MulticlassAveragePrecision crashes on .compute() if empty about torchmetrics HOT 2 CLOSED

Hi! thanks for your contribution!, great first issue!

from torchmetrics.

Hi @gui-miotto, thanks for raising this issue.
In general we do not make any guarantees regarding metrics in torchmetrics what happens when .compute is called before .update (or forward). That the accuracy metric happens to just give a warning and return 0 is just arbitrary. One could make the argument that the accuracy metric should actually crash because the score is clearly undefined if no input have been seen. The same argument is true for average precision score.

Specifically, for average precision metric you can actually just set the thresholds argument and you will get a warning and returned value of zero:

print(MulticlassAveragePrecision(num_classes=3, thresholds=100).compute())  # prints tensor(0.)

why does this not crash? Because the default value of thresholds=None means that we are dynamically calculating the thresholds values based on input to the metric. No input to the metric no meaningful thresholds to compute. Does it then make sense to when thresholds=100 to return 0? Not really because, again all the calculations are undefined.

I therefore keep the stance that in general we cannot make sure that metrics where no input is provided can actually succeed with the calculation, which is exactly what the warning we are giving to the user before the program crashes indicates.

from torchmetrics.