Issue with class Gamma

[anthonybritto]

I'm trying to construct the following model:

mod = GeneralizedLinearEstimator(Gamma(), skg.penalties.L1(alpha=0.5))

However, when I run mod.fit(exog, endog) I am met with

AttributeError: 'Gamma' object has no attribute 'intercept_update_step'

Looking at the documentation, I wonder if this is due a simple typo: I see that Gamma has the method intercept_update_self defined; should this be intercept_update_step instead?

Is there a possible workaround for now? I naively tried to define

g = Gamma()
g.intercept_update_step = g.intercept_update_self

but the error persists.

[mathurinm]

Hi @anthonybritto, thanks for raising the issue.
it is indeed a typo, we will fix it. But the issue is actually something else: since you're not specifying a solver, skglm will use AndersonCD, which does not work for the Gamma datafit.

Instead you should use the ProxNewton solver, as follows:

solver = ProxNewton(verbose=2)
mod = GeneralizedLinearEstimator(Gamma(), skg.penalties.L1(alpha=0.5), solver=solver)

feel free to reopen if it does not fix your issue.

EDIT: you'll also need to disable intercept fitting for a while, as the current implementation is probably not correct.
EDIT: in fact as pointed out by @Badr-MOUFAD, intercept fitting should be correct for ProxNewton

[mathurinm]

You'll need to checkout the branch #189 that fixes the issue, before we merge it into main.

EDIT: in fact it's currently implemented correctly in main, using a different method than intercept_update_step