Logging using a torchmetric object that returns dictionary

LightningModule
1

Hi everone,

Some of metrics in torchmetrics returns dictionary when I call its compute() method.
For example, torchmetrics.SQuAD() returns
{'exact_match': tensor(0., device='cuda:0'), 'f1': tensor(3.2220, device='cuda:0')}
when I call compute().

In [mode]_step method in lightning module, I tried to log with SQuAD() but I couldn’t.
self.log("score", METRIC_OBJECT, on_epoch=True, prog_bar=True, logger=True)
Error message: The .compute() return of the metric logged as ‘score’ must be a tensor.

Is there a way to log these metrics without overloading on_[mode]_epoch_end?

2

@iamseokhyun There is also a self.log_dict() method, that accepts a dictionary to log. Can you try that?