Why is fbeta default threshold 0.2?

fastai fastai dev
1

I’m working on multilabel text classification and noticed that the default value for the fbeta metric (https://docs.fast.ai/metrics.html#fbeta) is 0.2.
Shouldn’t the default value be 0.5 like in accuracy_thresh? My reasoning is that since sigmoid=True the scores should be interpreted as probabilities and thus the 0.2 threshold is too low.

2

I have the same question as Martin, and I feel a default value of 0.5 may be better. I think it may have been set to 0.2 because it gave the best results with lesson 3’s planet dataset where the concept was explained.

3

Fell free to use any value you want. We defaulted it to one we thought was good but it’s of course dependent of the task you have :wink:

1 Like
4

Thanks, Sylvain!