Manually averaging metrics when logging

[digital-idiot]

I have a metric from torchmetric as follows:

Accuracy(
    num_classes=self.model.out_channels,
     average='none',
     ignore_index=self.ignore_index
)

Obviously I can not log this, however I don't want to set average to any aggregation. I want to log its mean in training_step but want to preserve the class wise metric to till end of epoch where I display it to terminal. I want the metric to reset at epoch end only, so can't call compute() in training step.

How to solve this?

[digital-idiot]

From this comment I had the idea that:

if the .compute()method is called the internal state is reset.

However, this behavior has changed, now calling compute() does not reset the state of the metrics. See PR #5409