Exporting backpropagation graph

autograd
1

Is there any way to export/save the graph used for backprop as a regular pytorch feedforward network?

I have a network with fixed weights that computes a loss function. I need just a small number of backprop iterations to modify my input into a desired result. If I could export the backprop graph then I could effectively unroll the small number of iterations to create one larger network that computes the result in one shot.

2

No, that isn’t possible.
The backward computations are not even necessarily visible as functions. (Personally, I think it would be worth making all backwards available to be more consistent, but that sentiment hasn’t caught on, possibly because most potential use cases seem to have limited attraction.)

Best regards

Thomas