[HuggingFace] AttributeError("'str' object has no attribute 'item'",)

Describe the bug I get this error every time I try fine tuning BERT with W&B: AttributeError("'str' object has no attribute 'item'",) …whether it’s your example in Google Colab or it’s my own fine-tuning experiment with my own machine & dataset.

To Reproduce

1. Go to https://colab.research.google.com/drive/1SQ-FOgji8AiyrQ08sIVfDiA8OUw4bC12?usp=sharing#scrollTo=JL9OqcavmYVb
2. Click Run All
3. See error

Screenshots

Additional context Python 3.7.9 wandb 0.10.12 torch 1.7.0+cu110 transformers 4.1.1

[acd424]

I found the the docs on migrating from previous packages useful - paragraph 4.

https://huggingface.co/transformers/migration.html

I added ‘return_dict=False’ to my model argument and it worked as it should.

[MehmetKUTAN]

When running the Bert algorithm, I have the same problem, I get the AttributeError (“‘str’ object has no attribute ‘item’”,) error in the loss.item () statement. solution: Since there is an unequaled value to the total_train_loss value, When I set the value for “loss” I did “return_dict = False”. Problem solved.