I understand that you want to effectively train a regression model with multiple output heads and the outputs are of different orders.
My advice would be to use weights while adding L2 regression losses. This would normalise the magnitude effect caused due to difference in order.
For example:
Loss = Loss1*W1+Loss2*W2;
% If Loss1 s order > Loss2’s order W1<W2 (example: W1 = 0.001, W2=1)
Here W1 and W2 can be treated as hyper parameters and can be tuned accordingly for effective training.
Please refer to the following link for implementing custom training loop and loss function with weights for multiple outputs:
Please try this and reach out to us for any further issues.
Sincerely
Varun
