isotree.restore.handle: Unpack isolation forest model after de-serializing

In isotree: Isolation-Based Outlier Detection

Unpack isolation forest model after de-serializing

Description

After persisting an isolation forest model object through 'saveRDS', 'save', or restarting a session, the underlying C++ objects that constitute the isolation forest model and which live only on the C++ heap memory are not saved along, and depending on parameter 'lazy_serialization', might not get automatically restored after loading a saved model through 'readRDS' or 'load'.

The model object however keeps serialized versions of the C++ objects as raw bytes, from which the C++ objects can be reconstructed, and are done so automatically upon de-serialization when using 'lazy_serialization=TRUE', but otherwise, the C++ objects will only get de-serialized after calling 'predict', 'print', 'summary', or 'isotree.add.tree' on the freshly-loaded object from 'readRDS' or 'load'.

This function allows to automatically de-serialize the object ("complete" or "restore" the handle) without having to call any function that would do extra processing when one uses 'lazy_serialization=FALSE' (calling the function is not needed when using 'lazy_serialization=TRUE').

It is an analog to XGBoost's ‘xgb.Booster.complete' and CatBoost’s 'catboost.restore_handle' functions.

If the model was buit with 'lazy_serialization=TRUE', this function will not do anything to the object.

Usage

isotree.restore.handle(model)

Arguments

model

An Isolation Forest object as returned by 'isolation.forest', which has been just loaded from a disk file through 'readRDS', 'load', or a session restart, and which was constructed with 'lazy_serialization=FALSE'.

Details

If using this function to de-serialize a model in a production system, one might want to delete the serialized bytes inside the object afterwards in order to free up memory. These are under 'model$cpp_objects$(model,imputer,indexer)$ser' - e.g.: 'model$cpp_objects$model$ser = NULL; gc()'.

Value

The same model object that was passed as input. Object is modified in-place however, so it does not need to be re-assigned.

Examples

### Warning: this example will generate a temporary .Rds
### file in your temp folder, and will then delete it

### First, create a model from random data
library(isotree)
set.seed(1)
X <- matrix(rnorm(100), nrow = 20)
iso <- isolation.forest(X, ntrees=10, nthreads=1, lazy_serialization=FALSE)

### Now serialize the model
temp_file <- file.path(tempdir(), "iso.Rds")
saveRDS(iso, temp_file)
iso2 <- readRDS(temp_file)
file.remove(temp_file)

cat("Model pointer after loading is this: \n")
print(iso2$cpp_objects$model$ptr)

### now unpack it
isotree.restore.handle(iso2)

cat("Model pointer after unpacking is this: \n")
print(iso2$cpp_objects$model$ptr)

### Note that this function is not needed when using lazy_serialization=TRUE
iso_lazy <- isolation.forest(X, ntrees=10, nthreads=1, lazy_serialization=TRUE)
temp_file_lazy <- file.path(tempdir(), "iso_lazy.Rds")
saveRDS(iso_lazy, temp_file_lazy)
iso_lazy2 <- readRDS(temp_file_lazy)
file.remove(temp_file_lazy)
cat("Model pointer after unpacking lazy-serialized: \n")
print(iso_lazy2$cpp_objects$model$ptr)

isotree documentation built on May 29, 2024, 11:24 a.m.