Tune a Deep Learning Estimator

Performs hyperparameter search and cross-validation, then fits a final model using the best-performing settings.

Usage

gridCvDeep(mappedData, labels, modelSettings, modelLocation, analysisPath)

Arguments

mappedData

Covariate data mapped with PatientLevelPrediction::MapIds().

labels

Data frame containing outcomes and fold assignments.

modelSettings

A modelSettings object.

modelLocation

Directory in which to save the fitted model.

analysisPath

Directory used for the resumable training cache.

Value

A list containing the saved estimator location, predictions, final parameters, hyperparameter-search summaries, and feature information.

Examples

if (FALSE) { # \dontrun{
# Requires FeatureExtraction and the package's Python dependencies,
# including PyTorch.
# See vignette("Installing") for setup instructions.
data("simulationProfile", package = "PatientLevelPrediction")
plpData <- PatientLevelPrediction::simulatePlpData(
  simulationProfile, n = 200, seed = 42
)
# Supply metadata omitted by simulatePlpData() for this bundled profile:
# gender, age, conditions, and drugs. Only age is continuous.
plpData$covariateData$analysisRef <- data.frame(
  analysisId = c(1, 2, 102, 402),
  isBinary = c("Y", "N", "Y", "Y"),
  missingMeansZero = c(NA, "Y", NA, NA)
)
population <- PatientLevelPrediction::createStudyPopulation(
  plpData,
  populationSettings = PatientLevelPrediction::createStudyPopulationSettings(
    riskWindowEnd = 90, minTimeAtRisk = 89
  )
)
splitData <- PatientLevelPrediction::splitData(
  plpData,
  population,
  splitSettings = PatientLevelPrediction::createDefaultSplitSetting(
    testFraction = 0, trainFraction = 1, nfold = 2, splitSeed = 42
  )
)

# Keep this toy example small: two configurations, two folds, and one epoch.
# These settings demonstrate fitting, not meaningful predictive performance.
modelSettings <- setResNet(
  numLayers = 1,
  sizeHidden = c(8, 16),
  hiddenFactor = 1,
  residualDropout = 0,
  hiddenDropout = 0,
  sizeEmbedding = 8,
  hyperParamSearch = "grid",
  estimatorSettings = setEstimator(
    learningRate = 0.001, batchSize = 64, epochs = 1, seed = 42
  )
)
analysisPath <- tempfile("deep-plp-example-")
dir.create(analysisPath)

# The lower-level interface needs mapped IDs and fold assignments in labels.
labels <- merge(splitData$Train$labels, splitData$Train$folds, by = "rowId")
mappedData <- PatientLevelPrediction::MapIds(
  covariateData = splitData$Train$covariateData,
  cohort = labels
)
result <- gridCvDeep(
  mappedData = mappedData,
  labels = labels,
  modelSettings = modelSettings,
  modelLocation = file.path(analysisPath, "model"),
  analysisPath = analysisPath
)
result$finalParam

# Clean up this example's data, cache, and final model.
Andromeda::close(mappedData)
Andromeda::close(splitData$Train$covariateData)
Andromeda::close(plpData$covariateData)
unlink(analysisPath, recursive = TRUE)
} # }