app.py

import gradio as gr
from gradio.components import Text
import joblib
import clean
import nltk
nltk.download('wordnet')
import numpy as np
import language_detection
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

print("all imports worked")
# Load pre-trained model
model = joblib.load('model_joblib.pkl')
print("model load ")
tf = joblib.load('tf_joblib.pkl')
print("tfidf load ")

# Load Hindi abuse detection model
hindi_tokenizer = AutoTokenizer.from_pretrained("Hate-speech-CNERG/hindi-abusive-MuRIL")
hindi_model = AutoModelForSequenceClassification.from_pretrained("Hate-speech-CNERG/hindi-abusive-MuRIL")
print("Hindi model loaded")

def predict_hindi_text(text):
    inputs = hindi_tokenizer(text, return_tensors="pt", padding=True, truncation=True)
    outputs = hindi_model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    scores = predictions[0].detach().numpy()
    return scores

# Define function to predict whether sentence is abusive or not
def predict_abusive_lang(text):
    print("original text ", text)
    
    lang = language_detection.en_hi_detection(text)
    print("language detected ", lang)
    
    if lang=='eng':
        cleaned_text = clean.text_cleaning(text)
        print("cleaned text ", text)
        text = tf.transform([cleaned_text])
        print("tfidf transformation ", text)
        prediction = model.predict(text)
        print("prediction ", prediction)
        if len(prediction)!=0 and prediction[0]==0:
            return ["Not Abusive", cleaned_text]
        elif len(prediction)!=0 and prediction[0]==1:
            return ["Abusive",cleaned_text]
        else :
            return ["Please write something in the comment box..","No cleaned text"]
    elif lang=='hi':
        print("using transformers for Hindi text")
        scores = predict_hindi_text(text)
        if scores[1] > scores[0]:  # If score for abusive class is higher
            return ["Abusive", text]
        else:
            return ["Not Abusive", text]
    else:
        return ["Uknown Language","No cleaned text"]
    
# Define the GRADIO output interfaces
output_interfaces = [
    gr.Textbox(label="Result"),
    gr.Textbox(label="Cleaned text")
]
app = gr.Interface(predict_abusive_lang, inputs='text', outputs=output_interfaces, title="Abuse Classifier", description="Enter a sentence and the model will predict whether it is abusive or not.")
#Start the GRADIO app
app.launch()