Pytorch - RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed

Question

I keep running into this error:

RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

I had searched in Pytorch forum, but still can’t find out what I have done wrong in my custom loss function. My model is nn.GRU, and here is my custom loss function:

def _loss(outputs, session, items):  # `items` is a dict() contains embedding of all items
    def f(output, target):
        pos = torch.from_numpy(np.array([items[target["click"]]])).float()
        neg = torch.from_numpy(np.array([items[idx] for idx in target["suggest_list"] if idx != target["click"]])).float()
        if USE_CUDA:
            pos, neg = pos.cuda(), neg.cuda()
        pos, neg = Variable(pos), Variable(neg)

        pos = F.cosine_similarity(output, pos)
        if neg.size()[0] == 0:
            return torch.mean(F.logsigmoid(pos))
        neg = F.cosine_similarity(output.expand_as(neg), neg)

        return torch.mean(F.logsigmoid(pos - neg))

    loss = map(f, outputs, session)
return -torch.mean(torch.cat(loss))

Training code:

    # zero the parameter gradients
    model.zero_grad()

    # forward + backward + optimize
    outputs, hidden = model(inputs, hidden)
    loss = _loss(outputs, session, items)
    acc_loss += loss.data[0]

    loss.backward()
    # Add parameters' gradients to their values, multiplied by learning rate
    for p in model.parameters():
        p.data.add_(-learning_rate, p.grad.data)

Solution

The problem is from my training loop: it doesn’t detach or repackage the hidden state in between batches? If so, then loss.backward() is trying to back-propagate all the way through to the start of time, which works for the first batch but not for the second because the graph for the first batch has been discarded.

there are two possible solutions.

1) detach/repackage the hidden state in between batches. There are (at least) three ways to do this (and I chose this solution):

 hidden.detach_()
 hidden = hidden.detach()

2) replace loss.backward() with loss.backward(retain_graph=True) but know that each successive batch will take more time than the previous one because it will have to back-propagate all the way through to the start of the first batch.

Example

Comments

[nikhilweee]

The PyTorch tutorial on LSTMs suggests something along the following lines model.hidden = model.init_hidden() You need to clear out the hidden state of the LSTM, detaching it from its history on the last instance.

[Tengerye]

Variable is deprecated now (pytorch.org/docs/stable/autograd.html#variable-deprecated).

[Tom Dörr]

Regarding solution 1: Why do we need to detach two times? Isn't detach_() an inplace operation that makes hidden = hidden.detach() unnecessary?

[stason]

It is. He must have meant one or the other.

[stason]

The pytorch-1x+ way implementation example link is here

[Peyman]

@nikhilweee your link doesn't include any init_hidden() anymore.

[nikhilweee]

@Peyman You're right. I guess my solution was only valid for PyTorch < 1.0