DeepPavlov(rubert-base-cased) for Rasa

Rasa Open Source
1

Hi! Can you help me pliz, Is there an easy way to use DeepPavlov/rubert-base-cased from Huggingface with Rasa? This is a BERT model with russian language weights.

2

@SemionushkinDenis1988 Thanks for the question! Yes, you can use the BERT model with the specified russian language weights. Although, I see that the model checkpoint is for pytorch backend which we don’t support as of now. However, you can convert the pytorch checkpoint into a Tensorflow checkpoint and then load the converted checkpoint in your pipeline. Please refer to this thread to understand how you can do this.

1 Like
3

Thank you!

4

One more question, how can i feed converted checkpoints to rasa?

5

I have an error like this: image no cache

6

I don’t think that’s an error. After this step the transformers lib should have tried to download the model. and put it in cache. You can specify a cache_dir parameter in the configuration to use a custom path for caching models from HF Transformers.

7

You can use the path to the converted checkpoints as the value for model_weights parameter.

8

what if i convert .pt model to .h5 format and add link at modeling_tf_bert.py?

i tried it and it works

9

Hi, did you finally managed to use ruBert?