Open FFFox-abc opened 1 year ago
Hey! Thanks for your interest in our work! Setting the loss weights really depends on the absolute value of the weighted losses. In general, different loss functions may give losses at very different scales, and we hope to scale them to an appropriate common scale. We also observed in our experiments (e.g. ablations) that some parts of the network may need to be updated more than others, so we tune the weights accordingly.
I noticed that the loss is composed of 5 part. In the paper said: " We set loss weights as λt = 0.01, λs = 10, λdist = 0.001, λgen = 0.005, and λdis = 0.005." is there some tricks to set? or just set accroding experience?