Experiment

Module containing the functions which create experiment components from Hydra configs.

This is essentially just calling hydra.utils.instantiate on the datamodule, network, trainer, and algorithm configs in a certain order.

This also adds the instance_attr custom resolver, which allows you to retrieve an attribute of an instantiated object instead of a config.

Experiment dataclass

Dataclass containing everything used in an experiment.

This gets created from the config that are parsed from Hydra. Can be used to run the experiment by calling run(experiment). Could also be serialized to a file or saved to disk, which might come in handy with submitit later on.

setup_experiment

setup_experiment(experiment_config: Config) -> Experiment

Instantiate the experiment components from the Hydra configuration.

All the interpolations in the configs have already been resolved by project.utils.hydra_utils.resolve_dictconfig. Now we only need to instantiate the components from their configs.

Do all the postprocessing necessary (e.g., create the network, datamodule, callbacks, Trainer, Algorithm, etc) to go from the options that come from Hydra, into all required components for the experiment, which is stored as a dataclass called Experiment.

NOTE: This also has the effect of seeding the random number generators, so the weights that are constructed are deterministic and reproducible.

instantiate_datamodule

instantiate_datamodule(
    datamodule_config: (
        Builds[type[DataModule]] | DataModule
    ),
) -> DataModule

Instantiate the datamodule from the configuration dict.

Any interpolations in the config will have already been resolved by the time we get here.

instantiate_algorithm

instantiate_algorithm(
    algorithm_config: Config, datamodule: DataModule
) -> LightningModule

Function used to instantiate the algorithm.

It is suggested that your algorithm (LightningModule) take in the datamodule and network as arguments, to make it easier to swap out different networks and datamodules during experiments.

The instantiated datamodule and network will be passed to the algorithm's constructor.

instantiate_network_from_hparams

instantiate_network_from_hparams(
    network_hparams: Dataclass, datamodule: DataModule
) -> Module

TODO: Refactor this if possible. Shouldn't be as complicated as it currently is.

Perhaps we could register handler functions for each pair of datamodule and network type, a bit like a multiple dispatch?