Remove the restriction of Bert Based models in TextClassification Task.

[Atharva-Phatak]

🚀 Feature

Remove the restriction of Bert Based models in TextClassification Task.

Motivation

For classification I would like to use flash but I want independence to differ from bert based models.

Pitch

• I would like to propose that TextClassifier model, accepts "nn.Module" object for model argument rather than name of a bert based model. This will give the user the ability to create custom models which could have custom layers.
• Another suggestion is instead of accepting again backbone name for tokenizer, let the user create the tokenizer from hugging face.
• If the above are implemented essentially, we wont require transformers in requirements.txt, thus reducing additional dependency.

Note: This goes without saying, I would like to work on a PR if this issue is accepted. 😃

[edgarriba]

@Atharva-Phatak what you mention makes total sense. Before asking you to send a PR - I would suggest to check this active PR about adding a token_classification task and add some comments there. See: https://github.com/PyTorchLightning/lightning-flash/pull/275

[Atharva-Phatak]

@edgarriba Thank you very much I will go through the mentioned PR.

[Atharva-Phatak]

@edgarriba I did take a look at #275, it seems with the new release the branch is outdated. So what should I do now ? Should I create a PR for this issue ?

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.