Hi!
Thanks for beautifull package mlr3torch, it's awesome!

If you train 'classif.ft_transformer' on data that has categorical variables - then after saving (with marshal() before) trained learner to a file and after reading (with unmarshal() after) from the file when executing predict you will see the error 'Error in (function (self) : external pointer is not valid'

If you train 'classif.ft_transformer' on data that has no categorical variables - then the error does not occur.

Code to reproduce the error:
https://gist.github.com/jurbanhost/f01388430718610661a239518bc6d696

Jury.