[ChrisW09]

Fixes #513

Problem

__init__ derives the state fit() reads from the three config parameters: the architecture config from model_config (self.config), the preprocessor from preprocessing_config (self._preprocessor, self._preprocessor_kwargs) and the optimizer from trainer_config (self._optimizer_type). SklearnBase.set_params() re-derived this state only when a config object replaced the config of an estimator that already held one. As a result:

• On a default-constructed estimator, config objects passed through set_params() were stored but never trained. GridSearchCV clones such an estimator for every candidate, so searches over model_config objects trained the default architecture.
• set_params(model_config=None), preprocessing_config=None or trainer_config=None kept the previous architecture, preprocessing and optimizer, and with all three set to None, clone() raised ValueError.

Fix

deeptab/models/base.py: a new SklearnBase._set_config_param(name, value) assigns one of the three config parameters and re-derives the state fit() reads from it, as __init__ does:

• a config object sets config/_config_kwargs, rebuilds the preprocessor from the PreprocessingConfig, or takes the TrainerConfig's optimizer;
• None (replacing a config) restores the defaults of an estimator constructed without it: the default model config, the default preprocessor with empty _preprocessor_kwargs, and Adam.

Both branches of set_params() use it. The legacy branch, which default-constructed estimators take, therefore keeps accepting flat keys such as layer_sizes next to a config object in the same call. Objects that aren't DeepTab configs are stored unchanged, as before. The set_params docstring says that assigning a config updates what fit() trains with.

Tests

tests/models/test_estimator_configuration.py::TestSetParamsConfigsReachFit:

• test_model_config_on_default_estimator_is_trained
• test_trainer_and_preprocessing_configs_on_default_estimator
• test_grid_search_over_model_config_objects_trains_selected_config
• test_setting_configs_to_none_restores_defaults (including clone() afterwards)
• test_flat_keys_still_apply_next_to_a_config_object: MLPRegressor().set_params(trainer_config=..., layer_sizes=[4]) keeps working.

The first four fail on main and pass with this change. The last one passes on main too and guards that behaviour.

A related defect with its own cause, model_config__* keys being dropped when the estimator has no model_config, is #613 (fixed in #614); the two PRs merge cleanly.

Checks

• just test: 3241 passed, 46 skipped, 28 xfailed; coverage 91.32%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)