tensor.py 文件源码

def backward(self, grad_output):
        grad_input = torch.cumsum(-grad_output, dim=self.dim)

        end_idx = grad_input.size(self.dim) - 1
        grad_sum = grad_input.narrow(self.dim, end_idx, 1)
        grad_input -= grad_sum.expand_as(grad_input)
        grad_input += grad_output
        return grad_input


# TODO: unfold