automatic detection of multioutput datasets #1001
Conversation
|
Neat! Thank you for this PR. I think it should work for checking multi output dataset. But I checked this link, I think that not all regressors in our default config are supporting multi-output regression. I think we need a workaround for this issue. |
|
There are several ways to go about this:
|
|
Thank you for your ideas here. I prefer the 3rd one but I think a better solution based on it is to automatically use |
|
While working on it, i think i spotted a bug: Line 215 in aea42a5 This should be dep_import_str as the key and not import_str, right? It seems this also changes some tests. |
|
If i run the failing tests manually, they work. It has to be some unsuitable combination due to the random seed.. i'm not sure how to debug this. |
Yes, that is a bug. Hmm, I thought I fixed it a while ago but might not merge to master/dev branch. |
| @@ -501,6 +500,35 @@ def _fit_init(self): | |||
| self._last_optimized_pareto_front_n_gens = 0 | |||
| self._setup_config(self.config_dict) | |||
|
|
|||
| if multi_output_target: | |||
There was a problem hiding this comment.
Modifying _config_dict may not work in the situation that use use a customized configurations instead of default one. So, I think a practical way is to modify the _compile_to_sklearn function (here). If multi_output_target is True, then
sklearn_pipeline=MultiOutputClassifier(estimator=sklearn_pipeline) or sklearn_pipeline=MultiOutputRegessor(estimator=sklearn_pipeline) . I think it maybe a more general solution for multioutput dataset.
There was a problem hiding this comment.
This seems better, to be honest i didn't find a good place where to put my code and only settled on the _fit_init function therefore.
There was a problem hiding this comment.
Okay i looked into it, but the code would be a mess. Several functions would have to take multi_output_target as a new argument (most of them in export_utils.py), since they don't have access to the data or the TPOT Object.
Imo _fit_init seems to be the least intrusive point to include the checks
There was a problem hiding this comment.
Thank you for looking into. You are right. I think TPOT exported codes should also include MultiOutputRegessor/MultiOutputClassifier, which should change a lot of codes in TPOT. I will look into it when I get some time next week.
There was a problem hiding this comment.
@jhmenke I am sorry for overlooking this. I did not get a chance to look into this issue those days due to my busy schedule. I agree that TPOT need some major changes for including MultiOutputRegessor/MultiOutputClassifier. I may get some time in March to add those changes. You are welcome to push any changes meanwhile.
There was a problem hiding this comment.
Could we use my PR as a temporary fix until you have time to thoroughly refactor the code? I can prepare an update with the current development branch.
There was a problem hiding this comment.
Sorry for the delay. I think we can use it for a temporary solution with a minor release.
There was a problem hiding this comment.
Okay i merged the current master/development into this. Should be good to go as an interim solution then.
Fix a bug in the PR
Sorry, my mistake! I changed it back.
| if model in single_output_classifiers: | ||
| if 'sklearn.multioutput.MultiOutputClassifier' not in self._config_dict.keys(): | ||
| self._config_dict['sklearn.multioutput.MultiOutputClassifier'] = {"estimator": {}} | ||
| self._config_dict['sklearn.multioutput.MultiOutputClassifier']['estimator'][model] = self._config_dict[model] |
There was a problem hiding this comment.
There is only one sklearn.multioutput.MultiOutputClassifier in self._config_dict and estimator is keeping updated until the last model in single_output_classifiers so that the rest models should be removed.
|
Is this the closest pull to multi-output? any plans to finish it? |
sorry, no time for this right now.. feel free to work with my branch. i think the last todo is cloning the multioutputclassifier for every estimator that only supports single outputs as per weixuanfu's review. |
|
Would love to see multioutput regression/classification working with tpot! Keep going! Haha |
What does this PR do?
Slight variation of #903 depending on the shape of the target array.
What are the relevant issues?
#971