MetricWrapper for Target Binarization about torchmetrics HOT 3 OPEN

Hi! thanks for your contribution!, great first issue!

from torchmetrics.

Hi @lgienapp, thanks for opening this issue.
I am not against adding this feature, but maybe we should consider adding:

• an general MetricInputTransformer class where user can provide custom functions for transforming the input
• then BinarizedTargetWrapper can be included as just a subclass MetricInputTransformer with pre-selected transforms

how does that sound?

from torchmetrics.

Establishing a general class sounds good. Just for clarification: the general MetricInputTransformer would still subclass the WrapperMetric to inherit all the "reset-sync" code from there (e.g., live in torchmetrics.wrappers.transformations)? Or be its own thing (e.g., live in torchmetrics.transformations.abstract), possibly duplicating the sync code from the wrapper base class?

In either case, I would propose an implementation like this (subclassing wrappers here), assuming that only positional params like preds and targets would be interesting to modify (and thus ignoring kwargs e.g. indices):

class MetricInputTransformer(WrapperMetric):
  
    def __init__(self, wrapped_metric: Union[Metric, MetricCollection], **kwargs: Any):
        super().__init__(**kwargs)
        self.wrapped_metric = wrapped_metric

    def transform(self, *args) -> Tuple[torch.Tensor]:
        raise NotImplementedError

    def update(self, *args, **kwargs: Any) -> None:
        self.wrapped_metric.update(*self.transform(*args), **kwargs)

    def compute(self) -> Any:
        return self.wrapped_metric.compute()

    def forward(self, *args, **kwargs: Any) -> Any:
        self.wrapped_metric.forward(*self.transform(*args), **kwargs)

from torchmetrics.