The CI is failing because we are following NumPyDoc standards which require the "summary line" to fit one line.

Ah, I see. Thank you. :)

There isn't many requirements on the last step. For example, one can have a stateless transformers in a pipeline:

from sklearn.pipeline import make_pipeline
import numpy as np

class DoubleIt:
    def transform(self, X, y=None):
        return 2*X

X = np.array([[1, 2, 3], [4, 5, 6]])
p = make_pipeline(DoubleIt(), DoubleIt())

p.transform(X)
# array([[ 4,  8, 12],
#       [16, 20, 24]])

As long as the methods you call on the pipeline exist in the final step, then everything still works.

Thanks @thomasjpfan, I didn't think of that... I will change the wording to also allow for only stateless transformers. :)

But look at this:

from sklearn.pipeline import Pipeline
from sklearn.linear_model import LinearRegression
import numpy as np

class DoubleIt:

    def transform(self, X, y=None):
        return 2*X

X = np.array([[1, 2, 3], [4, 5, 6]])
p = Pipeline([
            ('double1', DoubleIt()),
            ('double2', DoubleIt()),
            ('linreg', LinearRegression())     # same results with ('linreg', None), but not with "passthrough"
            ])

print(p.fit(X))

Error:
TypeError: All intermediate steps should be transformers and implement fit and transform or be the string 'passthrough' '<__main__.DoubleIt object at 0x7fbed9d1f310>' (type <class '__main__.DoubleIt'>) doesn't

Whenever the one of the steps implements a fit, the previous steps are expected to be stateful transformers, too. This is because def _validate_steps() only checks the steps if .fit() is called.

Should _validate_steps() be modified to be less strict about stateless transformers? And should it then also run on transform(), and all the other possible methods?

PS: I just opened an issue for this (#27014). :)

Given where the discussion in #27014 is headed, I think we need to enforce that pipeline.fit works. According to our estimator docs, an estimator must define fit. For pipeline.fit to work:

• All steps must define fit.
• All non-last steps must also define transform.

I like the new paragraph. It is easier to read and is still correct. Getting this right is not easy, at least I had to go and read the definitions of estimator, predictor, etc and then think to figure out if it was correct or not!

I find the start of the first sentence hard to read. What do you think of starting with the end "To build a composite estimator ...."?

Another question: would it be correct to say "... combined with classifiers, regressors, other predictors or estimators to build a composite estimator"? The reason I'm asking is that if I want to make a composite estimator that is a transformer, then I don't need a predictor as the last step no? The model I'm starting to form in my head is that the type of the composite estimator is equal to the type of the final step in the pipeline. Is that useful as a mental model?

@betatim Thanks for your review. Your thoughts helped me to clarify how to communicate this. I've changed how to start the first sentence and found some words to express, that Pipeline would also combine two transformers, and what the type of the pipeline then would be.

This is the result:

To build a composite estimator, transformers are usually combined with other
transformers or with :term:`predictors` (such as classifiers or regressors).
The most common tool used for this is a :ref:`Pipeline <pipeline>`. Pipelines
without a final predictor will fit and transform the observed data (:term:`X`)
and return it. In contrast, pipelines with a final predictor will fit the
predictor with the transformed data from the previous steps and can then also
be used for predictions. The methods to be called on the pipeline are the same
methods that the last contained estimator provides.

What do you think?

I like it! (sorry, I have nothing more to say :D)

this can also be rephrased. Maybe like:

I will suggest is mixture of both ;)

Edit: Because (name, estimator) is unclear. It doesn't show, that users can pick name themselves.

Same here, and note that the last step doesn't have to implement fit_transform.

Yes, this is why you can only call fit_transform on the pipeline, if the last step implements fit_transform.
The phrase "Only valid if ..." is used on all the other 8 methods that are exposed after the pipeline is fitted. I copied it here, because to me it makes sense.

Here fit and transform of all the estimators (including the last one) are both used on the data.

"transformers sequentially to the input data" could be misunderstood though. I will suggest an updated sentence. :)

that's not true, you can call fit_transform, and it will call .fit.transform if fit_transform doesn't exist.

Ah, now I see. Should we put something like:
Only valid if the final estimator either implements fit_transform or fit and transform.
?