Gradient Can't Be Computed with Variable in For Loop

1

Hi, I’m trying to implement an autoregressive time series model by creating an initial tensor of zeros called motion and then filling in the value of each element in the tensor using a for loop. But I’m getting an error when training that says: ‘one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor []], which is output 0 of AsStrideBackward0…’

I think this error has something to do with filling the value of the tensor motion in a for loop, but I’m not sure if there’s a better way to do the computation that avoids the error? I know that for loops and other arbitrary Python control flows are possible in probabilistic models. Is there a good way to implement these computations without the gradient error? The relevant code in the model is below.

  # - Local trend component.
  # Prior for time-global scale.
  drift_scale = pyro.sample('drift_scale', dist.HalfNormal(25).expand([data_dim]).to_event(1))
  with pyro.plate('time', len(data), dim=-1):
    drift = pyro.sample('drift', dist.Normal(0, drift_scale).to_event(1))

  # Initialize motion tensor.
  motion = torch.zeros(len(drift) + 1)
  # Autoregressive coefficient.
  rho = pyro.sample('rho', dist.Uniform(-1, 1))
  for i in range(len(drift)):
    motion[i+1] = motion[i] * rho * drift[i]
  motion = motion[1:].unsqueeze(-1)  # drop initial state and add data dimenion of size 1
2

there’s usually no particularly compelling reason to use in-place ops. instead it’s better to append to a list and use torch.stack and torch.cat and such. or define prev_motion and keep updating it. see e.g. here.

3

Ah ok, so don’t modify tensors in place. I went ahead and just appended the values to a list and converted the list to a tensor, and that worked without any issues. Thanks!