Add Mean Absolute Percentage Error metric

[pranjaldatta]

🚀 Feature

Add Mean Absolute Percentage Error to the list of regression metrics.

Motivation

It's a neat metric that is included in the scikit-learn package - considered to be a bit more intuitive than MAE and more sensitive to relative errors. More can be found on its Wikipedia article.

Pitch

Mean Absolute Percentage Error to be included among the regression metrics.

Alternatives

Additional context

I would be really grateful if someone could judge whether such a metric is necessary in the first place (considering it has some documented issues as noted in the wiki here). Although I guess, scikit-learn deals with the 0-related issues by adding an arbitrary constant to the denominator.

If the community finds that such a metric would be a helpful addition to this awesome package, I would love to contribute this!

[github-actions]

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

[SkafteNicki]

@pranjaldatta looks like a great addtion. Would you be up for sending a PR?
I see no problem in the limitations of this metrics, we just need to document some of them. Finally, we just provide the metrics, but it is up to the user to use them correctly

[pranjaldatta]

@SkafteNicki Thanks! I'll send a PR right away in 2/3 days. I'll make sure to reflect some of the limitations in the associated documentation.

[kowshikthopalli]

Is there a reason why MAPE in torchmetrics doesn't multiply with 100 to actually get the percent?

[orestis-z]

Is there a reason why MAPE in torchmetrics doesn't multiply with 100 to actually get the percent?

I agree, I think this should be mentioned explicitly in the docs to avoid confusion. It's a "mean absolute relative error". @pranjaldatta