Draft: Refactor API

[simon-hirsch]

Starting a PR to refactor the public API for the Estimator object.

Goals:

• [x] Separate data and modeling choices by introducing the equation dictionary that specifies the regression equation for each distribution parameter by passing a list of column identifiers (column indices, names) or string parameters intercept and all.
• [x] Have only one X and y in Estimator.fit(X, y)
• [ ] Basic support for pandas and polars
• [ ] Allow most estimation parameters to be passed either as single int, bool, float or as dictionary of {parameter: value}
• [x] Add intercept by default
• [ ] Allow to select the columns that should be scaled
• [ ] Prepare for the introduction of autoregressive effects on the distribution parameters

Discussions in #23 and #24

[simon-hirsch]

Implemented a first version that follows the API. Changes are mostly cosmetic, i.e. internally most stuff is working as before. Notable changes:

• Added equation parameter in the __init__()
• Adding intercept by default, aligned to sklearn in the fit_intercept parameter in the init
• Only one scaler needed from now on, as the scaler is working on X directly
• Forget can now be dependend on the distribution parameter. Interesting for tracking parameters with an intercept only. Question @BerriJ how do we handle discounting of the deviance? Add seperate forget_deviance? For now I've used self.forget[0] which is the forget of the mean.
• The equation parameter is designed in a way to support pandas and polars already, but the scaler will fail.
• Added a make_lags method already, but not implemented in the API jet.

Need to do now:

• [ ] Re-write some tests to cope with the new API
• [ ] Allow column selection to be scaled (or better to be excluded)?

[simon-hirsch]

Code snipped to play around with after building from source on the branch:

import rolch
import numpy as np
from sklearn.datasets import load_diabetes

X, y = load_diabetes(return_X_y=True)

online_gamlss_new = rolch.OnlineGamlss(
    distribution=rolch.DistributionT(),
    equation={0: "all", 1 : np.arange(4, 8)},
    fit_intercept=True,
    method="lasso",
    scale_inputs=True,
)
online_gamlss_new.fit(X, y)
online_gamlss_new.update(X[[-1], :], y[[-1]])

print(online_gamlss_new.betas)

[simon-hirsch]

Overall, I think this is a great improvement to the interface!

Thanks!

My main comment - or question - is about how you are handling pandas and polars. Is all that happens just a conversion to numpy? In this case, you may want to make the conversions extraneous to the estimator. skpro also has default conversions implemented, if you want to use the boilerplate or just the datatypes module, though you probably do not want to take the dependency.

The "nice" part would be if all internal calculations are native, though this may be out of scope.

Yes, I will be all handled by subsetting and conversion to numpy. We have quite a bit of numba supported code in the deeper workings of the package and I don't see a huge benefit in re-writing e.g. coordinate decent to pandas or polars. Another reason is that this keeps the dependency on these libraries rather light, while the main numpy API is rather stable, which makes maintenance easier :)