ex.py

import torch

from torch.autograd import Variable
class AdvFunction(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input):
      print(type(input)  # torch.cuda.FloatTensor, why?
      ctx.save_for_backward(input)
      return output
            
    @staticmethod
    def backward(ctx, grad_output):
      print(grad_output.requires_grad)  # False! Why?
      input, = ctx.saved_variables  
      print(input.requires_grad)   # It is true.. Why?

      return grad_input

# The following is for question three.
            
class AdvFunction(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input):
      # do some operations and get the output
      ctx.save_for_backward(input)
      return output
            
    @staticmethod
    def backward(ctx, grad_output):
      grad_output = grad_output.data
      # do some operations on grad_output
            
      grad_input = Variable(grad_output, requires_grad=False) # ?

      return grad_input