Using Autogluon for supervised regression problem. I wonder how to use Mean Absolute Percentage Error (MAPE) as a loss function and better still, supply a custom loss function to autogluon? This is how my versions look like.
autogluon==0.2.0 autogluon-contrib-nlp==0.0.1b20210201 autogluon.core==0.2.0 autogluon.extra==0.2.0 autogluon.features==0.2.0 autogluon.mxnet==0.2.0 autogluon.tabular==0.2.0 autogluon.text==0.2.0 autogluon.vision==0.2.0 gluoncv==0.10.3.post0 gluonts==0.7.3 mxnet==1.8.0.post0
I am supplying
eval_metric='mean_absolute_error' in the
TabularPredictor constructor, but is it actually the loss function that the engine is minimising or is it just for reporting purpose? I do not see any loss parameter like I usually supply in Tensorflow, so any documentation around how the signature of a custom loss function would look like and where to supply it would be great.