Hyperparameter tuning for CausalForestDML #390
Conversation
…ith sklearn logic for future easier extendibility.
vsyrgkanis
changed the title
econml/dml/causal_forest.py
Outdated
| out of sample R-score. After the function is called, then all parameters of `self` have been | ||
| set to the optimal hyperparameters found. |
There was a problem hiding this comment.
Maybe explicitly mention that although the parameters will have been updated, this estimator will not have been fit with this data.
There was a problem hiding this comment.
now only final stage params are tunable
There was a problem hiding this comment.
I still think it's worth mentioning that this estimator has not been fit with this data, so that the user knows that the they still have to call fit afterward.
There was a problem hiding this comment.
Or alternatively, in your notebooks I think you always use:
est.tune(...)
est.fit(...)Are there times a user would not want to do that, or should calling fit just be folded into tune so that it saves the user the trouble?
There was a problem hiding this comment.
That's what I'm doing in the notebooks I think.
I think it's good to have separate. Maybe someone wants to tune on a subset of the data or some small chunk. Also the tune does not need to take other keyword args like inference, cache_values etc.
There was a problem hiding this comment.
I added the extra docstring comment that the returned self is not yet fitted.
No description provided.