only the first update is stored, all others are ignored.

source

Abstract States

In a few cases it is useful to have a hierarchy of types. These are