app.py

import os
import json
import gradio as gr
import spaces
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification, pipeline
from sentence_splitter import SentenceSplitter
from itertools import product

# Get the Hugging Face token from environment variable
hf_token = os.getenv('HF_TOKEN')

cuda_available = torch.cuda.is_available()
device = torch.device("cuda" if cuda_available else "cpu")
print(f"Using device: {device}")

# Initialize paraphraser model and tokenizer
paraphraser_model_name = "NoaiGPT/777"
paraphraser_tokenizer = AutoTokenizer.from_pretrained(paraphraser_model_name, use_auth_token=hf_token)
paraphraser_model = AutoModelForSeq2SeqLM.from_pretrained(paraphraser_model_name, use_auth_token=hf_token).to(device)

# Initialize classifier model and tokenizer
classifier_model_name = "andreas122001/roberta-mixed-detector"
classifier_tokenizer = AutoTokenizer.from_pretrained(classifier_model_name)
classifier_model = AutoModelForSequenceClassification.from_pretrained(classifier_model_name).to(device)

# Initialize spelling correction pipeline
spelling_correction = pipeline("text2text-generation", model="oliverguhr/spelling-correction-english-base", device=0 if cuda_available else -1)

# Initialize sentence splitter
splitter = SentenceSplitter(language='en')

def classify_text(text):
    inputs = classifier_tokenizer(text, return_tensors="pt", truncation=True, max_length=512).to(device)
    with torch.no_grad():
        outputs = classifier_model(**inputs)
    probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(probabilities, dim=-1).item()
    main_label = classifier_model.config.id2label[predicted_class]
    main_score = probabilities[0][predicted_class].item()
    return main_label, main_score

def correct_spelling(text):
    corrected_text = spelling_correction(text, max_length=2048)[0]['generated_text']
    print(corrected_text)
    return corrected_text

@spaces.GPU
def generate_paraphrases(text, setting, output_format):
    sentences = splitter.split(text)
    all_sentence_paraphrases = []

    if setting == 1:
        num_return_sequences = 5
        repetition_penalty = 1.1
        no_repeat_ngram_size = 2
        temperature = 0.8
        max_length = 128
    elif setting == 2:
        num_return_sequences = 10
        repetition_penalty = 1.2
        no_repeat_ngram_size = 3
        temperature = 1.2
        max_length = 192
    elif setting == 3:
        num_return_sequences = 15
        repetition_penalty = 1.3
        no_repeat_ngram_size = 4
        temperature = 1.4
        max_length = 256
    elif setting == 4:
        num_return_sequences = 20
        repetition_penalty = 1.4
        no_repeat_ngram_size = 5
        temperature = 1.6
        max_length = 320
    else:
        num_return_sequences = 25
        repetition_penalty = 1.5
        no_repeat_ngram_size = 6
        temperature = 1.8
        max_length = 384
    
    top_k = 40
    top_p = 0.90
    length_penalty = 1.0
    
    formatted_output = "Original text:\n" + text + "\n\n"
    formatted_output += "Paraphrased versions:\n"
    
    json_output = {
        "original_text": text,
        "paraphrased_versions": [],
        "combined_versions": [],
        "human_like_versions": []
    }
    
    for i, sentence in enumerate(sentences):
        inputs = paraphraser_tokenizer(f'{sentence}', return_tensors="pt", padding="longest", truncation=True, max_length=max_length).to(device)
        
        # Generate paraphrases using the specified parameters
        outputs = paraphraser_model.generate(
            inputs.input_ids,
            attention_mask=inputs.attention_mask,
            num_return_sequences=num_return_sequences,
            repetition_penalty=repetition_penalty,
            no_repeat_ngram_size=no_repeat_ngram_size,
            temperature=temperature,
            max_length=max_length,
            top_k=top_k,
            top_p=top_p,
            do_sample=True,
            early_stopping=False,
            length_penalty=length_penalty
        )
        
        paraphrases = paraphraser_tokenizer.batch_decode(outputs, skip_special_tokens=True)
        
        corrected_paraphrases = [correct_spelling(paraphrase) for paraphrase in paraphrases]
        
        formatted_output += f"Original sentence {i+1}: {sentence}\n"
        for j, paraphrase in enumerate(corrected_paraphrases, 1):
            formatted_output += f"  Paraphrase {j}: {paraphrase}\n"
        
        json_output["paraphrased_versions"].append({
            f"original_sentence_{i+1}": sentence,
            "paraphrases": corrected_paraphrases
        })
        
        all_sentence_paraphrases.append(corrected_paraphrases)
        formatted_output += "\n"
    
    all_combinations = list(product(*all_sentence_paraphrases))
    
    formatted_output += "\nCombined paraphrased versions:\n"
    combined_versions = []
    for i, combination in enumerate(all_combinations[:50], 1):  # Limit to 50 combinations
        combined_paraphrase = " ".join(combination)
        combined_versions.append(combined_paraphrase)
    
    json_output["combined_versions"] = combined_versions
    
    # Classify combined versions
    human_versions = []
    for i, version in enumerate(combined_versions, 1):
        label, score = classify_text(version)
        formatted_output += f"Version {i}:\n{version}\n"
        formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
        if label == "human-produced" or (label == "machine-generated" and score < 0.98):
            human_versions.append((version, label, score))
    
    formatted_output += "\nHuman-like or Less Confident Machine-generated versions:\n"
    for i, (version, label, score) in enumerate(human_versions, 1):
        formatted_output += f"Version {i}:\n{version}\n"
        formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
    
    json_output["human_like_versions"] = [
        {"version": version, "label": label, "confidence_score": score}
        for version, label, score in human_versions
    ]
    
    # If no human-like versions, include the top 5 least confident machine-generated versions
    if not human_versions:
        human_versions = sorted([(v, l, s) for v, l, s in zip(combined_versions, [classify_text(v)[0] for v in combined_versions], [classify_text(v)[1] for v in combined_versions])], key=lambda x: x[2])[:5]
        formatted_output += "\nNo human-like versions found. Showing top 5 least confident machine-generated versions:\n"
        for i, (version, label, score) in enumerate(human_versions, 1):
            formatted_output += f"Version {i}:\n{version}\n"
            formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
    
    if output_format == "text":
        return formatted_output, "\n\n".join([v[0] for v in human_versions])
    else:
        return json.dumps(json_output, indent=2), "\n\n".join([v[0] for v in human_versions])

# Define the Gradio interface
iface = gr.Interface(
    fn=generate_paraphrases,
    inputs=[
        gr.Textbox(lines=5, label="Input Text"),
        gr.Slider(minimum=1, maximum=5, step=1, label="Readability to Human-like Setting"),
        gr.Radio(["text", "json"], label="Output Format")
    ],
    outputs=[
        gr.Textbox(lines=20, label="Detailed Paraphrases and Classifications"),
        gr.Textbox(lines=10, label="Human-like or Less Confident Machine-generated Paraphrases")
    ],
    title="Advanced Diverse Paraphraser with Human-like Filter",
    description="Enter a text, select a setting from readable to human-like, and choose the output format to generate diverse paraphrased versions. Combined versions are classified, and those detected as human-produced or less confidently machine-generated are presented in the final output."
)

# Launch the interface
iface.launch()