You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
SMAPE docs say: "smape [...] A tensor with non-negative floating point smape value between 0 and 1"
R^2 docs say: "By definition the score is bounded between 0 and 1"
Both of these are incorrect:
SMAPE ranges between [0,2]
R^2 ranges between (-∞,1]
This is true when generalized beyond linear models, as is the case for torchmetrics.
But tbh it might be helpful to add an optional flag that can clip the range to [-1,1], for the sake of tensorboard readability? I'll leave this decision to you.
P.S.
Here is an example tensorboard plot that is hard to read due to exploding negative R^2:
The text was updated successfully, but these errors were encountered:
PR #2771 fixes the incorrect statements in the docs but the clipping feature is not implemented as it is a very specific feature that depends on the usecase (sometimes it may be beneficial to know if the r2 score is highly negative indicating a very bad model)
📚 Documentation
SMAPE docs say: "smape [...] A tensor with non-negative floating point smape value between 0 and 1"
R^2 docs say: "By definition the score is bounded between 0 and 1"
Both of these are incorrect:
P.S.
Here is an example tensorboard plot that is hard to read due to exploding negative R^2:
The text was updated successfully, but these errors were encountered: