checkpoint_earlystop: Early stopping check

In egenn/rtemis: Machine Learning and Visualization

Early stopping check

Description

Returns list with relative variance over n.steps, absolute.threshold, last value, and logical "stop", if conditions are met and training should stop. The final stop decision is: check.thresh | (check.rthresh & check.rvar) if combine.relative.thresholds = "AND" or check.thresh | (check.rthresh | check.rvar) if combine.relative.thresholds = "OR"

Usage

checkpoint_earlystop(
  x,
  absolute.threshold = NA,
  relative.threshold = NA,
  minimize = TRUE,
  relativeVariance.threshold = NA,
  n.steps = 10,
  combine.relative.thresholds = "AND",
  min.steps = 50,
  na.response = c("stop", "continue"),
  verbose = TRUE
)

Arguments

x	Float, vector: Input - this would normally be the loss at each iteration
absolute.threshold	Float: If set and the last value of x is less than or equal to this (if minimize = TRUE) or greater than or equal to this (if minimize = FALSE), then return stop = TRUE. See output under Value. Default = NA
relative.threshold	Float: If set, checks if the relative change from the first to last value of x exceeds this number. i.e. if set to .9 and minimize = TRUE, if there is a 90 percent drop from x[1] to x[length(x)], then the function returns stop = TRUE. If minimize = FALSE, then checks if there is a 90 percent increase, accordingly.
minimize	Logical: See absolute.threshold. Default = TRUE
relativeVariance.threshold	Float: If relative variance over last n.steps is less than or equal to this, return stop = TRUE. See output under Value
n.steps	Integer; > 1: Calculate relative variance over this many last values of x
combine.relative.thresholds	Character: "AND" or "OR": How to combine the criteria relative.threshold and relativeVariance.threshold. Default = "AND", which means both must be TRUE to stop. The scenario is you might want to check relastiveVariance threshold only after a certain amount of learning has taken place, which you can't predict with min.steps but would rather quantify with relative.threshold.
min.steps	Integer: Do not calculate relativeVariance unless x is at least this length
na.response	Character: "stop" or "continue": what should happen if the last value of x is NA
verbose	Logical: If TRUE, print messages to console

Value

List with the following items:

• last.value Float: Last value of x

• relativeVariance Float: relative variance of last n.steps

• check.thresh Logical: TRUE, if absolute threshold was reached

• check.rvar Logical: TRUE, if relative variance threshold was reached

• stop Logical: TRUE, if either criterion was met - absolute threshold or relativeVariance.threshold

Author(s)

E.D. Gennatas

egenn/rtemis documentation built on Dec. 17, 2024, 6:16 p.m.