CODEWITHSUNDEEP
pythonmachine-learningneural-networknlp
Himanshu Singhania
Himanshu Singhania

Reputation: 21

Prediction probability of each label in BERT classifier

I am trying to implement a method to evaluate prediction probability as is done with the help of scikit learn's

confidence = model._predict_proba_lr(x_count).max() * 100

Is there a way to evaluate the same using BERT Models?

Currently using Bert Base Uncased.
Using Ktrain Library(using Keras internally)
Reference Code : https://github.com/amaiya/ktrain/blob/master/examples/text/20newsgroups-BERT.ipynb

Upvotes: 2

Views: 4609

Answers (2)

blustax
blustax

Reputation: 489

In ktrain, you can pass return_proba=True to the predictor.predict method to output probabilities. See the text classification tutorial for more details.

Upvotes: 2

Robin
Robin

Reputation: 1599

I don't know much about the specific code you are using. But for classification (I guess this is what you're looking for), BERT uses a linear + soft-max layer over the last layer of the pre-trained BERT. So computing probabilities should be straightforward. I found an example with ktrain apparently where you can specify that you want probabilities (line 23).

Upvotes: 0

Related Questions