What is the equivalent of Caffe's propagate_down: false?

[hyojin]

Let’s say I have a layer that takes multiple inputs.
In this layer (e.g. eltwise multiplication) I want to disable the backpropagation on one of the inputs.
In Caffe, I was able to do this by setting propagate_down: false for that input inside that layer.
How could this be achieved using Pytorch? Thanks a lot in advance.

[ruotianluo (Ruotian(RT) Luo)]

Suppose you dont want to backprop through inp2, then you could do:
op = layer(inp1, inp2.detach())

[hyojin]

Thanks. This makes sense, but I was unsure to use detach because of what is said in the following post:

Quick Detach() Question

If in2 is detached, you can’t call loss.backward at all, since net2(in2) are not taken into account by the computation of the gradient. When a variable is detached, the backward computations will not visit the branches that start from this variable (all the operations done on it). So as @albanD suggested, you need to create a new graph (hence the new variable requiring a gradient) if you want to compute a gradient with respect to the new network (net2) independently on previous operation (net1…

“When a variable is detached, the backward computations will not visit the branches that start from this variable (all the operations done on it).”

If I understand this statement correctly, this is not exactly what I want. I do want backprop on operations that involved this variable (although not with respect to this variable itself.)

[alexis-jacq (Alexis David Jacq)]

With detach, you will not backprop on operations that only involve the detached variable. But if an operation involves both your detached variable and a variable requiring grad computations, then you will derivate the operation, taking the detached variable as a constant.

[hyojin]

I see. Thanks for the explanation!