MetaCAT BERT update to tutorials 4.1 and 4.2

[shubham-s-agarwal]

Updates to Tutorials 4.1 and 4.2: Integration of BERT Implementation

• Tutorial 4.1: No changes were necessary since BERT utilizes its own pre-trained tokenizer. A note has been added advising users to proceed directly to Tutorial 4.2 if using BERT.

• Tutorial 4.2: BERT implementation has been added, showing training on the given dataset. Unlike BiLSTM, which fine-tunes a previously trained model pack, BERT is trained from scratch.

[mart-r]

4.1 looks good. Don't think there needs to be more in here for that.

4.2

• The MetaCAT configuration and the Train MetaCAT (sub)section are now duplicated. It would be good to have these appear once. Otherwise we'll end up having to maintain 2 copies of the same text/code.
  The two MetaCAT instances (BERT vs BiLSTM) should provide the exact same interface anyway. The only difference is that some config entries don't affect one or the other.
• Perhaps we can trim the included output a little? I know it's already long in the existing one, but just doing this comparison it was pretty annoying to have to scroll around the sizeable chunks of output.
• There is no description of the model_variant config option in the tutorial

EDIT:
Also, regarding 4.2. The changes to the config in the first code section in the For BERT section is the reason I was talking about creating a static method for getting a BERT-based meta-cat with defaults.
I.e in ConfigMetaCAT we could have something along the lines:

    @classmethod
    def get_default_bert_config(cls, category_name: str, model_name: str = 'bert', nclasses: int = 2) -> 'ConfigMetaCAT':
        config = cls()
        cls.model.model_name = model_name
        cls.model.nclasses = nclasses
        cls.general.category_name = category_name
        return cls

[shubham-s-agarwal]

For 4.2:

• MetaCAT configuration and the Train MetaCAT have subsections for BERT and Bi-LSTM in them.
• Since model_variant isn't commonly changed, I haven't added it in the tutorial. There is a link to all config variables in the section.
• Regarding the config section for BERT, we have to explicitly change the model_name, category_name and other variables since we are starting the training from scratch. For Bi-LSTM, we are loading the model pack; if we were training Bi-LSTM from scratch, we would have the do the same.

[mart-r]

• Since model_variant isn't commonly changed, I haven't added it in the tutorial. There is a link to all config variables in the section.

But you're using it in the tutorial - passing it to TokenizerWrapperBERT.load. Someone going through the tutorial won't have any idea what this is or what value it holds.

In 4.2 I don't think we need 2 different (almost identical) mc.train_from_json code blocks. We could just have 1 and get the suffix for the save file from config_metacat.model.model_name automatically.

Other than that, I think this looks good.

[shubham-s-agarwal]

Fixed both!

[mart-r]

Thought I already approved yesterday, but I think I was waiting for GHA workflow just in case.

Looks good to me.