How to calculate training loss
abrosua opened this issue · 3 comments
I'm currently trying to recreate the PyTorch version of LiteFlowNet for both the inference and training pipeline, and I have a question regarding the training loss calculation.
It's shown in the train prototxt that the L2 training losses are calculate at every level for each stage (M, S and R), thus resulting with 15 losses value.
How to calculate the training loss? Is it:
a.) the Sum product of all the 15 values (with their respective loss weight)
OR
b.) Calculate the mean loss value for each level (M, S and R), then calculate the sum of it (the 5 loss values from every level).
Or maybe by other method.
Thank you very much for your attention, best regards.
I am not so sure whether I fully understood your question. Are you asking how to compute the L2 losses or the total loss?
I'm asking for the total loss computation
Total loss = The weighted sum (using the respective loss weights as the training prototxt) of all the losses of each stage (M, S, R) at each level.