app.py

import gradio as gr
import torch
from transformers import RobertaForSequenceClassification, RobertaTokenizer

# Load the saved model and tokenizer
model_path = "Charankarnati18/TASK3"  # Your HuggingFace model repo
tokenizer = RobertaTokenizer.from_pretrained(model_path)
model = RobertaForSequenceClassification.from_pretrained(model_path)

# Set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Label mapping
label_map = {0: "Non-toxic", 1: "Slightly Toxic", 2: "Highly Toxic"}

def predict_toxicity(text):
    # Tokenize input
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=64)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    # Make prediction
    model.eval()
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probabilities = torch.softmax(logits, dim=1)
        prediction = torch.argmax(logits, dim=1).item()
    
    # Convert probabilities to percentages
    probs = probabilities[0].cpu().numpy() * 100
    
    # Create results dictionary
    results = {
        "Prediction": label_map[prediction],
        "Non-toxic": f"{probs[0]:.2f}%",
        "Neutal": f"{probs[1]:.2f}%",
        "Toxic": f"{probs[2]:.2f}%"
    }
    
    return results

# Create Gradio interface
demo = gr.Interface(
    fn=predict_toxicity,
    inputs=gr.Textbox(placeholder="Enter text to analyze for toxicity...", lines=5),
    outputs=gr.JSON(),
    title="Text Toxicity Analyzer",
    description="This app analyzes text and classifies it as non-toxic, slightly toxic, or highly toxic.",
    examples=[
        ["This is a wonderful day!"],
        ["I don't really like this product."],
        ["You are an absolute idiot and I hate you."]
    ],
    theme=gr.themes.Base()
)

# Launch the app
demo.launch()