tensor.sign? no backward?

[edwardzcl]

See https://github.com/itayhubara/BinaryNet.pytorch/blob/master/models/binarized_modules.py

I want to know what is the backward process of this function. Only use tensor.sign()? What is the backward values of it？

[hainingfang]

since it uses tensor.data = Binarize(tensor.data.org) in the code, so PyTorch won't record the grad op in the graph, in other words, no need to worry about the backward op, it won't happen at all.

[whubaichuan]

I think the gradient will be set to 1 because of the code

    def backward(ctx,grad_output):
        #STE 
        grad_input=grad_output.sign()
        return grad_input,None,None,None

[zouyingcao]

I think the gradient will be set to 1 because of the code

    def backward(ctx,grad_output):
        #STE 
        grad_input=grad_output.sign()
        return grad_input,None,None,None

There is a question that there is no grad_output.sign() in the current code version. Just grad_input=grad_output there. Is there any error about code update?

[whubaichuan]

I think the gradient will be set to 1 because of the code

    def backward(ctx,grad_output):
        #STE 
        grad_input=grad_output.sign()
        return grad_input,None,None,None

There is a question that there is no grad_output.sign() in the current code version. Just grad_input=grad_output there. Is there any error about code update?

No, because 1.sign() = 1