Hyperparameter GridSearch with CausalForestDML

[jaronowitz]

Is there a way to use scikit learn's GridSearchCV with CausalForestDML? In the EconML user guide under How do I select the hyperparameters of the final model (if any)? is an example for this with NonParamDML but CausalForestDML does not have a model_final parameter. I'd like to fine tune n_estimators, max_depth, etc parameters of CausalForestDML.

When I try

est = CausalForestDML(inference='bootstrap', cv=5, 
                          model_t=WeightedMultiTaskLassoCV(), 
                          model_y=WeightedLassoCV())
    
est_dict[t] = GridSearchCV(estimator=est, param_grid=cv_parameters)
                        
col_list = make_col_list_polynomial_3(t)
T = T_df[col_list]
    
est_dict[t].fit(Y=y, T=T, X=X, W=W)

I get the following error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<timed exec> in <module>()

/opt/conda/anaconda/lib/python3.6/site-packages/sklearn/utils/validation.py in inner_f(*args, **kwargs)
     61             extra_args = len(args) - len(all_args)
     62             if extra_args <= 0:
---> 63                 return f(*args, **kwargs)
     64 
     65             # extra_args > 0

/opt/conda/anaconda/lib/python3.6/site-packages/sklearn/model_selection/_search.py in fit(self, X, y, groups, **fit_params)
    763         n_splits = cv_orig.get_n_splits(X, y, groups)
    764 
--> 765         base_estimator = clone(self.estimator)
    766 
    767         parallel = Parallel(n_jobs=self.n_jobs,

/opt/conda/anaconda/lib/python3.6/site-packages/sklearn/utils/validation.py in inner_f(*args, **kwargs)
     61             extra_args = len(args) - len(all_args)
     62             if extra_args <= 0:
---> 63                 return f(*args, **kwargs)
     64 
     65             # extra_args > 0

/opt/conda/anaconda/lib/python3.6/site-packages/sklearn/base.py in clone(estimator, safe)
     69                                 "estimator as it does not implement a "
     70                                 "'get_params' method."
---> 71                                 % (repr(estimator), type(estimator)))
     72 
     73     klass = estimator.__class__

TypeError: Cannot clone object '<econml.dml.causal_forest.CausalForestDML object at 0x7f6c1284cd68>' (type <class 'econml.dml.causal_forest.CausalForestDML'>): it does not seem to be a scikit-learn estimator as it does not implement a 'get_params' method.

[vsyrgkanis]

You cant use gcv itself but in the 0.9 release the CausalForestDML class will have a “tune” method that internally tunes parameters and you can define whatever grid of params you want similar to gcv.

[jaronowitz]

Thanks for the reply! Sounds like there is a lot of great new features in 0.9. Bring it on :)

[vsyrgkanis]

If you want to use it you can install the vasilis/grf_and_scorer_docs
Branch of the git repo which has not yet been merged to master. i.e.

pip install git+https://github.com/microsoft/EconML.git@vasilis/grf_and_scorer_docs

[jaronowitz]

Hi @vsyrgkanis I was checking out the methods for CausalForrestDML (https://econml.azurewebsites.net/_autosummary/econml.dml.CausalForestDML.html) now that .9 is released. Is the tune method included in this release? Thanks for the help

[vsyrgkanis]

Unfortunately, it didn't get in the base 0.9 release. It'll get in, in an upcoming point release maybe next week. See here for the PR:
#390

[vsyrgkanis]

This is now addressed by #390 and will be included in the 0.9.2 release

[vsyrgkanis]

0.9.2 has been released on pypi and hyperparam tuning is available