In-place operation causing RuntimeError: Trying to backward through the graph a second time

autograd
1

Hi there, I’m new to PyTorch and I’ve been trying to fix a runtime error without success. I have defined a custom layer that has the following forward pass:

  def forward(self, x):

      w_times_x = torch.mm(x, self.weights.t() + (torch.mul(self.alphas, self.h_ij)).t())

      yout = torch.add(w_times_x, self.bias)

      self.h_ij = .99 * self.h_ij + .01 * torch.ger(x[0], yout[0]).t()

      return yout

Only self.weights and self.alphas are parameters of the model. The tensor self.h_ij is simply accumulating the product of units’ i and j activations over forward passes (for that I’m doing batch size of 1 such that the training samples are presented sequentially so that the updated value of self.h_ij and the end of the forward pass is used during the forward pass of the next training sample).

The training loop is the following:

for epoch in range(num_epochs):
    for batch_inputs, batch_outputs in dataloader:

        optimizer.zero_grad()

        predictions = model(batch_inputs)
        loss = loss_function(predictions, batch_outputs)

        loss.backward()
        optimizer.step()

The problem I’m having is that after the first .backward() call I get a RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). At the end of this error message it’s suggested I use .backward(retain_graph=True). If I do that I get a different error message RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation.

What am I doing wrong? I now that if I call self.h_ij.detach_() at the end of the forward method I got get the first error but this will remove (if I understand correctly) self.h_ij from the computational graph, so it won’t be taken in consideration when computing the gradients wrt to self.alpha.

Another way I can avoid the error is by simply “rebuilding” the tensor (like in the code bellow) but I think this must be braking the computational graph somehow.

def rebuild_h_ij(self):

    _aux = self.h_ij.tolist().copy()

    self.h_ij = torch.zeros((self.size_out, self.size_in), requires_grad = False)

    self.h_ij = torch.tensor(_aux, requires_grad = False, dtype = torch.float32)

Some help with this would be rather appreciated. Thx in advance.

2

Hi William!

w_times_x depends on your trainable parameters, so yout does as well,
and after you accumulated into h_ij, h_ij does to.

The first time through your training loop, h_ij doesn’t yet depend on your
layer parameters when you compute yout, so backpropagating through
yout doesn’t backpropagate through h_ij. But the second time through
your training loop, h_ij depends on the parameters through the
computation graph that was created during the first loop iteration which
was freed when you called loss.backward() in the first iteration.

Hence your first error.

If you use retain_graph = True, that graph is saved instead of
freed. But when you try to use it, autograd detects that your call to
optimizer.step() has modified various parameters – that are used
in the first-iteration computation graph – inplace, which would prevent
autograd from computing the gradient correctly were it to try.

Hence your second error.

You have to figure out what the correct logic of your training should be.
After you’ve gone through your training loop 100 times, say, when you
call loss.backward() in the 101st iteration, do you really want to
backpropagate through the computation of h_ij 100 times, once (again)
for each time it was computed in the 100 previous iterations?

Best.

K. Frank