Load Huggingface Transformers model using TFAutoModel

[patricebechard]

Description of Problem: I want to use models provided by the HuggingFace transformers library that are not available when using the HFTransformersNLP Component.

Overview of the Solution: Right now, models are loaded using a dictionary of architectures in rasa/nlu/utils/hugging_face/registry.py. Using the AutoModel and AutoTokenizer features from the transformers library, we can get rid of these limitations, as well as not having to define a model_name in the config file. We can only use the model_weights info to select the proper model and weights of the wanted tokenizers and featurizers.

Examples (if relevant): Docs for how to use the "auto" features is available here

Blockers (if relevant):

Definition of Done:

• Changes are made to files in rasa/rasa/nlu/utils/hugging_face/
• Tests are added
• Feature described the docs
• Feature mentioned in the changelog

[sara-tagger]

Exalate commented:

sara-tagger commented:

Thanks for submitting this feature request

🚀
@JustinaPetr will get back to you about it soon!

✨

[ftarlaci]

Exalate commented:

ftarlaci commented:

Hi,

Is there any update on this enhancement? I would also like to use some of HF models that are currently not available in the HFTransformersNLP Component. Thank you!

[dakshvar22]

Exalate commented:

dakshvar22 commented:

The reason for not being able to use AutoModel and AutoTokenizer is because of a few implementation details on how HFTransformersNLP is interacting with the rest of the pipeline in Rasa Open Source. I would be curious to know what kind of models are you folks trying to use which are currently not supported.

[patricebechard]

Exalate commented:

patricebechard commented:

Hi @dakshvar22,

We were interested in looking into french models (FlauBERT, CamemBERT) for our project. Pretrained weights for these models are not available through the normal BertModel provided by HuggingFace, which was the main blocker on our side. We did some experimentation with the multilingual models, but they tend not to perform really well compared to french-only models. Instead of enabling only these specific models within HFTransformersNLP, we figured it would be best to have a more general way of dealing with models from the transformers library.

[dakshvar22]

Exalate commented:

dakshvar22 commented:

Thanks for the info @patricebechard . Are you looking at them for intent classification and entity recognition or some other tasks? We recently changed the default model weights that HFTransformersNLP uses. We now use a language agnostic model which gives much better sentence representations. This change will be part of the upcoming 2.1 release. I'll encourage you to try that as well.

[Ghostvv]

Exalate commented:

Ghostvv commented:

@dakshvar22 if we cannot use auto model, shall we close the issue?

[dakshvar22]

Exalate commented:

dakshvar22 commented:

The plan is to review this once we update the transformers dependency. Once we are sure that it is not possible to use auto model with the latest dependencies as well, we should close this issue.

[Ghostvv]

Exalate commented:

Ghostvv commented:

shall we move it to blocked then?

[mleimeister]

Exalate commented:

mleimeister commented:

@TyDunn Just to raise it to your awareness, I've started a side project to explore this. The current state is here: #10624

If I get to a point where it looks viable, I can check with Enable how this would play together with the scheduled TF/transformers upgrade.

[sync-by-unito]

➤ Maxime Verger commented:

💡 Heads up! We're moving issues to Jira: https://rasa-open-source.atlassian.net/browse/OSS.

From now on, this Jira board is the place where you can browse (without an account) and create issues (you'll need a free Jira account for that). This GitHub issue has already been migrated to Jira and will be closed on January 9th, 2023. Do not forget to subscribe to the corresponding Jira issue!

➡️ More information in the forum: https://forum.rasa.com/t/migration-of-rasa-oss-issues-to-jira/56569.