(ping @pprett, @glouppe)

Travis disagrees ;)

Thanks ! I forget to check that testing file before pushing.

I fix the test. GBRT is indeed mi-using the api (use predict function instead of decision_score) with its dummy classifier. It's why DummyClassifier could not be used as a drop in replacement.

@arjoly In general I'm -1 on this -- I see that code duplication is an argument but having little dependencies is also something that is valuable. Now any change in DummyRegressor needs to be checked against the gradient boosting code.

The main reason why those initial estimators in the gradient boosting modules are so quirky is because they are internal components of the gradient boosting module where I can make assumptions about inputs etc. I optimized the GBRT code for single datapoint predictions and one of the major performance killers there is redundant input checking (e.g. check_arrays is super costly).

I see that I made some poor design decisions here but I think the init feature is so infrequently used that I dont want to break pickle compatilibity just to iron out an infrequently used feature -- I would rather deprecate init.

The main reason why those initial estimators in the gradient boosting modules are so quirky is because they are internal components of the gradient boosting module where I can make assumptions about inputs etc. I optimized the GBRT code for single datapoint predictions and one of the major performance killers there is redundant input checking (e.g. check_arrays is super costly).

Then it tells me that all these estimator should be private (starting with _) with comments explaining all the gotchas.

I see that I made some poor design decisions here but I think the init feature is so infrequently used that I dont want to break pickle compatilibity just to iron out an infrequently used feature -- I would rather deprecate init.

There is an issue that complains about not being able to use init for a starting estimate. Though, I think it's a good idea to deprecate this feature.

At the moment, I find the implementation complex to understand.

@pprett Ok, then I am going to close this pr. I definitely think that the gbrt code should be simpler with more private things.