如何在 sklearn 中使用自定义估计器进行交叉验证？

Question

I have written a custom estimator class with a fit and transform method. I am able to create a model, train and predict using the model.

However, while doing cross-validation, I run into this error: TypeError: cannot deepcopy this pattern object .

This is how CustomEstimator looks like:

class DefaultEstimator(BaseEstimator, TransformerMixin):
    def __init__(self, preprocessor, pipelines):
      self.preprocessor = preprocessor
      self.pipelines = pipelines

    def fit(self, X, y=None):
      for each_pipeline in self.pipelines:
          each_pipeline.fit(self.preprocessor.apply(X), y)
      return self

   def transform(self, X):
     transformed_data = []
     for each_pipeline in self.pipelines:
        transformed_data.append(each_pipeline.transform(self.preprocessor.apply(X)))
     return sp.hstack(transformed_data)

Does anyone have an idea on approaching this issue?

Answer 1

I would suggest having the preprocessor inside the pipeline itself. Cross_val_score would try to copy the params of the estimator, it would break when the estimator cannot return the params while calling get_params() .

I am not sure whether your pipeline parameter is a Sklearn pipeline because the pipeline object is not iterable.

Answer 2

As suggested in few comments, this error is because self.processor can't be deep-cloned.

So, the workaround for this error is to remove preprocessing step from this class and move it as independent preprocessing step or inside the pipeline itself.