Added multilabel reader in LibSVMFile. Fixed a bug in so_multiclass.cpp

[Jiaolong]

The previous one #2039 is closed due to some unrelated testing error in Travis report.

The changes also include updating revision of data submodule, where multilabel dataset were added.

[Jiaolong]

I list the changes here again in order to link to the related issues:

(1) This PR is mainly for #1987: add reader/writer for MultilabelLabels. A new function has been added in LibSVMFile.

(2) An example has been added, see io_libsvm_multilabel.cpp. Scene and yeast datasets have been tested.

(3) A tiny bug in so_multiclass.cpp was fixed. It occurred when compiling in ubuntu with mosek installed.

[vigsterkr]

@Jiaolong we need to at least TODO the previous discussion: i.e. let the file content decide whether it's multi- or single-labelled, not the API... this means to mark in the the file as well not just a PR comment that is being lost as soon as you delete/merge it.

[Jiaolong]

Yes, I am still working on this branch. But I am not totally clear about the idea. Could you elaborate me a little bit more?

So, for example, can we just use a sparse matrix to output the labels? No matter single or multiple, just unify the output.

[Jiaolong]

I mean to use a unified function, e.g.,

void get_sparse_matrix(SGSparseVector<float64_t>*& matrix_feat, int32_t & num_feat, int32_t & num_vec,  SGSparseVector<float64_t>*& matrix_label, int32_t & num_classes);

Then, the user doesn't need to specify the type of the file but get the labels from matrix_label, including single and multiple labels.

[vigsterkr]

@Jiaolong mmm yeah, exactly that's what i wrote yesterday....
but we need to give helper functions so that one can convert SGSparseVector<float64_t>*& matrix_label into float64* if it's a single labelled.... so that actually one can supply that to CMulticlassLabels or CBinaryLabels

and no that helper function should not be implemented in LibSVMFile...

[Jiaolong]

OK, cool! To make sure I have catch your idea, I put the changes as following:

(1) Change old get_sparse_matrix (float64_t* labels, ...) into get_sparse_matrix (SGSparseVector<float64_t>* mat_labels, ...), so we get unified get_sparse_matrix (SGSparseVector<float64_t>* mat_labels, ...) for both single and multiple labels

(2) We parse the file and know what is the type, and if it is single-labeled, we convert sparse matrix into float pointers. The convert (helper) function is implemented in CBinaryLabels and CMulticlassLabels

(3) So for the users, they only call CBinaryLabels and CMulticlassLabels while they don't actually need to know what's going on inside CLibSVMFile.

Am I right?

[Jiaolong]

I am sorry, the branch was polluted.