HLP trying to get custom kernels to work with sparse SVM

[amueller]

This segfaults atm :-/
Also, needs more testing.

Not really on my todo list, could be a basis for someone to work on this.

[fabianp]

looks good to me.This adds support for custom kernels in the sparse case and I don't see any side-effects.

Merging ...

[fabianp]

ok to merge @amueller (I just noticed the WIP the title)?

[amueller]

@fabianp No, sorry, it is not. As I mentioned in the description above, this produces random segfaults and needs more work. See #918.
One needs to get the number of support vectors, for which there is no function yet, to copy the dual coefficients etc.

This is not really on my todo list atm, feel free to have a go.

[amueller]

Basically the cython side is not able to handle precomputed kernels in the sparse case.

[amueller]

@fabianp any help would be appreciated with this.
It seems to me there is some low-level wrapper coding to do.
In the sparse case, there is no way to create the kernel-indices (yet).
Looking at Kernel::precomputed_kernel in svm.cpp, I tried just stacking them to the front, but I get very weird results.
When I look into svm_csr_node, everything seems fine, but when I look into kernel_precomputed the values have changed?! any idea where this comes from?

[amueller]

There is also this comment:
/* in the case of precomputed kernels we do not use
dense_to_precomputed because we don't want the leading 0. As
indices start at 1 (not at 0) this will work */
which I must admit is not entirely clear to me. This is in the svm_csr_model function.

[amueller]

Closing as we don't really need sparse kernels. We need being able to use sparse X to precompute a dense kernel.