Open cdcseacave opened 1 week ago
Hi cdcseacave, may I ask have you resolved this problem?
After a lot of trial and error I think the gradient is computing well now, however I still do not know a good procedure to tackle this task: how to modify/add the derivative in the backward step such that it reflects the forward math changes, and how to best test the changes (I tried using torch.autograd.gradcheck
but first of all I had to convert everything to doubles otherwise it does lack accuracy, and even like that it is not obvious if there is a problem how do I spot where in the code it occurs).
Great project, thank you! This is more a cry for help. I'm trying to add new functionality to the forward renderer, which works great, but next I have huge problems understanding how to write the backward pass.
Do you have any suggestions on how to best approach this task?