Multi-output regression using Hugging Face library

1

Hi,
I have been using the Hugging Face library, specifically the ‘robertaforsequenceclassification’ model. I wanted to train a multi-output regression model and for that I have set the number of out classes to two. Essentially, I want to predict a geo-coordinate as output. But, the hugging face document mentions that for regression tasks, the number of classes needs to be 1, in which case an RMSE loss is automatically used. While if number of classes is greater than one, a cross entropy loss is used. Is there a way to use the ‘robertaforsequenceclassification’ model for multi-output regression?

2

Hello @Shreyas_Mangalgi

Did you find any way to do multi-output regression with the Transformers library?

Thanks!