Error predicting with language model: 'L' object has no attribute 'join'

In the 10_nlp notebook I run into an error when predicting on the language model which has been fit_one_cycle for 1 epoch:

(Compared to the original notebook I moved the prediction cell up to save some time, but that shouldn’t have an effect.)

When checking the type of learn.dls.train_ds.tokenizer it is a L consisting of a fastai2.text.core.SpacyTokenizer and fastai2.text.core.Tokenizer, so in fact does not support .join

Any ideas where this might be going wrong?

I checked the source code and it seems the source has changed to:

sep = self.dls.train_ds.tokenizer[-1].sep
return sep.join(decoder(tokens))

where self.dls.train_ds.tokenizer[-1].sep indeed gives a string.

I am working on Paperspace, any idea why this change hasn’t been propagated there?

Did you make sure to git pull from the upstream branch?

I am pretty sure that git pull only pulled the notebooks but not the fastai codebase.

Given that the error is thrown in the codebase, I hadn’t pulled the notebooks. Nevertheless, giving it a try now …

EDIT: The issue persists.