PT Pump Up Client

Use Cases

Train Semantic Role Labeller

from pt_pump_up.benchmarking import TrainerFactory
from pt_pump_up.benchmarking.training_strategies.SemanticRoleLabellingStrategy import SemanticRoleLabellingStrategy
from datasets import load_dataset

propbank_br = load_dataset("liaad/Propbank-BR", "flatten")

for model_name in ['neuralmind/bert-base-portuguese-cased', 'neuralmind/bert-large-portuguese-cased', 'PORTULAN/albertina-100m-portuguese-ptpt-encoder' 'PORTULAN/albertina-900m-portuguese-ptpt-encoder']:

    repository_name = f"SRL-{model_name.split('/')[1]}"

    trainer = TrainerFactory.create(
        nlp_task="SRL",
        repository_name=repository_name,
        model_name=model_name,
        label_names=propbank_br['train'].features['frames'].feature.names,
        max_epochs=30,
        lr=1e-5,
        train_dataset=propbank_br['train'],
        eval_dataset=propbank_br['test'],
    )

    trainer.train()

    SemanticRoleLabellingStrategy.create_pipeline(
        hf_repo=repository_name,
        model=trainer.model,
        tokenizer=trainer.tokenizer,
    )

Train Sentiment Analyser

from datasets import load_dataset

# Usage of the PT-Pump-Up library is not mandatory, but it will make your life easier.
# It reuses code previously developed for similar NLP tasks. That is already tested and validated.
from pt_pump_up.benchmarking import TrainerFactory


# Load dataset from huggingface/datasets.
# StanfordNLP/IMDB is a dataset for sentiment analysis in English.
# It is as simple as it can be. It has only two columns: text and label.
imdb = load_dataset("stanfordnlp/imdb")

# There are 4 transformers models that can be adapted for sentiment analysis in Portuguese:
# - neuralmind/bertimbau version base (110M) and large (335M) (Bigger/Computational Expensive Architecture)
# - PORTULAN/albertina version 100m and 900m. Avaiable in PT-PT and PT-BR
for model_name in ['neuralmind/bert-base-portuguese-cased', 'neuralmind/bert-large-portuguese-cased', 'PORTULAN/albertina-100m-portuguese-ptpt-encoder' 'PORTULAN/albertina-900m-portuguese-ptpt-encoder']:
    # You should specify the repository name for each model to be trained.
    # It will be available in the Hugging Face Hub under that name.
    # Ex: f"dataset-SRL-{model_name.split('/')[1]}" produced https://huggingface.co/liaad/propbank_br_srl_bert_base_portuguese_cased
    repository_name = "<<REPOSITORY_NAME>>"

    trainer = TrainerFactory.create(
        # Sentiment Analysis is a Text Classification task.
        nlp_task="Text Classification",
        repository_name=repository_name,
        model_name=model_name,
        # label_names is a list of strings with the possible labels in the dataset.
        # If the dataset is correctly loaded, you can access the label names with dataset['train'].features[<<LABEL_COLUMN_NAME>>].feature.names
        # In this case, the label column name is 'label'.
        # If not proprely loaded, you can use a list of strings with the possible labels. Ex: ['Positive', 'Negative'], assuming that the labels are 'Positive' and 'Negative'.
        label_names=imdb['train'].features['label'].names,
        max_epochs=30,
        lr=1e-5,
        train_dataset=imdb['train'],
        eval_dataset=imdb['test'],
    )

    trainer.train()