Implementing a dynamic "dict of metrics" efficiently

[orweiser-code]

specificaly - apply one metric (e.g. MeanMetric) on a dict of values, where the dict keys are not known in advance (i.e. should add elements depending on the input dict of values to the update / forward functions.

simply creating a class that holds a dict of MeanMetric objects is very inefficient in a ddp setting.

implementing a metric that changes states shape dynamically seems unsupported.

what is your suggested approach, and do you intend to implement something like this soon?