data.py

import torch
from datasets import load_dataset, Dataset
from transformers import BertTokenizerFast
import pandas as pd
from imblearn.under_sampling import RandomUnderSampler
import logging
import os


def balance_data(dataset):
    df = dataset.to_pandas()
    
    logging.info(f"Balancing {df['label'].value_counts()}")

    rus = RandomUnderSampler(random_state=42, replacement=True)

    X_resampled, y_resampled = rus.fit_resample(
        df['text'].to_numpy().reshape(-1, 1), df['label'].to_numpy())
    
    df = pd.DataFrame(
        {'text': X_resampled.flatten(), 'label': y_resampled})
    
    logging.info(f"After balancing: {df['label'].value_counts()}")

    return Dataset.from_pandas(df)


def tokenize(dataset):
    tokenizer = BertTokenizerFast.from_pretrained("neuralmind/bert-large-portuguese-cased")

    dataset = dataset.map(lambda example: tokenizer(
        example["text"], truncation=True, padding="max_length", max_length=512))
    
    return dataset

# This function supports the Notebook version of LID. No usage elsewhere.
def tokenize_single_document(text):
    tokenizer = BertTokenizerFast.from_pretrained("neuralmind/bert-large-portuguese-cased")

    return tokenizer(text, truncation=True, padding="max_length", max_length=512)

def load_dataloader(domain):

    logging.info(f"Loading {domain} dataset...")
    
    if domain == 'dslcc':
        dataset = load_dataset("arubenruben/portuguese_dslcc")
    else:
        dataset = load_dataset("Random-Mary-Smith/port_data_random", domain)
    
    DEBUG = (os.getenv('DEBUG', 'False') == 'True')

    dataset['train'] = balance_data(dataset['train'])

    dataset['test'] = dataset['test'].select(range(min(len(dataset['test']), 10_000)))
    
    for split in ['train', 'test']:    
        if DEBUG:
            logging.info("DEBUG MODE: Loading only 100 samples")
            dataset[split] = dataset[split].select(range(min(len(dataset[split]), 50)))

    dataset = tokenize(dataset)
    dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'label'])

    # Create Dataloaders
    train_dataloader = torch.utils.data.DataLoader(dataset['train'], batch_size=int(os.getenv('BATCH_SIZE')), shuffle=True)
    test_dataloader = torch.utils.data.DataLoader(dataset['test'], batch_size=int(os.getenv('BATCH_SIZE')), shuffle=False)

    return train_dataloader, test_dataloader