Use with PL 1.2.3

[hadim]

Is that already possible to use the metrics package with PL1.2.3 or do we need to work on master or wait with next PL release?

[SkafteNicki]

It should already be possible, everything should work as you know it from pytorch_lightning.metrics.

[FlorianMF]

Using

self.train_acc(pred,target)
self.log('train_acc', self.train_acc, ..., on_epoch=True)

does not work though, because "reduce_on_epoch_end" does check whether the output element is a pl.Metric.
As the TorchMetrics instances aren't, the function tries to convert it to a tensor which leads to a RunTimeError.

[SkafteNicki]

Thanks @FlorianMF for correcting me.
When this PR Lightning-AI/pytorch-lightning#6540 gets merged on the lightning side it should work.

[SkafteNicki]

@FlorianMF yes, that is the idea. pytorch_lightning.metrics will still exist, but will be deprecated from v1.3 and removed in v1.5 of lightning. Also, all metrics in lightning will just (until removed) call torchmetrics underneath.

[Borda]

yes, Lightning-AI/pytorch-lightning#6540 and Lightning-AI/pytorch-lightning#6530 were merged 🐰

[Borda]

base on the comment above and Lightning-AI/pytorch-lightning#6540 and Lightning-AI/pytorch-lightning#6530 were merged 🐰 you shall be able to use it from master and atm we ain on full re-routing for PL 1.3 ⚡

[hadim]

Will we be able to use torchmetrics with 1.2.4 (or any releases before 1.3)?

[Borda]

v1.2.4 is out and it is not there, but if there is such hunger we can think about it for v1.2.5