export_graphviz should work with a sklearn.tree._tree.Tree

[jnothman]

export_graphviz currently has code that checks if it's operating on a Tree instance or an estimator with a tree_ attribute, but also has code that forces it to only operate on an estimator. One or the other of these should change. Either we force it to be an estimator, or we test that it's okay to be a Tree!

[mabelvj]

Could I try this? Whenever they decide what's the best option.

[jnothman]

I don't have a strong preference for either option. Neither will break backwards compatibility and we could always add Tree support later. currently all our ensembles are ensembles of tree classifiers/regressors, I think, not of trees themselves. @glemaitre might have an opinion as the incumbent tree expert.

[glemaitre]

I would say estimator. However, I need to go trough #1046 to understand why the code is currently as it is. I keep inform soon.

[glemaitre]

estimators_ in GradienBoosting are DecisionTreeRegressor and not Tree anymore. So IMO, there is needs to support anymore Tree directly. I would agree with @amueller comment mentioning that tree_.Tree is an internal thing and that we could keep the user away from it, now that this is possible.

[jnothman]

_tree.Tree is not quite an internal thing, IMO. It's needed in some applications, and it's accessible through a public attribute. Still, I'm happy to not support it here.

[amitesh197]

Thank you everyone