I am using torch.nn.functional.poisson_nll_loss as a loss function. Now I would like to take into account the weights for each instance (for example, exposure in a risk application). I am thinking all I need to do is set reduction='none' and then dot the vector of losses with the weights column. Does that sound right?
Yes, your idea sounds reasonable. Iām not familiar with your exact use case, but you should consider normalizing the reduced mean with the used weight as seen here to avoid creating a data-dependent loss spike.
1 Like