trop_control() sets the options of the weighted low-rank solver in C. The
estimate, the cross-validation, and the bootstrap all use this solver.
Usage
trop_control(
max_iter = 100L,
max_inner = 10L,
tol = 1e-06,
threads = trop_threads(),
warm = TRUE,
solver = c("prox", "block"),
throttle = FALSE
)Arguments
- max_iter
The maximum number of outer iterations. With
solver = "block", each outer iteration is one sweep of the blocks. Withsolver = "prox", the maximum number of proximal steps ismax_iter * max_inner.- max_inner
The number of proximal steps in each outer iteration. Each step is a step of FISTA, an accelerated proximal gradient method.
- tol
The convergence tolerance. The solver stops when the largest absolute change of the iterates between two steps is less than
tol.- threads
The number of 'OpenMP' threads. The default comes from
trop_threads().- warm
Logical. If
TRUE, the fit for a cell starts from the solution for the previous cell of the same unit. IfFALSE, each fit starts from zero. Both options give the same results in the limits oftol. The number of threads does not change the results.- solver
Use
"prox", the default, or"block"."prox"is an accelerated proximal gradient method on the low-rank term. At each step, it calculates the unit and time effects again, and it uses an adaptive restart."block"is the block coordinate method of the reference implementations. Both methods find the minimum of the same convex objective."prox"usually calculates fewer singular value decompositions.- throttle
Logical. Windows can use power throttling (EcoQoS) to make processes that run in the background for a long time slower. An R session with a long calculation can be such a process.
With
FALSE, the default, each thread that runs the C code of fasttrop tells Windows not to use power throttling on it. When the C code stops, Windows controls the thread again. WithTRUE, Windows uses its usual policy. On other operating systems, this option has no effect.
