EvalResult doesn't do mean_of_gpus if using TensorMetric

[xiadingZ]

I want to use new AccuracyMetric, it can automatically sync in ddp, but it doesn't divide by word_size. In manually mode, I can divide it by word_size by hand in validation_epoch_end. But if I use EvalResult, how to do this? It only do mean across batches, but no across gpus.
This is original code:

    def validation_epoch_end(self, outputs):
        avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
        avg_acc = torch.stack([x['acc'] for x in outputs]).mean()
        avg_acc = avg_acc / self.trainer.world_size

[Borda]

@SkafteNicki @justusschock mind have a look?

[SkafteNicki]

@xiadingZ this was solved in a recent PR (#2568). You can now set the reduce_op='avg' when you construct the metric, and it will calculate the mean and not the sum.