Error: there are no graph nodes that require computing gradients

autograd
1 
                # wrap them in Variable
                inputs = Variable(inputs.cuda())
                labels = Variable(labels.cuda())

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                outputs = model(inputs)

                _, preds = torch.max(outputs, 1)

                preds = preds.float()
                loss = criterion(preds, labels)
                loss.backward()

I’m trying to solve a classifying problem using code above. But I got the error “RuntimeError: there are no graph nodes that require computing gradients”. How should I solve it? Does it due to the function"torch.max"?Thanks in advance.

2

Hi,

The argmax operation is not differentiable, if you check the two outputs of the max operation, you will see that the value has requires_grad=True (this one is differentiable) but the indices has requires_grad=False (this one is not).
You cannot learn this unfortunately, to do classification, you will need a criterion that works directly with the outputs values like nn.CrossEntropyLoss

1 Like