Add support for hierarchical dict

🚀 Feature

Motivation

Since v0.7.0, LightningModule accepts dict hparams, however, still TensorBoardLogger raises an error with hierarchical dict. Considering the compatibility of the other package, especially Hydra #807, hierarchical dict should be accepted by any loggers.

Pitch

• Flatten hierarchical dict before hparam logging

Alternatives

The function _convert_params in loggers/base.py will be changed like:

def _convert_params(self, params: Union[Dict[str, Any], Namespace]) -> Dict[str, Any]:
    # in case converting from namespace
    if isinstance(params, Namespace):
        params = vars(params)

    # added part
    params = flatten_dict(params)  # convert {'a': {'b': 'c'}} -> {'a/b': 'c'}

    if params is None:
        params = {}

    return params

[jeremyjordan]

Could we implement this in such as way to enable the use of writer.add_scalars() for the Tensorboard version? When you use this method it plots the values on a single chart (versus grouping charts with individual metrics when you organize tags like parent/child).

eg.

versus

[Borda]

I saw the PRs #1130 #1128, then I feel there are some conflicts between the PRs. For example, if non-primitive types are converted to str by #1130, #1128 does not have any effects.

Could you pls comment on these PRs thow…

The integrative discussion around hparams is needed I think.

Could you open an issue and raise/start such discussion (or jsut continue in the Hydra one?)