#11950 has started to do it for a refactoring of K-Means and early benchmarks are promising.

Saw the discussion in PR ( #12184 ) seems to be forming a consensus around using a newer version of SciPy on master making this feasible. Does that mean PRs in this direction be welcome?

I think we would want to wait for the minor 0.20.1 bugfix release first (cf #12184), to reduce potential backporting issues. After which PRs would be definitely very welcome :)

BTW, are you aware of projects that have been doing this reliably for a while for a range scipy versions and OSs (e.g. on conda forge)? For instance, do we know if scipy 0.17 - 1.1+ does maintain a backward compatible ABI for the BLAS functions exposed in Cython ? There is a discussion implying it in scipy/scipy#6440 (comment) but nothing too official.

Sounds like a good question for SciPy devs (though I guess you did that already 😄). Maybe worth drawing their attention to this issue again?

There is a small(?) obstacle to this. We don't make calls to BLAS only from cython but also from C++, in sklearn/svm/src/liblinear/tron.cpp to be precise. There we can't use scipy cython bindings of BLAS functions.

One solution would be to port tron.cpp to cython. The file is not too big. Would you consider it worth ?

I don't know if there's an other way. Do you have any idea ?

Could it be possible to call Cython BLAS functions from C (and I guess C++) ? https://cython.readthedocs.io/en/latest/src/userguide/external_C_code.html#using-cython-declarations-from-c

I saw that but to do that you have to embed python into c, which would cause a huge overhead I guess.

which would cause a huge overhead I guess.

Yeah, taking time python -c "import numpy" as a benchmark, I would guess 100-200 ms (possibly less) to get the interpreter started. Doesn't sound great indeed.

Stupid question, why are we vendoring liblinear in sklearn/svm?

Stupid question, why are we vendoring liblinear in sklearn/svm?

LinearSVC uses liblinear,

Linear Support Vector Classification.
Similar to SVC with parameter kernel='linear', but implemented in terms of
liblinear rather than libsvm, so it has more flexibility in the choice of
penalties and loss functions and should scale better to large numbers of
samples.

If you question is why in svm instead of linear_module I don't know, maybe putting it alongside libsvm made more sense with repect to vendored code...