Allow updates to the outcome/partial resid before sample_one_iteration from R/Python interface

[jaredsmurray]

Description

Allow updates to the outcome in/before sample_one_iteration from R/Python interface

Ideas for Implementation / Expected Behavior

In the simples case we want to fit y = w'\beta + f(x) + eps; the relevant guts of the MCMC loop would be something like

beta = sample_beta(...)

forest_dataset$update_outcome(y - W%*%beta)

forest_model$sample_one_iteration(
    forest_dataset, outcome, forest_samples, rng, feature_types, 
    outcome_model_type, leaf_prior_scale, var_weights, 
    global_var_samples[i], cutpoint_grid_size
)

[andrewherren]

Great suggestion! Since the outcome object always maintains a partial residual with respect to forest and random effects predictions, I think we can probably add two methods to the outcome object for subtracting and adding values from that partial residual, so the workflow would look something like:

for (i in 1:mc) {
    if (i > 1) outcome$add_vector(W%*%beta)

    beta = sample_beta(...)

    outcome$subtract_vector(W%*%beta)

    forest_model$sample_one_iteration(
        forest_dataset, outcome, forest_samples, rng, feature_types, 
        outcome_model_type, leaf_prior_scale, var_weights, 
        global_var_samples[i], cutpoint_grid_size
    )
}

would that work?

[andrewherren]

This functionality has been implemented as sketched above. A demo of how this looks for an additive linear model is available in the Custom Sampling Routine vignette.