Quantile Regression Loss

[vincentqb]

When interested in fitting an unknown probability distribution using a neural network model, we are sometimes interested in quantile regression. I suggest we add the quantile regression loss to the losses we have in pytorch.

• Example implementation here.
• Relates to Time series audio#267

cc @albanD @mruberry

[mruberry]

Great idea, @vincentqb!

[maxmarketit]

I have implemented QuantileLoss and QinputLoss in this notebook.

I followed the pytorch Loss Class and I do not know anything about torch.jit.is_scripting() part so it is all commented out. Anyone might help here.

It works fine in the example and I tested it with p(p>2) input features and even q(q>2) output features(for QinputLoss, and I figure I should have generalized the QuantileLoss too with tensor output like batch size * output variables * quantiles).

Take a look at it and be free to say any suggestions.

https://github.com/maxmarketit/Auto-PyTorch/blob/develop/examples/quantiles/Quantiles.ipynb

ps) QinputLoss is from Tagasovska, N., & Lopez-Paz, D. (2018). Single-Model Uncertainties for Deep Learning. https://arxiv.org/abs/1811.00908v3.

[Javicadserres]

Hi, I don't know if there has been any update on this topic but I have implemented the Quantile loss function vectorized here with Pytorch.