TrainConfig

class maze.train.trainers.common.config_classes.TrainConfig(env: omegaconf.DictConfig, model: maze.train.trainers.common.config_classes.ModelConfig, algorithm: maze.train.trainers.common.config_classes.AlgorithmConfig, runner: maze.runner.Runner)

Top-level configuration structure.

The structured configuration support of hydra is limited currently (1.0-rc2).

E.g.

  • Merging different configuration files did not work as expected (e.g. algorithm and env-algorithm)

  • Although the entry-point expects a TrainConfig object, it just receives a DictConfig, which can cause unexpected behaviour.

Note that due to this limitations, this class merely acts as type hinting mechanism. Behind the scenes we receive raw DictConfig objects and either need to invoke the Registry functionality or hydra.utils.instantiate to instantiated objects of specific types where required.