Based on debugging done during the labs
Q: When trying to pip install the
transformers
ortokenizers
library, I get an error in building thebdist_wheel
oftokenizers
A: Try to install the Rust compiler using the command curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
, reopen a different shell session and try the installation again. If it doesn't work, you can try to install from source.
*Q: When I try to load a
pipeline
or atransformer
, the loading stops and I get an errorAttributeError: 'FloatProgress' object has no attribute 'style'
.
A: Try to update your pip and ipywidgets versions: pip install --upgrade pip ipywidgets
*Q: In W3E, when I try to retrain the BPE tokenizer changing the vocabulary size, I get an error
PanicException: no entry found for key
.
A: The error is likely to be due to the procedure of removing subwords that are not frequent anymore after a BPE merge, which may cause gaps in the vocabulary when iterated. Simply redefining a new tokenizer object and using that for training should solve the issue.