Hello!
I was wondering how someone would implement a loss function that uses a percentage value of “how far is the predicted tensor from the target tensor”, without knowing the actual target. I tried implementing fake target tensors but I guess the math behind that is not correct, since what I tried was subtracting the percentage from my predicted tensor. Meaning the target tensor would constantly change in proportion to the predicted tensor, which seems wrong to me. But how would I go differently about that? I also tried replacing the value of the predicted tensor with the percentage, but since the percentage calculation is not integrated into the loss calculation, the loss is constantly rising to infinity (at least this is what I believe). All help is appreciated, since I’m very new to pytorch!
Have a lovely day, everyone!