app.py

import logging
import os
from typing import List, Tuple

import gradio as gr
import pandas as pd
import spacy
import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

try:
    nlp = spacy.load("pt_core_news_sm")
except Exception:
    os.system("python -m spacy download pt_core_news_sm")
    nlp = spacy.load("pt_core_news_sm")

model = AutoModelForTokenClassification.from_pretrained("Emanuel/porttagger-news-base")
tokenizer = AutoTokenizer.from_pretrained("Emanuel/porttagger-news-base")
logger = logging.getLogger()
logger.setLevel(logging.DEBUG)


def predict(text, nlp, logger=None) -> Tuple[List[str], List[str]]:
    doc = nlp(text)
    tokens = [token.text for token in doc]

    logger.info("Starting predictions for sentence: {}".format(text))

    input_tokens = tokenizer(
        tokens,
        return_tensors="pt",
        is_split_into_words=True,
        return_offsets_mapping=True,
        return_special_tokens_mask=True,
    )
    output = model(input_tokens["input_ids"])

    i_token = 0
    labels = []
    scores = []
    for off, is_special_token, pred in zip(
        input_tokens["offset_mapping"][0],
        input_tokens["special_tokens_mask"][0],
        output.logits[0],
    ):
        if is_special_token or off[0] > 0:
            continue
        label = model.config.__dict__["id2label"][int(pred.argmax(axis=-1))]
        if logger is not None:
            logger.info("{}, {}, {}".format(off, tokens[i_token], label))
        labels.append(label)
        scores.append(
            "{:.2f}".format(100 * float(torch.softmax(pred, dim=-1).detach().max()))
        )
        i_token += 1

    return tokens, labels, scores


def text_analysis(text):
    tokens, labels, scores = predict(text, nlp, logger)
    pos_count = pd.DataFrame(
        {
            "token": tokens,
            "etiqueta": labels,
            "confiança": scores,
        }
    )
    pos_tokens = []
    for token, label in zip(tokens, labels):
        pos_tokens.extend([(token, label), (" ", None)])

    output_highlighted.update(visible=True)
    output_df.update(visible=True)

    return {
        output_highlighted: output_highlighted.update(visible=True, value=(pos_tokens)),
        output_df: output_df.update(visible=True, value=pos_count),
    }


css = open("style.css").read()
top_html = open("top.html").read()
bottom_html = open("bottom.html").read()

with gr.Blocks(css=css) as demo:
    gr.HTML(top_html)
    text = gr.Textbox(placeholder="Enter your text here...", label="Input")
    examples = gr.Examples(
        examples=[
            [
                "A população não poderia ter acesso a relatórios que explicassem, por exemplo, os motivos exatos de atrasos em obras de linhas e estações."
            ],
            ["Filme 'Star Wars : Os Últimos Jedi' ganha trailer definitivo; assista."],
        ],
        inputs=[text],
        label="Select an example",
    )
    output_highlighted = gr.HighlightedText(label="Colorful output", visible=False)
    output_df = gr.Dataframe(label="Tabular output", visible=False)
    submit_btn = gr.Button("Send")
    submit_btn.click(
        fn=text_analysis, inputs=text, outputs=[output_highlighted, output_df]
    )
    gr.HTML(bottom_html)


demo.launch(debug=True)