my loss function involves a term f(x) from which I want to compute part of the derivative by hand making use of the chain rule, i.e. df(x)/dw = df(x)/dx * dx/dw. The reason is that I have an explicit and pre-computable expression for df(x)/dx. The code is as follows
Tensor x;
Tensor f = compute_f(x);
Tensor df = compute_df_fx(x);
How can I make sure that the df term is used in the back propagation step?
Package compute_f() and compute_df_fx() as a custom autograd.Function.
Roughly speaking compute_f() will become your Function’s forward()
method, and compute_df_fx() will become backward().
If it is more efficient or more convenient for your use case, you can compute compute_df_fx() in the forward pass (in forward(), store in ctx (the context),
and then use it in the backward pass (in backward()).
Thanks very much for your suggestion. Meanwhile, I also came across another solution here which seems easier. Do you know if Variables are supported in C++?