Open eddiebergman opened 2 months ago
As an additional thing to fix, this function calls read_tasks_and_dev_stages_from_disk
which essentially reads all results from disk, uses just the ids it retrieves and the proceeds to call anther function summarize_results
which will then load all results again.
Last point, nothing in NePS is using this function and there's no publicised user API for this. I guess we'll need to rediscover the point of this feature since it's pretty undocumented anywhere.
My guess it's some sort of convinience feature but I don't feel this provides much benefit and perhaps more confusion than if someone just made a different working_directory
More issues with this function, pretty sure this would just not work if subdir was given.
What even is a user prior dir... what's meant to be in there that's different from any other run directory?
Not really sure how the whole task and dev id's really work but from what I gather from this function the path goes something like this:
sub_dir
:workingdir/seed/sub_dir/task_id/dev_id
.workingdir/seed/task_id/dev_id
However this function only takes it over the first valid seed dir found, with no knowledge of what this is.
https://github.com/automl/neps/blob/dbc307697168a4b2099bc117e200e42ab36c2fd9/neps/utils/data_loading.py#L223-L251
I guess the way to go (if seed dir is still relevant) is to specifically ask for what seed is requested in the function params.