Pinball loss for quantile predictions

[trdvangraft]

Is your feature request related to a problem? Please describe.

For a regression problem, you are often not interested in the actual outcome but more in the boundary where the outcomes lay between. This can be achieved through for example taking a quantile output. This however often requires taking the quantile loss function (Pinball loss), which is not natively supported within Fastai. I think that fastai can benefit from this loss function or at the very least give more flexibility to the user.

Describe the solution you'd like
I would like to implement the quantile loss function similarly how it is done in the following post.
TLDR: def quantile_loss(q, y_p, y): e = y_p-y return tf.keras.backend.mean(tf.keras.backend.maximum(q*e, (q-1)*e))

Describe alternatives you've considered
Afaik there are no alternatives to the quantile loss function as suggested in this issue.

Additional context
I am completely able to implement this function myself, however I thought it would be nice to test if there was any interest into this function

[Sandy4321]

great idea
thanks
when it will be ready ?

[Sandy4321]

it is not implemented yet
https://docs.fast.ai/losses.html#DiceLoss
class BaseLoss
class CrossEntropyLossFlat
class FocalLoss
class FocalLossFlat
class BCEWithLogitsLossFlat
BCELossFlat
MSELossFlat
L1LossFlat
class LabelSmoothingCrossEntropy
class LabelSmoothingCrossEntropyFlat
class DiceLoss

[ThomasBourgeois]

That would clearly be a great add-on.

[KossaiSbai]

If people are still interested in having that function implemented, more than happy to take this on