TunerHyperband class that implements hyperband tuning. Hyperband is a
budget oriented-procedure, weeding out suboptimal performing configurations
early in a sequential training process, increasing tuning efficiency as a
consequence.
For this, several brackets are constructed with an associated set of
configurations for each bracket. Each bracket as several stages. Different
brackets are initialized with different amounts of configurations and
different budget sizes. To get an idea of how the bracket layout looks like
for a given argument set, please have a look in the details.
To identify the budget for evaluating hyperband, the user has to specify
explicitly which hyperparameter of the learner influences the budget by
tagging a single hyperparameter in the paradox::ParamSet with "budget".
An alternative approach using subsampling and pipelines is described below.
Naturally, hyperband terminates once all of its brackets are evaluated, so a bbotk::Terminator in the tuning instance acts as an upper bound and should be only set to a low value if one is unsure of how long hyperband will take to finish under the given settings.
Source
Li L, Jamieson K, DeSalvo G, Rostamizadeh A, Talwalkar A (2018). “Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization.” Journal of Machine Learning Research, 18(185), 1-52. https://jmlr.org/papers/v18/16-558.html.
Details
This sections explains the calculation of the constants for each bracket.
A small overview will be given here, but for more details please check
out the original paper (see references).
To keep things uniform with the notation in the paper (and to safe space in
the formulas), R is used for the upper budget that last remaining
configuration should reach. The formula to calculate the amount of brackets
is floor(log(R, eta)) + 1. To calculate the starting budget in each
bracket, use R * eta^(-s), where s is the maximum bracket minus the
current bracket index.
For the starting configurations in each bracket it is
ceiling((B/R) * ((eta^s)/(s+1))), with B = (bracket amount) * R.
To receive a table with the full brackets layout, load the following function
and execute it for the desired R and eta.
hyperband_brackets = function(R, eta) { result = data.frame() smax = floor(log(R, eta)) B = (smax + 1) * R # outer loop - iterate over brackets for (s in smax:0) { n = ceiling((B/R) * ((eta^s)/(s+1))) r = R * eta^(-s) # inner loop - iterate over bracket stages for (i in 0:s) { ni = floor(n * eta^(-i)) ri = r * eta^i result = rbind(result, c(smax - s + 1, i + 1, ri, ni)) } } names(result) = c("bracket", "bracket_stage", "budget", "n_configs") return(result) } hyperband_brackets(R = 81L, eta = 3L)
Parameters
etanumeric(1)
Fraction parameter of the successive halving algorithm: With every step the configuration budget is increased by a factor ofetaand only the best1/etaconfigurations are used for the next stage. Non-integer values are supported, butetais not allowed to be less or equal 1.samplerparadox::Sampler
Object defining how the samples of the parameter space should be drawn during the initialization of each bracket. The default is uniform sampling.
Archive
The mlr3tuning::ArchiveTuning holds the following additional columns that are specific to the hyperband tuner:
bracket(integer(1))
The console logs about the bracket index are actually not matching with the original hyperband algorithm, which counts down the brackets and stops after evaluating bracket 0. The true bracket indices are given in this column.bracket_stage(integer(1))
The bracket stage of each bracket. Hyperband starts counting at 0.budget_scaled(numeric(1))
The intermediate budget in each bracket stage calculated by hyperband. Because hyperband is originally only considered for budgets starting at 1, some rescaling is done to allow budgets starting at different values. For this, budgets are internally divided by the lower budget bound to get a lower budget of 1. Before the learner receives its budgets for evaluation, the budget is transformed back to match the original scale again.budget_real(numeric(1))
The real budget values the learner uses for evaluation after hyperband calculated its scaled budget.n_configs(integer(1))
The amount of evaluated configurations in each stage. These correspond to ther_iin the original paper.
Hyperband without learner budget
Thanks to mlr3pipelines, it is possible to use hyperband in combination with learners lacking a natural budget parameter. For example, any mlr3::Learner can be augmented with a mlr3pipelines::PipeOp operator such as mlr3pipelines::PipeOpSubsample. With the subsampling rate as budget parameter, the resulting mlr3pipelines::GraphLearner is fitted on small proportions of the mlr3::Task in the first brackets, and on the complete Task in last brackets. See examples for some code.
Custom sampler
Hyperband supports custom paradox::Sampler object for initial
configurations in each bracket.
A custom sampler may look like this (the full example is given in the
examples section):
# - beta distribution with alpha = 2 and beta = 5 # - categorical distribution with custom probabilities sampler = SamplerJointIndep$new(list( Sampler1DRfun$new(params[[2]], function(n) rbeta(n, 2, 5)), Sampler1DCateg$new(params[[3]], prob = c(0.2, 0.3, 0.5)) ))
Runtime scaling w.r.t. the chosen budget
The calculation of each bracket currently assumes a linear runtime in the
chosen budget parameter is always given. Hyperband is designed so each
bracket requires approximately the same runtime as the sum of the budget
over all configurations in each bracket is roughly the same. This will not
hold true once the scaling in the budget parameter is not linear
anymore, even though the sum of the budgets in each bracket remains the
same. A basic example can be viewed by calling the function
hyperband_brackets below with the arguments R = 2 and eta = 2. If we
run a learner with O(budget^2) time complexity, the runtime of the last
bracket will be 33% longer than the first bracket
(time of bracket 1 = 2 * 1^2 + 2^2 = 6; time of bracket 2 = 2 * 2^2 = 8).
Of course, this won't break anything, but it should be kept in mind when
applying hyperband. A possible adaption would be to introduce a trafo,
like it is shown in the examples.
Logging
When loading the mlr3hyperband package, two loggers based on the lgr
package are made available. One is called mlr3, the other bbotk. All
mlr3 methods log into the mlr3 logger. All optimization methods form the
packags bbotk, mlr3tuning and mlr3hyperband log into the bbotk
logger. To hide the mlr3 logging messages run:
lgr::get_logger("mlr3")$set_threshold("warn")
Super class
mlr3tuning::Tuner -> TunerHyperband
Methods
Public methods
Inherited methods
Method new()
Creates a new instance of this R6 class.
Usage
TunerHyperband$new()
Method clone()
The objects of this class are cloneable with this method.
Usage
TunerHyperband$clone(deep = FALSE)
Arguments
deepWhether to make a deep clone.
Examples
if(requireNamespace("xgboost")) { library(mlr3) library(mlr3learners) library(paradox) library(mlr3tuning) library(mlr3hyperband) # Define hyperparameter and budget parameter for tuning with hyperband ps = ParamSet$new(list( ParamInt$new("nrounds", lower = 1, upper = 4, tag = "budget"), ParamDbl$new("eta", lower = 0, upper = 1), ParamFct$new("booster", levels = c("gbtree", "gblinear", "dart")) )) # Define termination criterion # Hyperband terminates itself terminator = trm("none") # Create tuning instance inst = TuningInstanceSingleCrit$new( task = tsk("iris"), learner = lrn("classif.xgboost"), resampling = rsmp("holdout"), measure = msr("classif.ce"), search_space = ps, terminator = terminator, ) # Load tuner tuner = tnr("hyperband", eta = 2L) # \donttest{ # Trigger optimization tuner$optimize(inst) # Print all evaluations as.data.table(inst$archive)# } }#>#> [04:29:21] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:21] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:21] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:21] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:22] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:22] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:22] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:22] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:22] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:22] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:23] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:23] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:23] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior. #> [04:29:23] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'multi:softprob' was changed from 'merror' to 'mlogloss'. Explicitly set eval_metric if you'd like to restore the old behavior.#> eta booster nrounds bracket bracket_stage budget_scaled budget_real #> 1: 0.03745117 dart 1 2 0 1 1 #> 2: 0.59355379 dart 1 2 0 1 1 #> 3: 0.23697755 gbtree 1 2 0 1 1 #> 4: 0.90629727 dart 1 2 0 1 1 #> 5: 0.03745117 dart 2 2 1 2 2 #> 6: 0.59355379 dart 2 2 1 2 2 #> 7: 0.03745117 dart 4 2 2 4 4 #> 8: 0.91766051 dart 2 1 0 2 2 #> 9: 0.97284422 gblinear 2 1 0 2 2 #> 10: 0.81908245 dart 2 1 0 2 2 #> 11: 0.91766051 dart 4 1 1 4 4 #> 12: 0.72175973 gbtree 4 0 0 4 4 #> 13: 0.86661570 dart 4 0 0 4 4 #> 14: 0.23845311 gblinear 4 0 0 4 4 #> n_configs classif.ce uhash #> 1: 4 0.08 8095ddab-98fc-43b0-941b-eff009a394c2 #> 2: 4 0.08 e66e53cd-6dea-4dbb-9e66-52cb98e2ecf7 #> 3: 4 0.08 a636914d-b302-4147-a3bc-51abc1cc1c9b #> 4: 4 0.08 00e531b5-a64f-4f63-817a-54a552a99b82 #> 5: 2 0.08 14a355cd-fd1d-412c-b3a4-fe0a1f0dea6d #> 6: 2 0.08 adffedc6-bdf9-4d00-b2d8-6791dae7a8b9 #> 7: 1 0.08 b216fdcc-dc5c-4394-bb6e-2ad6a6250147 #> 8: 3 0.04 148f7095-69c6-4912-aa8f-f6645b1ce078 #> 9: 3 0.40 540e1f69-ed8f-411e-9d14-379f4f910204 #> 10: 3 0.04 0163bf92-2bae-400c-9dd7-1fa2c869dbed #> 11: 1 0.04 a3d7c27a-6914-45db-82b4-b248e7adc9de #> 12: 3 0.04 2e6d8e02-4d1b-40a7-8b16-78c56d711389 #> 13: 3 0.04 c3fedf01-6fc5-48bc-a83d-a5254b5d6345 #> 14: 3 0.40 3255a53a-76d8-4b93-8977-b648e5f75c0d #> timestamp batch_nr x_domain_nrounds x_domain_eta x_domain_booster #> 1: 2021-03-21 04:29:22 1 1 0.03745117 dart #> 2: 2021-03-21 04:29:22 1 1 0.59355379 dart #> 3: 2021-03-21 04:29:22 1 1 0.23697755 gbtree #> 4: 2021-03-21 04:29:22 1 1 0.90629727 dart #> 5: 2021-03-21 04:29:22 2 2 0.03745117 dart #> 6: 2021-03-21 04:29:22 2 2 0.59355379 dart #> 7: 2021-03-21 04:29:22 3 4 0.03745117 dart #> 8: 2021-03-21 04:29:23 4 2 0.91766051 dart #> 9: 2021-03-21 04:29:23 4 2 0.97284422 gblinear #> 10: 2021-03-21 04:29:23 4 2 0.81908245 dart #> 11: 2021-03-21 04:29:23 5 4 0.91766051 dart #> 12: 2021-03-21 04:29:23 6 4 0.72175973 gbtree #> 13: 2021-03-21 04:29:23 6 4 0.86661570 dart #> 14: 2021-03-21 04:29:23 6 4 0.23845311 gblinear