Custom multiclass objectives in R

[tvatter]

Hi everyone,

I am wondering whether it is possible to create custom multiclass objectives for xgboost in R.

Below is an MWE:

rm(list=ls())
require(xgboost)
set.seed(0)

# Simulate some dummy data
n <- 1e2
num_features <- 2
num_class <- 3
y <- apply(rmultinom(n, 1, rep(1, num_class)), 2, function(yy) which(yy != 0))
x <- matrix(rnorm(n*num_features), n, num_features)

# Format for xgboost
dtrain <- xgb.DMatrix(x, label = y)

# A dummy custom objective to understand what is going on
dummy_obj <- function(preds, dtrain) {
  browser()
}

# Feed to xgboost
params_dummy <- list(objective = dummy_obj, 
                     num_class = num_class)
model_dummy <- xgb.train(params_dummy, data = dtrain, nrounds = 1)

On my computer, executing the code above allowed me to do:

Called from: obj(pred, dtrain)
Browse[1]> preds
  [1] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 [40] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 [79] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Browse[1]> Q

Having read the source code of multiclass_obj.cc, I was expecting for preds a vector of length n x num_class, or a matrix with n rows and num_class columns, or something like that.

So, my question is: am I missing something obvious (very likely) or is it currently not possible to implement custom multiclass objectives?

Below are my environment informations, although I doubt they will be useful in this case.

---

Operating System:

OS X 10.12.6

Compiler:

Not relevant

Package used (python/R/jvm/C++):

R

sessionInfo():

R version 3.4.1 (2017-06-30)
Platform: x86_64-apple-darwin16.6.0 (64-bit)
Running under: macOS Sierra 10.12.6

Matrix products: default
BLAS: /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib
LAPACK: /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libLAPACK.dylib

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats graphics grDevices utils datasets methods base

other attached packages:
[1] xgboost_0.6-4 tvmisc_0.1.0

loaded via a namespace (and not attached):
[1] compiler_3.4.1 magrittr_1.5 Matrix_1.2-11 tools_3.4.1 rstudioapi_0.6
[6] stringi_1.1.5 grid_3.4.1 data.table_1.10.4 lattice_0.20-35

[khotilov]

There's an issue with multiclass case in this respect, because when objective was not explicitly set when num_class > 1 then it is assigned to be "multi:softmax"

xgboost/src/learner.cc

Line 224 in c09ad42

cfg_["objective"] = "multi:softmax";

And the "multi:softmax" transforms predictions into an ndata-sized vector:

xgboost/src/objective/multiclass_obj.cc

Line 96 in c09ad42

if (!prob) tmp.resize(ndata);

I'll think about how to handle this better. But as a workaround, you may rebuild your R package after adding the outputmargin = TRUE parameter to the predict call in

xgboost/R-package/R/utils.R

Line 135 in c09ad42

pred <- predict(booster_handle, dtrain)

Then you should be able to work with untransformed predictions within a custom objective.