passing an extra argument to GenericUnivariateSelect without scope tricks

问题

EDIT:

here is the complete traceback if I apply the make_scorer workaround suggested in the answers...

`File "________python/anaconda-2.7.11-64/lib/python2.7/site-packages/spyder/utils/site/sitecustomize.py", line 880, in runfile
    execfile(filename, namespace)

  File ""________python/anaconda-2.7.11-64/lib/python2.7/site-packages/spyder/utils/site/sitecustomize.py", line 94, in execfile
    builtins.execfile(filename, *where)

  File ""________/main_"________.py", line 43, in <module>
    "_________index.fit(X,Y ,g=g,L=L)

  File ""________/Core.py", line 95, in fit
    X_preprocessed=self.preprocessing.fit_transform(X,y)

  File ""________python/anaconda-2.7.11-64/lib/python2.7/site-packages/sklearn/pipeline.py", line 303, in fit_transform
    return last_step.fit_transform(Xt, y, **fit_params)

  File ""________/python/anaconda-2.7.11-64/lib/python2.7/site-packages/sklearn/base.py", line 497, in fit_transform
    return self.fit(X, y, **fit_params).transform(X)

  File "Base/Base.py", "________
    score_func_ret = self.score_func(X, y)

TypeError: __call__() takes at least 4 arguments (3 given)`

I am working on a sklearn pipeline.

custom_filter=GenericUnivariateSelect(Custom_Score,mode='MinScore',param=0.9)   
custom_filter._selection_modes.update({'MinScore': SelectMinScore})
MyProcessingPipeline=Pipeline(steps=[...
                           ('filter_step', None),
                           ....])
ProcessingParams.update({'filter_step':custom_filter})
MyProcessingPipeline.set_params(**ProcessingParams)

where SelectMinScore is a custom BaseFilter.

I need to perform univariate feature selection based on a Custom_Score, which must receive an extra argument, called XX in here

def Custom_Score(X,Y,XX=_XX ):
      # do stuff
      return my_score

Unfortunately, AFAIK the sklearn API does not allow extra arguments to be passed a parameter of a parameter of a step of a pipeline.

I have tried

MyProcessingPipeline({'filter_step':custom_filter(XX=_XX)})

but that breaks argument passing cascade (I believe).

So far, I have solved this by writing a wrapper, where _XX is the data I need which unfortunately need to be in the scope of the function at definition time. So I have ended up defining the function within my main function so that _XX exists and it can be passed.

def Custom_Score_Wrapped(X,Y):
            return Custom_Score(X,Y,XX=_XX )

I think this is a really dirty workaround.

What is the right way to do this?

回答1:

You can simply pass the extra arguement while calling the make_scorer() function. For example, you check this link. In the example it makes use of fbeta_score. Now fbeta requires an additional parameter, beta which is set while calling the make_scorer() function like this :

ftwo_scorer = make_scorer(fbeta_score, beta=2)

So in your case, this should work:

def Custom_Score(X,Y,XX):
  # do stuff
  return my_score

my_scorer = make_scorer(Custom_Score,XX=_XX)

来源：https://stackoverflow.com/questions/46559634/passing-an-extra-argument-to-genericunivariateselect-without-scope-tricks

标签

python

scikit-learn

pipeline

scoping