Make sentiment BERT feature run faster

[xehu]

Currently, the sentiment BERT feature is memory and time inefficient. It actually makes it hard to run minimal examples such as this one: https://colab.research.google.com/drive/1y4lIl3aoEFCMTK9Kn7R-M-sOEa_BI8Rz?usp=sharing

(The FB here fails due to running out of RAM on CoLab).

We might want to consider taking some steps to make this more efficient, including:

moving operations to GPU if possible

import torch

def get_sentiment(texts):
    """
    Analyzes the sentiment of the given list of texts using a BERT model and returns a DataFrame with scores for positive, negative, and neutral sentiments.

    :param texts: The list of input texts to analyze.
    :type texts: list of str
    :return: A DataFrame with sentiment scores.
    :rtype: pd.DataFrame
    """
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model_bert.to(device)

    texts_series = pd.Series(texts)
    non_null_non_empty_texts = texts_series[texts_series.apply(lambda x: pd.notnull(x) and x.strip() != '')].tolist()

    if not non_null_non_empty_texts:
        return pd.DataFrame(np.nan, index=texts_series.index, columns=['positive_bert', 'negative_bert', 'neutral_bert'])

    encoded = tokenizer(non_null_non_empty_texts, padding=True, truncation=True, max_length=512, return_tensors='pt').to(device)

    with torch.no_grad():
        output = model_bert(**encoded)

    scores = output[0].cpu().detach().numpy()
    scores = softmax(scores, axis=1)

    sent_dict = {
        'positive_bert': scores[:, 2],
        'negative_bert': scores[:, 0],
        'neutral_bert': scores[:, 1]
    }
    
    non_null_sent_df = pd.DataFrame(sent_dict)

    sent_df = pd.DataFrame(np.nan, index=texts_series.index, columns=['positive_bert', 'negative_bert', 'neutral_bert'])
    sent_df.loc[texts_series.apply(lambda x: pd.notnull(x) and x.strip() != ''), ['positive_bert', 'negative_bert', 'neutral_bert']] = non_null_sent_df.values

    return sent_df

releasing memory explicitly

del encoded, scores, non_null_sent_df
torch.cuda.empty_cache()

taking a batching approach, like we do with SBERT vectors.