Only save performance of hyperparameter combination in training cache

[lhjohn]

In gridCVDeep thegridSearchPrediction object grows to several GBs on larger data. https://github.com/OHDSI/DeepPatientLevelPrediction/blob/2aba7580f8410d623dfb6f012ce6fed95bdf417f/R/Estimator.R#L368 The training cache uses this object and saves it as RDS file, which can evetnually take 5+ minutes.

We should only save the performance, which currently is only AUC.

[lhjohn]

This won't be as easy as I thought, since we need to keep track of the best prediction as well. The best prediction will be attached as CV prediction to the result.

[lhjohn]

Here is a script that converts old caches to new caches. This is done in place, so a backup of the old cache maybe useful.

path <- file.path("D:/backup/paramPersistence.rds")

paramPersistence <- readRDS(path)

for (gridId in 1:length(paramPersistence$gridSearchPredictions)) {

  if (!is.null(paramPersistence$gridSearchPredictions[[gridId]])) {
    gridPerformance <- PatientLevelPrediction::computeGridPerformance(paramPersistence$gridSearchPredictions[[gridId]]$prediction, paramPersistence$gridSearchPredictions[[gridId]]$param)

    paramPersistence$gridSearchPredictions[[gridId]] <- list(
      prediction = paramPersistence$gridSearchPredictions[[gridId]]$prediction,
      param = paramPersistence$gridSearchPredictions[[gridId]]$param,
      gridPerformance = gridPerformance
    )

  }
  print(paste0("Add grid performance for index: ", gridId))
}

# remove all predictions that are not the max performance
indexOfMax <- which.max(unlist(lapply(paramPersistence$gridSearchPredictions, function(x) x$gridPerformance$cvPerformance)))
for (i in seq_along(paramPersistence$gridSearchPredictions)) {
  if (!is.null(paramPersistence$gridSearchPredictions[[i]])) {
    if (i != indexOfMax) {
      paramPersistence$gridSearchPredictions[[i]]$prediction <- list(NULL)
    }
  }
  print(paste0("Remove prediction for index: ", i))

}

saveRDS(paramPersistence, path)