Thanks for the review @SagiPolaczek !
Even though we talked answering here for future reference:
the stored arguments are not useful in a common scenario (such as continue the training - cause you to run the exact same script anyway and recreate the arguments). model on the other hand is useful to load a model for inference.
Yes, we need unittest. I will open an issue.
since it's three lines, I'm not sure.
I wanted to support the simplest building blocks in data pipeline (such as a function that gets sample_dict) and not to force all to inherit from OpBase.
Thanks for the review @SagiPolaczek ! Even though we talked answering here for future reference: