I want to force a monotonic relationship between a set of features and the model’s outcome, and I would like to do it by controlling the model’s partial derivatives into the loss function, like a regularization term.
I tried the solution proposed here, but I got an error:
one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [2, 1]], which is output 0 of TBackward, is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).
Is there a way to include the model’s partial derivatives into the loss?