evaluator init kwarg in train.py

[eggerdj]

What should we add?

Description

The following trainers are initialized from a trainer

• TransitionStatesTrainer
• ReweightingTrainer
• RecursiveTransitionStates
• RecursionTrainer

These trainers therefore do not have an evaluator_init in the config from which they initialize since the evaluator is provided by the sub_trainers. This poses a problem in train.py since we cannot pass the initialization arguments for the evaluator. In particular this block of code is an issue

        if evaluator_init_kwargs_str is not None:
            if "evaluator" in conf["trainer_init"]:
                evaluator_cls = EVALUATORS[conf["trainer_init"]["evaluator"]]

                evaluator_init = conf["trainer_init"].get("evaluator_init", dict())
                evaluator_init.update(evaluator_cls.parse_init_kwargs(evaluator_init_kwargs_str))
                conf["trainer_init"]["evaluator_init"] = evaluator_init
            else:
                raise ValueError(
                    f"evaluator_init_kwargs{train_idx} given but no evaluator "
                    f"in trainer {trainer_name}."
                )

Proposed enhancement

To fix this we could modify the from_config methods of the trainers listed above such that they optionally accept an evaluator in their config. If an evaluator is specified then we override the evaluator of the sub_trainer.