Call backward of function in backpropagation #1776
Description
Unfortunately the following code snippet leads to an endless loop.
Is there a way to do backward computations on functions inside a backpropagation step?
It would be nice to have to implement quantized convolutions (e.g. like in https://arxiv.org/pdf/1612.01064.pdf)
import torch
class exampleFct(torch.autograd.Function):
def forward(self, dataIn):
self.save_for_backward(dataIn)
print("forward...")
return dataIn**2
def backward(self, grad_output):
dataIn, = self.saved_tensors
x_in_back = torch.autograd.Variable(torch.Tensor([2]), requires_grad=True)
y = x_in_back**2
print("backward...")
y.backward(torch.Tensor([1]))
print("unreachable")
print(x_in_back.grad.data)
return grad_output*2*dataIn
x = torch.autograd.Variable(torch.Tensor([[3, 4]]), requires_grad=True)
m = exampleFct()
m(x).backward(torch.Tensor([1, 1]))
print(x.grad.data)