CODE SetFit w/ SBERT for Text Classification (Few-Shot Learning) multi-class multi-label (SBERT 44)

[@azinposhtyar9222]

Thanks for the great tutorial, it's awesome! Would you please share the google Colab notebook for the videos?

[@brianmvukwe5506]

How do you use this with an input of 2 sentence(for duplicate detection). How do the parameters in the SetFitTrainer's column_mapping change?

[@TadjouLionelIT]

Thansk for your video about SetFit. I've also try playing with it for Multiclass (10 classes) classifier on speech acts recognition (MRDA dataset) with 78% of accurary. But I tried to use SoftMaxLoss for the same multiclass finetuning and I got some errors. Could please make a video explaining when and how to use SoftMaxLoss and TripletLoss with Setfit. Thanks in advance!

[@code4AI]

Will take some time for new videos ... For immediate clarification two sources of information: towardsdatascience.com/triplet-loss-advanced-intro-49a07b7d8905 and Classification Loss Functions: Comparing SoftMax, Cross Entropy, and More: wandb.ai/capecape/classification-techniques/reports/Classification-Loss-Functions-Comparing-SoftMax-Cross-Entropy-and-More--VmlldzoxODEwNTM5 with referenced: Losses Learned -- Optimizing Negative Log-Likelihood and Cross-Entropy in PyTorch (Part 1) sebastianraschka.com/blog/2022/losses-learned-part1.html

[@TadjouLionelIT]

@@code4AI Thanks for your reply! I'll looked at those links for better understanding.

[@MirrorNeuron]

@04:17 I am getting this error in Line 4 ImportError: cannot import name 'SetFitModel'

[@code4AI]

It works just fine. After you "pip install setfit" and import the model, it works (as seen in the live video) ... May depend on your specific virtual environment, I just used the free COLAB NB form Google.

[@MirrorNeuron]

@@code4AI I am in Py36 in AWS notebook instance

[@Opera-1553]

@@MirrorNeuron you need to create a virtual environment (say pytho3.8.5) and use that to add a new kernel in your AWS Sagemaker, and it will work.

[@korilifs]

Can I only get a similarity scores of my set to a certain prompt with this system rather then classifying sentences according to predefined labels?

[@code4AI]

SetFit being a few-shot learning approach, I showed in the video at about 11 min that when you run inference on your trained model, you get a classification for your prompt. We trained it on a sentiment classifier and when we prompt two sentences, we do not get a similarity score to a set, but a classification: either sentiment "good" or sentiment "bad", according to our predefined labels (eg 0 or 1) in our training dataset. So you end up with a trained model, that you can apply in any familiar way you want.

[@BatBallBites]

Initially you choose 32 samples for few shot learning but in your training loop it is showing that you are using 3480 samples for training ?

[@Hipno70]

its a crapy tutorial