Open tanglef opened 2 years ago
Related test failure: https://github.com/benchopt/benchmark_ridge/actions/runs/3572133863/jobs/6004677349#step:6:430
I'd skip glmnet in the tests and add a warning docstring in the glmnet
solver. @Badr-MOUFAD do you have time to take a look at it?
It was among the reasons why the CI test failed in https://github.com/benchopt/benchmark_ridge/pull/4
As far as I know, glmnet
uses a normalized datafit. Yet,
I tried adding/removing the scaling by n_samples
but didn't seem to solve the problem.
glmnet
ridge does not seem to always give the same results as others on simple datasets (even outside of Benchopt, that's why I present the R code): To solve the ridge problem: $\min \frac{1}{2}||y-X\beta||_2^2 + \frac{\lambda}{2}||\beta||_2^2$ the penalty inglmnet
is: $\lambda_R=\frac{\sigma(y)}{n}\lambda$ with $\sigma(y)$ the french standard deviationBut on datasets like
bodyfat
:@cassiofragadantas found this post that also says there are discrepancies not explained by rescaling So how would we treat this in Benchopt ? (a case where the solver does not converge to the same coefficients as others sometimes)