JonasGeiping / cramming

Cramming the training of a (BERT-type) language model into limited compute.
MIT License
1.29k stars 100 forks source link

Fix `save_training_checkpoint` #14

Closed JeanKaddour closed 1 year ago

JeanKaddour commented 1 year ago

I had to make these two minor changes to get save_intermediate_checkpoints to work.

The first fix is that my tokenizer object did not have a name attribute, but tokenizer.name_or_path exists.

The second fix is that there was no path variable in the save_training_checkpoint(...) function but a variable named directory.

After making these two changes, saving checkpoints worked.

JonasGeiping commented 1 year ago

Great, thanks!