In my PyTorch train iterator, how do I resolve the ValueError: only one element tensors can be converted to Python scalars?

Question

How do I solve the ValueError: only one element tensors can be converted to Python scalars ?

I am closely following a tutorial on building a Question Answering Bot in PyTorch . However, at training, my code is unable to save the checkpoints, giving me aforementioned ValueError. The error happens at torch.save(torch.tensor(train_loss_set), os.path.join(output_dir, 'training_loss.pt'))

Below is my code corresponding to the train iterator:

num_train_epochs = 1

print("***** Running training *****")
print("  Num examples = %d" % len(dataset))
print("  Num Epochs = %d" % num_train_epochs)
print("  Batch size = %d" % batch_size)
print("  Total optimization steps = %d" % (len(train_dataloader) // num_train_epochs))

model.zero_grad()
train_iterator = trange(num_train_epochs, desc="Epoch")
set_seed()

for _ in train_iterator:
    epoch_iterator = tqdm(train_dataloader, desc="Iteration")
    for step, batch in enumerate(epoch_iterator):
        if step < global_step + 1:
            continue
            
        model.train()
        batch = tuple(t.to(device) for t in batch)
        
        inputs = {'input_ids':       batch[0],
                'attention_mask':  batch[1], 
                'token_type_ids':  batch[2],  
                'start_positions': batch[3], 
                'end_positions':   batch[4]}
        
        outputs = model(**inputs)
        
        loss = outputs[0]
        train_loss_set.append(loss)
        loss.sum().backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)

        tr_loss += loss.sum().item()
        optimizer.step()
        model.zero_grad()
        global_step += 1

        if global_step % 1000 == 0:
            print("Train loss: {}".format(tr_loss/global_step))
            output_dir = 'checkpoints/checkpoint-{}'.format(global_step)
            if not os.path.exists(output_dir):
                os.makedirs(output_dir)
            model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
            model_to_save.save_pretrained(output_dir)
            torch.save(torch.tensor(train_loss_set), os.path.join(output_dir, 'training_loss.pt'))
            print("Saving model checkpoint to %s" % output_dir)

Edit print(train_loss_set[:10]) returns the following:

[tensor([5.7099, 5.7395], device='cuda:0', grad_fn=<GatherBackward>), tensor([5.2470, 5.4016], device='cuda:0', grad_fn=<GatherBackward>), tensor([5.1311, 5.0390], device='cuda:0', grad_fn=<GatherBackward>), tensor([4.4326, 4.8475], device='cuda:0', grad_fn=<GatherBackward>), tensor([3.4740, 3.9955], device='cuda:0', grad_fn=<GatherBackward>), tensor([4.8710, 4.5907], device='cuda:0', grad_fn=<GatherBackward>), tensor([4.4294, 4.3013], device='cuda:0', grad_fn=<GatherBackward>), tensor([2.7536, 2.9540], device='cuda:0', grad_fn=<GatherBackward>), tensor([3.8989, 3.3436], device='cuda:0', grad_fn=<GatherBackward>), tensor([3.3534, 3.2532], device='cuda:0', grad_fn=<GatherBackward>)]

Could this have to do with the fact that I'm using DataParallel?

Answer 1

It's a weird behavior of pytorch.

Basically you can't create a Tensor using a list(s) of Tensors.

But there's 3 things you can do.

1. Is, you don't need torch.tensor when saving a list of tensors so this should work.

torch.save(train_loss_set, os.path.join(output_dir, 'training_loss.pt'))

2. Use torch.stack instead.

torch.save(torch.stack(train_loss_set), os.path.join(output_dir, 'training_loss.pt'))

3. This kinda counter intuitive but you can convert the tensors inside to ndarray . And you can use torch.tensor

train_loss_set.append(loss.cpu().detach().numpy())