Dynamically wrap fit in the CallbackSupportMixin to initialize the callback context #20
Conversation
jeremiedbb
left a comment
jeremiedbb
left a comment
There was a problem hiding this comment.
Let's try another alternative in a new PR, that is an extension of this one, where fit accepts a callback_context arg. We assume that fit has the arg and the decorator initializes the context and passes it to fit.
sklearn/callback/_mixin.py
Outdated
|
|
||
| def __init_subclass__(cls, **kwargs): | ||
| super().__init_subclass__(**kwargs) | ||
| if hasattr(cls, "fit"): |
There was a problem hiding this comment.
Not sure this defensive condition is useful. If one adds the mixin to a non-estimator class, it's his problem :)
There was a problem hiding this comment.
Yes, I was anticipating cases where we might have callbacks on transform, so we would just add another if statement with a _transform_callback decorator, to tackle classes with transform but no fit or vice versa.
sklearn/callback/_mixin.py
Outdated
| def __init_subclass__(cls, **kwargs): | ||
| super().__init_subclass__(**kwargs) | ||
| if hasattr(cls, "fit"): | ||
| cls.fit = _fit_callback(cls.fit) |
There was a problem hiding this comment.
We'll probably need to wrap all methods that fit, i.e. fit, fit_transform, fit_predict, partial_fit.
Although I'm not sure what will be the behavior then in a case like fit calling fit_transform internally.
sklearn/callback/_mixin.py
Outdated
| means the maximum number of subtasks is not known in advance. | ||
| @functools.wraps(fit_method) | ||
| def wrapper(estimator, *args, **kwargs): | ||
| estimator._callback_fit_ctx = CallbackContext._from_estimator(estimator) |
There was a problem hiding this comment.
To be safe w.r.t. my previous comment we can check first if the attribute doesn't already exist
Co-authored-by: Jérémie du Boisberranger <jeremie@probabl.ai>
Co-authored-by: Jérémie du Boisberranger <jeremie@probabl.ai>
…finally for eval_on_fit_end in _fit_context
Co-authored-by: Jérémie du Boisberranger <jeremie@probabl.ai>
… the same wrapper
…it-learn into mixin_wrapper_for_callbacks
…s merged with its parent
2 participants
The callback context relative to the
fitis intialized in a decorator aroundfit, and thefitfunction is dynamically decorated through__init_subclass__in theCallbackSupportMixinclass.