Fine-tune Sentence Transformers the OG Way (with NLI Softmax loss)

Поділитися
Вставка
  • Опубліковано 22 січ 2025

КОМЕНТАРІ • 15

  • @eugenechua7239
    @eugenechua7239 Рік тому +1

    Hey james, will you be sharing the notebook?

  • @NitishKumar-pg6qw
    @NitishKumar-pg6qw Рік тому

    How can we put two sentences closer in vector space ? Can we use this approach?

  • @benp.9225
    @benp.9225 Рік тому

    In the Paper they talk about "trainable" weights for the softmax classification. Can you not access those weights after training from the SBERT Modell (e.g. to make predictions)? Or is the Cosine-Distance the only way to use the model?

  • @yangwang9688
    @yangwang9688 2 роки тому

    So you don't update the weight of `ffnn`?

  • @souravsamanta3354
    @souravsamanta3354 2 роки тому

    @james Briggs: I am not clear about this aspect: When you are doing the training using the Pytorch way, you perform concatenation operations explicitly in your code... but if training is done using the Sentence Transformer framework, we don't see any concatenation... Is it being handled automatically by the library ???

    • @jamesbriggs
      @jamesbriggs  2 роки тому +1

      Yes that’s right, the library handles it automatically

    • @souravsamanta3354
      @souravsamanta3354 2 роки тому

      @@jamesbriggs Thank you so much for the clarification and the wonderful tutorial... I have one more followup question... In the Pytorch implementation, you have added a FFNN after the concatenated tensor... but for Sentence BERT there is no dense layer after the pooling operation... Is my understanding correct ?

  • @пчелобавурод
    @пчелобавурод Рік тому

    Why don't you freeze the layers?

  • @1lyf
    @1lyf 2 роки тому

    Hello James,
    I tried to find tune the model. I have a 3050 with 4gb and when I try to fit the model with batch size of 16 i get cuda out of memory error. I just ran your exact code shown in the video. The difference is that my data is just 5000 rows. Could you please advise how to solve this?

    • @jianfengzhu2100
      @jianfengzhu2100 2 роки тому

      I had the same issue. I reduced the data size and then it went through. I am wondering how to fix the problem.

  • @holthuizenoemoet591
    @holthuizenoemoet591 Рік тому

    Love this intro!

  • @zacharyadams3772
    @zacharyadams3772 2 роки тому

    Can you make that title card into a png I can use for my background?

  • @cameron.willis
    @cameron.willis 3 роки тому

    Another great video! Thanks James! I'm hoping my team can take your Udemy course soon!