fine tune mBERT for sentiment analysis using ktrain

[mcormick123]

Hi, thanks for your great work! I've been using ktrain to fine tune multilingual bert and got an amazing result! i've looked on several tutorials already but some are vague, and i really want to understand more how ktrain works.

Would like to ask if i'm doing the process correctly to fine tune the mBERT from huggingface using ktrain? Here's my process:

1. Load dataset file which contains text and label
2. I've read a blog that suggested to get the 95th percentile to get the maxlen, I thought you should tokenize it first and get the maxlen from there? What's the default in ktrain ?
3. Split the train, val, and test sets
4. Execute:

   MODEL_NAME = 'bert-base-multilingual-uncased'
   // in setting the maxlen here, i used the 95th percentile
   t = text.Transformer(MODEL_NAME, maxlen = 250, class_names = ['0','1','2'])
   train_final = t.preprocess_train(X_tr, y_tr, mode='train')
   val_final = t.preprocess_test(X_val, y_val)
   model = t.get_classifier()
   learner = ktrain.get_learner(model, train_data=train_final, val_data=val_final, batch_size=16)
   learner.fit_onecycle(2e-5, 2)

5. Use the learner to predict the test set and evaluate its accuracy and f1 score
6. Use the learner to predict unseen data. Should I preprocess first the data? or is it good to go to the predictor?

Please, actually, i'm about to report this to my professor as a class output and imagining my prof questioning me stuffs about my code terrifies me already HAHA! explanation would be fully appreciated! thank you so much!

[amaiya]

Hello: The default maxlen is 400 I believe, but running it once to observe the stats and then running it again choosing 250 (which you said was 95% percentile) sounds reasonable.

To predict on unseen data in an application, a Predictor instance is typically used. To evaluate the test set, you can also use the Learner instance instead of the Predictor.