Lost type in PyTorch optimization

Question

I'm trying to implement a simple minimizer in PyTorch, here is the code:

for i in range(10):
    print('i =', i, ' q =', q)
    v_trans = transform_dq(v, q)
    loss = mse(v_trans, v_target)
    loss.backward()
    q = q - eta * q.grad

print('Final q = ', q)

Where v and q are tensors, and eta = 0.01.

The problem is that at the line q = q - eta * q.grad I get an error on the second iteration of the loop:

TypeError: unsupported operand type(s) for *: 'float' and 'NoneType'

It looks like the update for q changes the graph in an unwanted way (q is not a leaf of the graph anymore and hence it doesn't have a grad). If that is the case, then how to implement this simple minimizer?

Answer 1

First, you need to reset q's gradients before each iteration.
Second, you should update q outside "gradient scope":

with torch.no_grad():
    q = q - eta * q.grad