app.py

import gradio as gr
from keras.models import load_model
from transformers import TFBertModel, TFXLMRobertaModel
import numpy as np
import tensorflow as tf
from transformers import AutoTokenizer

app_title = "Portuguese Counter Hate Speech Detection (NFAA)"

app_description = """
This app is the culmination of the kNOwHATE consortium project, which aimed to tackle Online Hate Speech in the Portuguese comunity. It serves as an user-friendly interface to classify text and identify instances of Hate Speech.
This app leverages state-of-the-art Natural Language Processing models developed in the scope of this project to classify harmful text.
Select a model from the dropdown menu and input your text to see the classification results. Explore the examples of Hate Speech and Non-Hate Speech offered, and join us in fostering a safer and more respectful online community.
For more information about the kNOwHATE project and its initiatives, visit our website [here](https://knowhate.eu) and to explore and use these models visit our Hugging Face page [here](https://huggingface.co/knowhate).
"""

# 1 0 2
app_examples = [
    ["Essa gente tem é de deixar de ser apaparicada pelo Estado e começar a cumprir os seus deveres como cidadãos", 
     "Nepia o que faz com que as pessoas generalizem é o ódio intrínseco que têm contra uma etnia, ng é responsável pela sua xenofobia", 
     "knowhate/twt-bertimbau/twt-bb-b16e5-avg767.keras"],
    ["Nem vou comentar o hate e misoginia que tenho visto aqui no tt em relação à Anitta", 
     "E xenofobia também. Tugas no seu melhor", 
     "knowhate/twt-bertimbau/twt-bb-b16e5-avg767.keras"],
    ["A Festa tá no Climax, chama o zuca pra Dançar.", 
     "Já reparaste no contador da luz? Vai trabalhar malandro", 
     "knowhate/twt-bertimbau/twt-bb-b16e5-avg767.keras"]
]

model_list = [
    "knowhate/twt-bertimbau/twt-bb-b16e5-avg767.keras"
]

def predict(text, target, chosen_model):    
    model1 = load_model(chosen_model, custom_objects={"TFBertModel": TFBertModel})
    checkpoint = "neuralmind/bert-base-portuguese-cased"
    tokenizer = AutoTokenizer.from_pretrained(checkpoint, use_fast=True)
    tokpair = tokenizer(text, target, truncation=True, padding='max_length', return_tensors='np')
    
    outp = model1(tokpair)
    
    proto_tensor = tf.make_tensor_proto(outp)
    allscores = tf.make_ndarray(proto_tensor)[0]
        
    scores_dict = {
        'Neutral': allscores[0],
        'Counter Speech': allscores[1],
        'Hate Speech': allscores[2]
    }
    
    return scores_dict

inputs = [
    gr.Textbox(label="Context", value= app_examples[0][0]),
    gr.Textbox(label="Target", value= app_examples[0][1]),
    gr.Dropdown(label="Model", choices=model_list, value=model_list[0])
]

outputs = [
    gr.Label(label="Result"),
]

gr.Interface(fn=predict, inputs=inputs, outputs=outputs, title=app_title, 
             description=app_description, examples=app_examples, theme=gr.themes.Base(primary_hue="red")).launch()