Preliminary

gpredomicsR is an R package that wraps the gpredomics rust library, which implements accelerated heuristics to find sparse interpretable models described in Prifti et al, Gigascience, 2020. The main algorithm is a genetic algorithm which applies numerous operators to evolve individuals, which are classification models. The fit of the models is evaluated on a penalized AUC.

Compiling

Open an R session at root and:

devtools::build(".")
install.packages("../gpredomicsR_0.0.0.9000.tar.gz", repos = NULL, type = "source")

Using

Creating the environment

After loading the package, we need to load the parameters and create a RunningFlag object. param is a rust pointer and is need to call rust internal functions. However, it is possible to get it as an R object with param$get().

library(gpredomicsR)
running_flag <- RunningFlag$new()
param <- Param$load("param.yaml")
rparam <- param$get()

Running the algorithm

The main heuristic based on genetic algorithm ga() is called with the param and running_flag objects. This will return a Experiment object which will store the results of the algorithm. The results in R are a list generations, each being a Population object, which contains the individuals of each generation. Several functions allow to get specific information, such as the number of generations, the best individual of a generation, etc.

# Create an experiment and run the genetic algorithm
exp <- fit(param, running_flag)
exp$generation_number()
exp$individual(99,0)
exp.results <- parseExperiment(exp)

Analyzing the results

# get the resulting attributes from all the evolve modesl
perf.df <- analyzeEvolution(exp, attributes = c("auc", "fit", "k", "n"), plot = FALSE)
analyzeEvolution(exp, attributes = c("auc", "fit", "k", "n"), plot = TRUE)
analyzeEvolutionAllModels(exp, attributes = c("auc", "fit", "k", "n"), plot = TRUE)
  

NB: you can additionally setup a GLogger object to add some display of info during ga() call

glog <- GLogger$new()

NB2: the glog object can only be created once. NB3: once created the verbosity level can be adjusted with : glog$set_level("warning") for instance.

NEW

• you may directly set a logger at a correct level glog$set_level("trace")
• you may use ga/ga2/ga_no_overfit/ga2_no_overfit sub algorithms of the ga family
• you may get all individuals at once in a a generation with pop$get_all_individuals(98) (which will give you a Dataframe of all the individuals)

GPU usage

You should have GPU set to true (directly in param.yaml or with param$set_bool("gpu",TRUE)). You may need additional libraries (Apple Silicons are natively supported by WGPU which gpredomics is using, Linux needs Vulkan libraries, with drivers, see gpredomics README.md for details).