Open simonpcouch opened 3 months ago
Okay, a few more notes here.
It seems like the architecture here is:
1) tunable()
gets called on a boost_tree()
model spec (or other model type)
2) tunable.boost_tree()
kicks in
2a) it first calls tunable.model_spec()
,
...which only excludes rows that are both main arguments and have no tunable information associated, so there's a bunch of engine arguments with NULL call_info
2b) then, it adds engine-specific parameter information. One way could be add_engine_parameters()
:
xgboost_engine_args
is an inlined tibble in the source:
...the other possible way is to manually insert call_info
entries:
Barring a rewrite that stores tunable information in the model environment (see #826), I think our solution might be to filter out rows with NULL call_info
at the end of existing non-model_spec
tunable()
methods.
Created on 2024-04-04 with reprex v2.1.0