I think you can uncomment "chi2": [{"gamma": 1.}],.
Your dictionary is somehow the same as here. Could you use only one dictionary?

Dear Tom, thanks for you comments.
You are absolutely right, I can uncomment "chi2": [{"gamma": 1.}]. I will do so.
The dictionary you are referring to has no default parameters, only parameter names. Do you think there should be one global dictionary with kernel names, parameter names and default values?

You could reuse your dictionary (with default values) where the old one (without default values) was used (here).

>>>"gamma" in frozenset(["gamma", "degree", "coef0"])
True
>>>"gamma" in {"gamma": None, "degree": 3, "coef0": 1}
True

Sorry for the delay. I just implemented the changes: I added the default kernel parameters to the dictionary in pairwise.py, removed the dictionary in kernel_approximation.py and import the one in pairwise.py.

This looks good to me.
Could you also add a small non-regression test?
Is it what you had in mind @amueller ?

Should the test address anything else then calling the Nystroem approximation without specifying the gamma parameter for the rbf and chi^2 kernel (this is where Travis was failing).
Should I add the test to sklearn.tests.test_kernel_approximation.py?

I have the feeling that the non-regression test is already covered here. Lines 164 - 169 call the kernels with default parameters:

# test that available kernels fit and transform
kernels_available = kernel_metrics()
for kern in kernels_available:
    trans = Nystroem(n_components=2, kernel=kern, random_state=rnd)
    X_transformed = trans.fit(X).transform(X)
    assert_equal(X_transformed.shape, (X.shape[0], 2))

This test failed, because it called the chi2 kernel with the default gamma value for the rbf kernel (gamma:None). Now the default gamma for chi2 (1.) is taken from the dictionary.

Please let me know if I should add another test.

Thanks,

Fernando

Not entirely the change I had in mind, and KERNEL_PARAMS is public, right? So maybe we shouldn't change it?
A good test would be to run RBFSampler with different chi2 and different gamma and see that the results actually change (in the same way as when rescaling the data). That wasn't the case before, right?

The fix I suggested was passing a dict of arguments in the RBFSampler.__init__ and deprecating the other parameters.
You just introduced "another" place where defaults are kept and they need to be in sync with the defaults elsewhere. We could add tests that ensure consistency but I feel this is asking for trouble.

[sorry for prolonged absence]

Closing as the original issue #5229 was resolved in #9166