Coordinate descent should work on float32 data (in addition to float64 data)

[arthurmensch]

For large scale application, constraining ElasticNet input to be float64 is a waste of space, as same quality results should be obtainable with float32 or event float16 input.

Use of cython in coordinate descent module make the type flexibility a bit tricky, but should be doable.

[GaelVaroquaux]

Use of cython in coordinate descent module make the type flexibility a bit tricky, but should be doable.

We need to use fused types for this (read on cython fused types).

[ogrisel]

Float16 is too low: BLAS does not support it. But float32 (in addition to float64) is a good idea.

BLAS comes in 2 flavors (for operations on real numbers as in non-complex): single precision (float32) and double precision (float64). The single precision operation are prefixed with s (e.g. saxpy and sgemm) while double precision operations are prefixed with d (e.g. daxpy and dgemm).

[MechCoder]

I can work on this, if noone is already!

[arthurmensch]

Sure, but we need to remove c files from the repository before, so as not
to blow repository size with generated c files from fused type.
Le 20 oct. 2015 23:11, "Manoj Kumar" notifications@github.com a écrit :

I can work on this, if noone is already!

—
Reply to this email directly or view it on GitHub
#5464 (comment)
.

[MechCoder]

Oh the idea is to make the build script generate c from pyx files locally instead of them being the repository?

[Sorry, if there was a discussion but I missed it]

[arthurmensch]

Yes, checkout PR #5492. Still WIP !

[amueller]

@MechCoder but you can work on it independently. We just need #5492 be done before merging this.

I didn't realize you could do if typeof(x) is double with fused types, but apparently that's the way to do it.

[yenchenlin]

Hello @MechCoder ,
I've already go through #6430, may I take a stab at this issue?