NaN predictions about agtboost

[Blunde1]

@ldesreumaux Thank you for raising the issue!

I am not able to reproduce the result that you are showing.
With a fresh install from the developer version together with your code

devtools::install_github("Blunde1/agtboost/R-package")
library(agtboost)
data(caravan.train, package = "agtboost")
train <- caravan.train
gbt.train(train$y, train$x, loss_function="logloss", verbose=10, nrounds=500, learning_rate=0.1)

I got the following output

it: 1  |  n-leaves: 2  |  tr loss: 0.2143  |  gen loss: 0.2145
it: 10  |  n-leaves: 2  |  tr loss: 0.1969  |  gen loss: 0.201
it: 20  |  n-leaves: 2  |  tr loss: 0.1913  |  gen loss: 0.1981
it: 30  |  n-leaves: 2  |  tr loss: 0.1887  |  gen loss: 0.1977
C++ object <000001e8ce835610> of class 'ENSEMBLE' <000001e8cdeae5a0>

I obtained almost identical results at consecutive runs.
May I ask which system/compiler you are using, or if you have worked on the internal C++ code?
The tree-depths you obtain are vastly different from mine, so something strange is deffinitely going on here.

[ldesreumaux]

Indeed, the issue cannot be reproduced with the development version. I can only reproduce it when agtboost is installed with:

install.packages("agtboost")

And it does not seem to be platform-dependent: I reproduced it on Windows 10 and openSUSE 12.3 with R 4.0.5.

[Blunde1]

Thank you, I was now able to reproduce it on R 4.0.5 (2021-03-31). Strangely CRAN and development version should be the same. I will look into it.

[Blunde1]

Hoping this will be fixed in v0.9.2.
Waiting for CRAN mirrors to be updated to test with docker