Runtime error: loss.backward() with autograd Variable

Hi, I’m trying to call backward() on a Variable as part of training an RNN, and I’m getting this error message:

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.

Here is the code I’m running:

import tqdm as tqdmx
from tqdm import tqdm_notebook as tqdm

def train_rnn_model(network, criterion, optimizer, trainLoader, valLoader, n_epochs = 10, use_gpu = False):
    if use_gpu:
        network = network.cuda()
        criterion = criterion.cuda()
    # Training loop.
    for epoch in range(0, n_epochs):
        correct = 0.0
        cum_loss = 0.0
        counter = 0
        # Make a pass over the training data.
        t = tqdm(trainLoader, desc = 'Training epoch %d' % epoch)
        #initialize hidden state
        index, (imgIds, paddedSeqs, seqLengths) = next(enumerate(t))
        hiddenState = Variable(torch.Tensor(1, paddedSeqs.size(1), 512).zero_(), requires_grad=True)

        network.train()  # This is important to call before training!
        for (i, (imgIds, paddedSeqs, seqLengths)) in enumerate(t):
            batch_size = paddedSeqs.size(1) #gets batch size
            batchSequenceLength = paddedSeqs.size(0) #gets length of longest sequence
            # Wrap inputs, and labels into torch.autograd.Variable types.
            tmpseqs = paddedSeqs.clone()
            inputs = Variable(tmpseqs) #just the initial input from the training data
            #read in target/labels to remove [START], and add [END]
            #change it via numpy first
            tmp = paddedSeqs.clone().numpy()
            tmp = np.delete(tmp, (0), axis=0)
            #5001 is start, 5002 is end. 
            newrow = np.array([5002] * batch_size)
            tmp = np.vstack([tmp, newrow])
            #convert back to tensor
            paddedSeqs = torch.from_numpy(tmp)
            #create label Variable for later
            labels = Variable(paddedSeqs)

            if use_gpu:
                inputs = inputs.cuda()
                hiddenState = hiddenState.cuda()
            # Forward pass:
            predictions, hiddenState = network(inputs, hiddenState)
            #before passing to loss
            # Collapse the batch and sequence-length dimensions in order to use nn.Linear.

            predictions = predictions.view(batch_size * batchSequenceLength, -1)
            labels = labels.permute(1, 0)
            labels = ((labels.contiguous()).view(-1))
            if use_gpu:
                predictions = predictions.cuda()
                labels = labels.cuda()
            #print('preds', predictions)
            #print('labels', labels)
            loss = criterion(predictions, labels)

            # Backward pass:
            # Loss is a variable, and calling backward on a Variable will
            # compute all the gradients that lead to that Variable taking on its
            # current value.

            # Weight and bias updates.
            # logging information.
            cum_loss +=[0]
            max_scores, max_labels =
            correct += (max_labels ==
            counter += (inputs.size(0) * inputs.size(1))
            t.set_postfix(loss = cum_loss / (1 + i), accuracy = 100* correct / counter)

vocabularySize = len(trainData.vocabulary['word2id'])
model = TextGeneratorModel(vocabularySize)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr = 0.0005)

# Train the previously defined model.
train_rnn_model(model, criterion, optimizer, trainLoader, valLoader, n_epochs = 10)


I have no idea why I’m getting that error message, but I’m guessing it has something to do with the hiddenState Variable. When I don’t reassign the hiddenState variable in the forward pass call (i.e.

predictions, _ = network (inputs, hiddenState)
it trains without an error. Obvious, it doesn’t update the hiddenState though…

I’m completely stumped. What am I missing?


This thread probably has answer to your question: [Solved] Training a simple RNN

Let me know if you need any other help :slight_smile: Have fun pytorching.