Make metrics on tests more stable?

[ismael-mendoza]

It's strange that the precision decrease from 0.5 to 0.25 even with different noise realizations. What do you think @thuiop ?

Originally posted by @ismael-mendoza in https://github.com/LSSTDESC/BlendingToolKit/pull/237#r739253386

[thuiop]

I repost my answer here :

Well the magnitude of the change is not abnormal, right now if there is one more or one less detection the value will change a lot. We should think about how to make this metric more useful.

[ismael-mendoza]

agreed - I didn't think it warranted holding up this issue much longer, thanks for your suggestion

[ismael-mendoza]

The solution is probably to make the test slower by having more sources. Not sure if this is a good idea.

[ismael-mendoza]

Test suite was changed by #412, metrics are now computed on aggregate over many batches on simple inputs so likely will always work otherwise something really wrong with algorithms implementation.