mvpa_model: Create an MVPA Model

In bbuchsbaum/rMVPA: Multivoxel Pattern Analysis in R

Create an MVPA Model

Description

Create an MVPA model based on a caret-based classification or regression model.

Usage

mvpa_model(
  model,
  dataset,
  design,
  model_type = c("classification", "regression"),
  crossval = NULL,
  feature_selector = NULL,
  tune_grid = NULL,
  tune_reps = 15,
  performance = NULL,
  class_metrics = TRUE
)

Arguments

model	A caret-based classification or regression model.
dataset	An 'mvpa_dataset' instance.
design	An 'mvpa_design' instance.
model_type	A character string indicating the problem type: "classification" or "regression".
crossval	An optional 'cross_validation' instance.
feature_selector	An optional 'feature_selector' instance.
tune_grid	An optional parameter tuning grid as a 'data.frame'.
tune_reps	The number of replications used during parameter tuning. Only relevant if 'tune_grid' is supplied.
performance	An optional custom function for computing performance metrics.
class_metrics	A logical flag indicating whether to compute performance metrics for each class.

Details

If 'performance' is supplied, it must be a function that takes one argument and returns a named list of scalar values. The argument the function takes is a class deriving from 'classification_result' appropriate for the problem at hand. See example below.

Examples

mod <- load_model("sda")
traindat <- neuroim2::NeuroVec(array(rnorm(6*6*6*100), c(6,6,6,100)), neuroim2::NeuroSpace(c(6,6,6,100)))
mask <- neuroim2::LogicalNeuroVol(array(rnorm(6*6*6)>-.2, c(6,6,6)), neuroim2::NeuroSpace(c(6,6,6)))

mvdset <- mvpa_dataset(traindat,mask=mask)
design <- data.frame(fac=rep(letters[1:4], 25), block=rep(1:10, each=10))
cval <- blocked_cross_validation(design$block)
mvdes <- mvpa_design(design, ~ fac, block_var=~block)

custom_perf <- function(result) {
  c(accuracy=sum(result$observed == result$predicted)/length(result$observed))
}
mvpmod <- mvpa_model(mod, dataset=mvdset, design=mvdes, crossval=cval, performance=custom_perf)
ret <- run_searchlight(mvpmod)
stopifnot("accuracy" %in% names(ret))

bbuchsbaum/rMVPA documentation built on April 23, 2024, 7:35 a.m.