How to use L2 Loss , L1 Loss or MSE Loss while training Deepar Model ?

[sayedathar11]

Hiii ,
Our team is using GluonTS Deepar model for time series forecasting our target has range 0 to 1 , As per our information , GluonTS Deepar uses StudentTOutput and it doesn't uses L1 Loss , L2 Loss or Mse Loss , we wanted to ask how to change loss function in GluonTS Deepar ? Is there any fucntionality that allows us to select L1 Loss or L2 Loss or Mse Loss or we can only go with StudentTOutput ?
@mbohlkeschneider @lostella can you please help with it ?

[hardik-raja-dsw]

hi, we had the same query related to loss function, any help on same would be appreciated, Thanks in advance

[mbohlkeschneider]

Hi @sayedathar11 and @hardik-raja-dsw,

The Laplace and Gaussian distribution essentially reduces to L1 and L2 loss when fixing the variance, respectively. For L1/Laplace, this is done here: https://github.com/awslabs/gluon-ts/blob/master/src/gluonts/mx/distribution/laplace.py#L123.

You can implement something analogous for the Gaussian distribution to get an L2 loss.

Hope that helps.

[sayedathar11]

@mbohlkeschneider That works , Thanks alot for your reply