predict-individual.Rmd

```{r init}
source("lib/function_library.R", echo = F)

# Load other R source files in the lib directory.
load_all_code("lib")

# Load a bunch of required packages; optionally installing missing ones as needed.
load_all_packages(auto_install = F)

conf = list(verbose = F,
            dir = ".",
            data_dir = "data",
            tex_dir = "tex",
            visual_dir = "visuals")
```

```{r predict-setup}

# Setup multinode/multicore processing if available.
conf$cluster = ck37r::parallelize(allow_multinode = T)

#conf$method = "method.AUC"
conf$method = "method.NNLS"

# conf$outer_cv_folds = 10
conf$outer_cv_folds = 20
conf$cv_folds = 10

# If the cluster object is null it means we're setup for multicore (or nuthin).
if (is.null(conf$cluster) || is.na(conf$cluster)) {
  # Use non-CV if we're running this in RStudio on a laptop.
  if (.Platform$GUI == "RStudio") {
    fn = "sl_fn"
  } else {
    fn = "cv_sl_fn"
  }
  # Manually change this to F to force non-parallel SuperLearner.
  if (T) {
    # Multicore version.
    conf$sl_fn = ck37r::gen_superlearner(parallel = "multicore",
                   outer_cv_folds = conf$outer_cv_folds)[[fn]]
  } else {
    # Sequential version.
    conf$sl_fn = SuperLearner
  }
  rm(fn)
} else {
  # Multinode version.
  conf$sl_fn = ck37r::gen_superlearner(parallel = "snow",
                  cluster = conf$cluster,
                  outer_cv_folds = conf$outer_cv_folds)$cv_sl_fn
}
```

## Individual Prediction

```{r predict}
# Backup our current configuration before we load the RData file.
conf_backup = conf
# Load data from the ariane-tbi-predict.Rmd file, to avoid duplicating code.
load(paste0(conf$dir, "/data/create-dataset.RData"))
# Restore our backup configuration.
conf = conf_backup
rm(conf_backup)

if (foreach::getDoParName() == "doSNOW") {
  # Set multinode-compatible seed.
  clusterSetRNGStream(conf$cluster, iseed = 1)
} else {
  # Set multicore-compatible seed.
  set.seed(1, "L'Ecuyer-CMRG")
}

X = dataset$X
Y = dataset$outcomes$progression

libs = individual_library_seq(dim(X), type = "regression", glm = F)
length(libs)

# Review final data dimensions.
dim(X)

# Temporaryily reduce library size.
# rpartPrune not working :/
# libs = libs[1:4]

# Test glmnet and XGB
libs = libs[c(3, 11)]

results = fit_model_libs(Y, X, libs, family=gaussian(), sl_fn = conf$sl_fn,
                     cv_folds = conf$cv_folds, cluster = conf$cluster)

save(results, libs, conf, dataset, file = paste0(conf$data_dir, "/predict-indiv-results.RData"))
```

Process results.

```{r process-results, eval=F}
load(paste0(conf$data_dir, "/predict-indiv-results.RData"))

# TODO: update this for PPMI!!!

# Create AUC plots for each SL result, plus a summary table.
lib_names = c("Mean", "Pruned Tree", "Elastic Net", "Elastic Net",
              "Support Vector Machine (radial)",
              "Multivariate Adaptive Regresson Spline (MARS)", "Polymars",
              "Elastic Net screener + OLS",
              "Random Forest", "Gradient Boosted Machine",
              "XGBoost (GBM)", "XGBoost (GBM)")
analyze_results = plot_all_results(results, Y, "indiv", libs,
                 plot_dir = paste0(conf$visual_dir, "/visuals/"),
                 tex_dir = paste0(conf$tex_dir, "/tex/"), lib_names = lib_names)
save(analyze_results, file=paste0(conf$data_dir, "/predict-indiv-analysis.RData"))
```

```{r cleanup}
ck37r::stop_cluster(conf$cluster)
```