How to log a Tensor Metric

[wellhowtosay]

I implement a Metric which returns a Tensor or a Scalar in compute, I try to log the Metric in *_step like:

def test_step(self, batch, batch_idx):
        loss, pred, inp, target = self.compute_loss_and_acc(batch)
        metric = self.val_csc_metric(pred, inp, target)
        self.log("CSC_metric", self.val_csc_metric, on_epoch=True, on_step=True,logger=True)

but meet this:
pytorch_lightning.utilities.exceptions.MisconfigurationException: The metric tensor([0.0541, 1.0000, 0.1026, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, inf]) does not contain a single element, thus it cannot be converted to a scalar.
Exception ignored in: <function tqdm.del at 0x7f8ba571f710>

I use self.log because I need it's on_epoch option to sync metric states among multi gpus at the end of eval/test epoch.

How can I fix it or is there a alternative.
Thanks !

[talhaanwarch]

I think it would be self.log("CSC_metric",metric , on_epoch=True, on_step=True,logger=True) because val_csc_metric is a function and you want to log output of that function

[tchaton]

Yes, doing metric = self.val_csc_metric(pred, inp, target) performs a forward and metric is actually the result for the current batch.

However, you can provide the metric directly as suggested by @talhaanwarch.