The basic idea is to have a metrics package, we can start with ROC/AUC (first on GPU, then if necessary on CPU). It should mimic the SKLearn API and results:

http://scikit-learn.org/stable/modules/generated/sklearn.metrics.roc_auc_score.html

Requirements:

1. GPU implementation. It can use Thrust and other libraries as first step.
2. Bindings for other languages (currently CTypes but soo

In most of the other projects I contribute to, we use pydocstyle to check conformity to PEP257. Once #259 is in, we can enable it by simply switching the line $ make flake to $ make pep. It's just a lot of errors, so it can be done when we are bored :)

(see branch docs)

• the fact that, when we can, we avoid doing hcat(X, 1)
• use IterativeSolvers.cg and not IterativeSolvers.lsqr as it allows specifying the operator as a linear map which is efficient and avoids copying when having to add a column for X; anyway it should be identical apart from pathological cases
• robust regression with concomitant scale estimation, not do

On the following line, I believe that the sparsity is destroyed by the subtraction.

https://github.com/jolars/sgdnet/blob/7e9261a83263a616cc07053c47f5cde5c2333cdb/src/utils.h#L70

It is not necessarily a big problem since it is only temporarily destroyed for each column, but could still probably be better implemented through a for loop.