Model Saving Directory Not Existing

[superctj]

• fibber version: 0.2.5
• Python version: 3.7
• Operating System: Linux

Description

I wanted to train a Bert classifier on the IMDB dataset with the setup in fibber.metrics.bert_classifier. When the program executes model.save_pretrained(ckpt_path) in load_or_train_bert_clf, it triggers an assertion error in transformers library and I believe it is because ckpt_path does not exist.

A screenshot of the traceback is attached for your reference.

What I Did

To replicate the error, simply load the IMDB dataset from fibber and train a Bert classifier as shown below

My workaround is to create ckpt_path when it is not detected.