[MRG] FIX Nystroem with precomputed kernel

[venkyyuvy]

Reference Issues/PRs

Fixes #14641

What does this implement/fix? Explain your changes.

Added

• key precomputed in KERNEL_PARAMS
• non-regression test to validate Nystroem with precomputed kernel using polynomial_kernel

Any other comments?

[TomDLT]

LGTM

Please also add a bugfix entry in doc/whats_new/0.22.rst.

[amueller]

can you add a whatsnew?

[amueller]

actually, not sure if this is the right fix, maybe _get_kernel_params in Nystroem should be fixed.

I was wrooong!

[venkyyuvy]

can you add a whatsnew?

Now we can use the Nystroem with precomputed kernel matrix. Previous it was giving KeyError: 'precomputed'

[glemaitre]

actually, not sure if this is the right fix, maybe _get_kernel_params in Nystroem should be fixed.

@amueller
Do you mean that it makes more sense for _get_kernel_params() to return an empty tuple instead of registering 'precomputed' as a kernel? Is it because 'precomputed' kernel would not be a kernel?

        if self.kernel == 'precomputed':
            return params
        if not callable(self.kernel):
            for param in (KERNEL_PARAMS[self.kernel]):
                if getattr(self, param) is not None:
                    params[param] = getattr(self, param)
        else:
            if (self.gamma is not None or
                    self.coef0 is not None or
                    self.degree is not None):
                raise ValueError("Don't pass gamma, coef0 or degree to "
                                 "Nystroem if using a callable kernel.")

[TomDLT]

Right, if someone wants to iterate over KERNEL_PARAMS' keys, "precomputed" does not make sense. Then your fix would be better.

[venkyyuvy]

Do you mean that it makes more sense for _get_kernel_params() to return an empty tuple instead of registering 'precomputed' as a kernel? Is it because 'precomputed' kernel would not be a kernel?

@glemaitre

Please correct me If I am wrong,
Aren't we supposed to check, whether

(self.gamma is not None or
   self.coef0 is not None or
   self.degree is not None)

when kernel='precomputed' then raise a ValueError

"Don't pass gamma, coef0 or degree to Nystroem if using precomputed kernel."

On a side note, I saw here, we have been using this inside PAIRWISE_DISTANCE_FUNCTIONS .

        'precomputed': None,  # HACK: precomputed is always allowed, never called

That is the reason, why I had pursued ahead to add precomputed in KERNEL_PARAMS

[amueller]

I feel if we can avoid the hack that's better ;)

[adrinjalali]

Please also test that the correct error is raised with a precomputed kernel and passed kernel parameters.

[venkyyuvy]

Thanks a lot @adrinjalali for your valuable comments. Please review my latest commit.

[adrinjalali]

LGTM, thanks @venkyyuvy

[TomDLT]

Nice, thanks !