I used an example of using BERT to classify reviews, described at the link. The code is written for using the CPU and it works fine, but slowly. In Colab Google, with a multilingual model, one epoch is considered 4 hours for me. If I replace the CPU with the CUDA everywhere in the code, then the error that YOU met with appears. I followed the guidelines given in the link, but then another error appears:
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-3-0b35a5f74768> in <module>()
268 'labels': batch[2],
269 }
--> 270 inputs.to(device)
271 outputs = model(**inputs)
272
AttributeError: 'dict' object has no attribute 'to'