So far I circumvented this issue by copying the head model in the folder containing the data I wanted to source localise.
This is actually what you need to do.
The high-level functions do not take directly the input files as arguments, they get the files from the database structure. This process "process_inverse_2018" takes recordings in input (argument sFiles, which contains pointers to "data" files) and then gets the default head model for these files (= the one displayed in green in the same folder as the channel file corresponding to each of the "data" files in sFiles).
Does it make sense?