[ENH] parallelization backends for grid and random search #150
Conversation
| Directly passed to ``utils.parallel.parallelize``. | ||
| Valid keys depend on the value of ``backend``: | ||
| - "None": no additional parameters, ``backend_params`` is ignored |
There was a problem hiding this comment.
I read through the backend_params multiple times and I do not understand how they work. How does the key and value of the dict look like?
We also have two parameters "backend" and "backend_params", that could easily be done in one parameter.
There was a problem hiding this comment.
backend selects which parellization backend you want to use, e.g., "joblib" or `"ray".
backend_params is a config dict for the backend, which keys you can pass depends on the chosen backend.
There was a problem hiding this comment.
I updated the docstrings, hope that is clearer now?
There was a problem hiding this comment.
How does the key and value of the dict look like? Does the dict have multiple keys? Is the value a list of the strings shown in the docstring text or just one string? Could you write an example into the docstring?
There was a problem hiding this comment.
the params dict can be a dict with multiple keys, I will add an example
fkiraly
changed the title
SimonBlanke
left a comment
SimonBlanke
left a comment
There was a problem hiding this comment.
fix tests before merge can be done
From Original PR #150: This PR introduces parallelization backends for GridSearch and RandomSearch. Both estimators now allow to specify backend and backend_params in the constructor, which allows selection of a parallelization backend and configuration parameters for it - the default being "None", i.e., plain loop. This uses the parallel utilities also used in sktime and skpro, with a mid-term plan to move these to scikit-base. --------- Co-authored-by: Franz Király <fkiraly@gcos.ai>
2 participants
This PR introduces parallelization backends for
GridSearchandRandomSearch.Both estimators now allow to specify
backendandbackend_paramsin the constructor, which allows selection of a parallelization backend and configuration parameters for it - the default being"None", i.e., plain loop.This uses the
parallelutilities also used insktimeandskpro, with a mid-term plan to move these toscikit-base.