🚀 Feature

When we use nested metrics (i.e. ClassificationReport) we have a structure like this:

{
   'macro_avg': {
         'precision': ...,
         'f1-score': ...,
         'recall': ...,
    }
}

Currently it goes through this code path when adding the metrics to the engine state metrics.

Which will fill the metrics dict in the next way:

{
   '$ENGINE_METRIC_NAME': {
         macro_avg': {
           'precision': ...,
           'f1-score': ...,
           'recall': ...
          },
    },
    macro_avg': {
         'precision': ...,
         'f1-score': ...,
         'recall': ...
     },
}

$ENGINE_METRIC_NAME is the one defined here:

    classification_report.attach(evaluator, "$ENGINE_METRIC_NAME")

The main problem in that is for when we want to log metrics using TensorBoard logger for example, we need to either specify what metrics we want or all.

my code below

def setup_tblogger(
...
    train_metric_names: list[str] | str = "all",
...
) -> TensorboardLogger:

    logger.attach(
        trainer,
        log_handler=OutputHandler(
            tag="train",
            metric_names=train_metric_names, # list or all
            global_step_transform=global_step_from_engine(trainer),
        ),
        event_name=Events.EPOCH_STARTED,
    )

And in the case of classification report, we end with something not much desirable, that is repeated metrics in the TensorBoard logger, since we will flatten and get something like this:

$ENGINE_METRIC_NAME/macro_avg/precision
macro_avg/precision

We could potentially list only the metrics we want, or a blacklist of metrics that we don't want (not supported), but in any way, would be just a workaround of the real problem.