I added this test case that catches the error:

def test_pipeline_param_error():
    clf = Pipeline(memory=None, steps=[('wrong_step_name', LogisticRegression())])
    assert_raise_message(TypeError,
                         "Use naming convention step__parameter.",
                         clf.fit, [[0], [0]], [0, 1],
                         logisticregression__sample_weight=[1, 1]
                         )

@okz12 as described in the original issue #13534 (comment) the problem we are trying to address here is that if a user passes sample_weight to a pipeline as-is, e.g.

pipe.fit([[0], [0]], [0, 1], sample_weight=[1, 1])

they end up with an error message that is not clear at all (Not enough values to unpack...). We want this error message to become something along the lines of

Passing sample_weight directly to a pipeline isn't supported. You can however pass 
sample_weight to specific steps of your pipeline, e.g. 
`pipe.fit(X, y, logisticregression__sample_weight=sample_weight)`.

I've added the error message and an accompanying test. I am checking on whether two underscores '__' are within the variable name, and this would catch other parameters besides sample_weight. I made the error generic referring to parameters rather than sample_weight in particular but kept the sample_weight parameter passing example. Let me know if this should be changed.

Also you might want to change the title to [MRG]

Not sure if we need to worry about changing a ValueError into a TypeError.

Or actually, whether we must be consistent and also raise a TypeError in the case of any other parameter names (including __)

I found examples in gaussian_process/kernels.py and base.py where it splits on '__' and raises a ValueError if the parameter name is invalid. Not sure if I should change pipeline.py to ValueError or the other two to TypeError.

Hm, I guess the most consistent way would be to raise a ValueError here instead of a TypeError?

We raise ValueError when invalid parameters are passed to the estimators, and that's what this check is about

Changing to ValueError breaks the test for fix #13472. The error message picked up is from my fix rather than the fix in #13472, which causes the error message to mismatch and the test-case to fail. I'm not entirely sure of the implications of this. Is it good that this fix captures both errors, in which case I can update the test case for gradient boosting.

Test Case for #13472:

def test_gradient_boosting_with_init_pipeline():
    # Check that the init estimator can be a pipeline (see issue #13466)

    X, y = make_regression(random_state=0)
    init = make_pipeline(LinearRegression())
    gb = GradientBoostingRegressor(init=init)
    gb.fit(X, y)  # pipeline without sample_weight works fine

    with pytest.raises(
            ValueError,
            match='The initial estimator Pipeline does not support sample '
                  'weights'):
        gb.fit(X, y, sample_weight=np.ones(X.shape[0]))

Output Error:

E           AssertionError: Pattern 'The initial estimator Pipeline does not support sample weights' not found in 'Pipeline.fit does not accept the sample_weight parameter. You can pass parameters to specific steps of your pipeline using the stepname__parameter format, e.g. `Pipeline.fit(X, y, logisticregression__sample_weight=sample_weight)`.'

sklearn/ensemble/tests/test_gradient_boosting.py:1396: AssertionError

Yes, IMO, change the gradient boosting check to conform to this.

You need to not modify the gradient boosting test, but the gradient boosting code, to check for the right pipeline failure error message here:

Merged, thanks @okz12 !