app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

class TextDetectionApp:
    def __init__(self):
        # Load DeBERTa model and tokenizer
        self.deberta_tokenizer = AutoTokenizer.from_pretrained("zeyadusf/deberta-DAIGT-MODELS")
        self.deberta_model = AutoModelForSequenceClassification.from_pretrained("zeyadusf/deberta-DAIGT-MODELS")

        # Load RoBERTa model and tokenizer
        self.roberta_tokenizer = AutoTokenizer.from_pretrained("zeyadusf/roberta-DAIGT-kaggle")
        self.roberta_model = AutoModelForSequenceClassification.from_pretrained("zeyadusf/roberta-DAIGT-kaggle")

        # Load Feedforward model
        self.ff_model = torch.jit.load("model_scripted.pt")

    def api_huggingface(self, text):
        """
        Generate predictions using the DeBERTa and RoBERTa models.

        Args:
            text (str): The input text to classify.

        Returns:
            tuple: Predictions from RoBERTa and DeBERTa models.
        """
        # DeBERTa predictions
        deberta_inputs = self.deberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
        deberta_outputs = self.deberta_model(**deberta_inputs)
        deberta_logits = deberta_outputs.logits
        deberta_scores = torch.softmax(deberta_logits, dim=1)
        deberta_predictions = [
            {"label": f"LABEL_{i}", "score": score.item()}
            for i, score in enumerate(deberta_scores[0])
        ]

        # RoBERTa predictions
        roberta_inputs = self.roberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
        roberta_outputs = self.roberta_model(**roberta_inputs)
        roberta_logits = roberta_outputs.logits
        roberta_scores = torch.softmax(roberta_logits, dim=1)
        roberta_predictions = [
            {"label": f"LABEL_{i}", "score": score.item()}
            for i, score in enumerate(roberta_scores[0])
        ]

        return roberta_predictions, deberta_predictions

    def generate_ff_input(self, models_results):
        """
        Generates input features for the Feedforward model from the API output.

        Parameters:
            models_results (tuple): Tuple containing the results of DeBERTa and RoBERTa models.

        Returns:
            torch.Tensor: Feedforward model input features tensor.
        """
        roberta, deberta = models_results
        input_ff = []
        try:
            if roberta[0]['label'] == 'LABEL_0':
                input_ff.append(roberta[0]['score'])
                input_ff.append(roberta[1]['score'])
            else:
                input_ff.append(roberta[1]['score'])
                input_ff.append(roberta[0]['score'])

            if deberta[0]['label'] == 'LABEL_0':
                input_ff.append(deberta[0]['score'])
                input_ff.append(deberta[1]['score'])
            else:
                input_ff.append(deberta[1]['score'])
                input_ff.append(deberta[0]['score'])

        except Exception as e:
            print(f"Error {e}: The text is long")

        input_ff = torch.tensor(input_ff, dtype=torch.float32)
        input_ff = input_ff.view(1, -1)
        return input_ff

    def detect_text(self, text):
        """
        Detects whether the input text is generated or human-written using the Feedforward model.

        Returns:
            float: The detection result.
        """
        with torch.no_grad():
            self.output = self.ff_model(self.generate_ff_input(self.api_huggingface(text)))[0][0].item()
        return self.output

    def classify_text(self, text, model_choice):
        """
        Classifies the input text using the selected model.

        Args:
            text (str): The input text to classify.
            model_choice (str): The model to use ('DeBERTa', 'RoBERTa', or 'Feedforward').

        Returns:
            str: The classification result.
        """
        if model_choice == 'DeBERTa':
            # Tokenize input
            inputs = self.deberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)

            # Run model
            outputs = self.deberta_model(**inputs)

            # Get classification results
            logits = outputs.logits
            predicted_class_id = logits.argmax().item()
            return f"DeBERTa Prediction: Class {predicted_class_id}"

        elif model_choice == 'RoBERTa':
            # Tokenize input
            inputs = self.roberta_tokenizer(text, return_tensors="pt", truncation=True, padding=True)

            # Run model
            outputs = self.roberta_model(**inputs)

            # Get classification results
            logits = outputs.logits
            predicted_class_id = logits.argmax().item()
            return f"RoBERTa Prediction: Class {predicted_class_id}"

        elif model_choice == 'Feedforward':
            # Run feedforward detection
            detection_score = self.detect_text(text)
            return f"Feedforward Detection Score: {detection_score}"

        else:
            return "Invalid model selection."


# Initialize the app
app = TextDetectionApp()

# Gradio Interface
iface = gr.Interface(
    fn=app.classify_text,
    inputs=[
        gr.Textbox(lines=2, placeholder="Enter your text here..."),
        gr.Radio(choices=["DeBERTa", "RoBERTa", "Feedforward"], label="Model Choice")
    ],
    outputs="text",
    title="Text Classification with Multiple Models",
    description="Classify text using DeBERTa, RoBERTa, or a custom Feedforward model."
)

iface.launch()