Skip to content

Latest commit

 

History

History
89 lines (55 loc) · 3.84 KB

README.md

File metadata and controls

89 lines (55 loc) · 3.84 KB
Raw

How to boost BERT performances

Executive summary

In our blog post we have explored a way to boost BERT performances when we deal with a small dataset for a classification problem. We used a sample of hatred speech dataset which we collected on Kaggle.

You can take a look at how we built the datasets (train and test) and get final results in the notebooks folder. We also provide the colab to do the experiment on your own.

The datasets are also in the data/ folder.

Results of the different experiments

Training and testing set

Dataset #Texts #Hatred
Train 100 50
Test 10,000 5,000

Metrics

Models F1 P R
Random 50.73(0.00) 50.44(0.00) 51.02(0.00)
bow 56.40(0.00) 63.07(0.00) 51.00(0.00)
FSL distilBERT 64.41(4.65) 65.34(2.87) 64.80(10.39)
FSL_mean distilBERT 67.86(0.00) 67.75(0.00) 67.98(0.00)
distill distilBERT experts 68.34(0.00) 67.58(0.00) 69.12(0.00)
FSL XtremeDistilTransformers 63.05(1.40) 63.70(1.69) 62.46(4.00)
FSL_mean XtremeDistilTransformers 63.87(0.00) 64.56(0.00) 63.20(0.00)
distill XtremeDistilTransformers experts 64.15(0.00) 64.52(0.00) 63.78(0.00)

Finally, we provide some command lines to apply the code straight from this repo.

Pre-processing

To pre-process one text you should just apply copy paste the following CLI:

python run_pre_processing.py --name test_set

Of course this means you have a text variable in your data/my_dataset.csv otherwise you should add variables --text_var my_text_variable.

You'll find a data/my_dataset_pp.csv in the data folder ready for training !

Training a BERT model

Now you can try to train the model with a transformer architecture. You need to have the dataset already processed as the data/my_dataset_pp.csv. To do so you should copy paste the following CLI:

python run_training.py --name my_dataset_pp

You can take a look at all different parameters, but normally the default hyperparameters are already set values used in the blog post. You will have a models/my_dataset_pp.h5 model trained and saved that you can use for inference.

Infer from a BERT model

Now you can use your models/my_dataset_pp.h5 for inferences. If you have a data/inference_set.csv with a text variable in it, you can apply your model with the following CLI:

python run_inferences.py --name_data inference_set --name_model my_dataset_pp

To be sure that the models/my_dataset_pp.h5 has been loaded you should see a: INFO:root:Model /YOUR_PATH_TO_THE_MODEL/my_dataset_pp.h5 restored Otherwise you'll see: WARNING:root:Model /YOUR_PATH_TO_THE_MODEL/my_dataset_pp.h5 not found WARNING:root:If training: new model from scratch WARNING:root:If classifying: the configuration does not fit the architecture and this model is not trained yet!

And you're inferences will be from a random model !

Then you'll have a data/inference_set_preds.csv file with all the predicted data.

Distil a model from experts inferences

Now let's say you have created the augmented dataset, data/random_set_with_inferences_pp.csv which is already pre-processed and have a variable prob_hatred in this dataset as the mean of all you experts predictions, you can train a model from it with the following CLI:

python run_training_distilled_BERT.py --name random_set_with_inferences_pp

You can take a look at all different parameters, but normally the default hyperparameters are already set values used in the blog post. You will have a models/random_set_with_inferences_pp_distilled.h5 model trained and saved that you can use for inference.