Nice work, but it seems the link is invalid. Could you check what is going on.

Besides, still waiting instructions to run the code and reproduce the results. Thanks.

I'm using python3.8.7 with the libraries in the requirement.txt

I have a problem when running: train_bert_binary_augmented_kilt.py

I got: ValueError: The 'target' has to be an integer tensor. at the following line: https://github.com/nicola-decao/KnowledgeEditor/blob/80b86e41a780161b5479862184d5d0b62f579acf/src/models/bert_binary_augmented_kilt.py#L257

Indeed batch["labels"] is a Tensor with dtype = torch.float32. I tried to cast the tensor to integer with batch['labels'].to(torch.int). It works but then i got:

ValueError: The 'preds' should be probabilities, but values were detected outside of [0,1] range.

So I changed logits with: torch.IntTensor([x > 0 for x in logits])

The same holds for these lines: https://github.com/nicola-decao/KnowledgeEditor/blob/80b86e41a780161b5479862184d5d0b62f579acf/src/models/bert_binary_augmented_kilt.py#L262-L265

Is this a viable fix? Thank you.

Can not run the code for train_bert_binary_augmented_kilt.sh First, it says ModuleNotFoundError: No module named 'pytorch_lightning.metrics'. Then, I import torchmetrics and change the original lines of pl.metrics.Accuracy() to torchmetrics.Accuracy(). The code can work. However, new problems occur in https://github.com/nicola-decao/KnowledgeEditor/blob/668aca6d01cc12e9a5fda17fe29a2334b6d0a359/src/models/bert_binary_augmented_kilt.py#L257 with the message below: ValueError: Thetargethas to be an integer tensor.

Please advise how to debug it.

Hii, I just wanted to highlight that when using the 'ddp' with the following command: python scripts/train_bert_binary_augmented_kilt.py --gpus 1 --accelerator ddp --num_workers 32 --batch_size 256 --max_steps 50000 --divergences kl

There is an error: RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. This error indicates that your module has parameters that were not used in producing loss. You can enable unused parameter detection by (1) passing the keyword argument find_unused_parameters=True to torch.nn.parallel.DistributedDataParallel; (2) making sure all forward function outputs participate in calculating loss. If you already have done the above two steps, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's forward function. Please include the loss function and the structure of the return value of forward of your module when reporting this issue (e.g. list, dict, iterable).

But after removing the --accelerator ddp, it is running fine. Is there any workaround to use it with DDP? I have tried setting find_unused_parameters=True to torch.nn.parallel.DistributedDataParallel;. But somehow this is also not working.

Hi: Thanks for your excellent work and providing the nice codebase. There is one question I have been bothered by. When testing retrain acc, do you test this while updating the model parameters? And do you update the model parameters with batch[-1:] and then use batch[:-1] to test the stability of the model when testing?

Any help is appreciated.