Add float32 support for SGDClassifier and SGDRegressor

[elanmart]

Hi,

I was wondreing why sklearn does not allow me to specify the dtype I'd like to use.

I'm working with rather large dataset, which fits into my RAM as float32, but when I'm trying to train a simple SGD on it, the model tries to copy my data into float64, causing MemoryError.

I can change this in my local sklearn build, but I guess there is a good reason why this is not a free parameter?

[amueller]

PR welcome ;)
It would be great to have both 32 and 64 bit everywhere, I think.
This means using fused types everywhere in Cython.

One of the reasons not to do that was the explosion in the generated C code, but I think with #5492 we need to be somewhat less careful about that.

[lorentzenchr]

#13243 implemented float32 support for SequentialDataset. This has to be propagated to _plain_sgd in _sgd_fast.pyx (and other places) in order to enable to preserve float32 X ndarrays in SGDClassifier and SGDRegressor.

[jjerphan]

@OmarManzoor: you might be interested in this slightly harder Cython issue.

[OmarManzoor]

@jjerphan Thank you for suggesting.

[OmarManzoor]

@jjerphan Do we need to change the _sgd_fast.pxd file in relation to this? On checking its usages it is mainly being used inside the _sag_fast module which I think already supports float32 and float64.

[OmarManzoor]

@jjerphan Do we need to change the _sgd_fast.pxd file in relation to this? On checking its usages it is mainly being used inside the _sag_fast module which I think already supports float32 and float64.

On further investigation I don't think we need to change this file.