app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from sentence_splitter import SentenceSplitter
import itertools
import spaces
device = "cuda"

tokenizer = AutoTokenizer.from_pretrained("NoaiGPT/777")
model = AutoModelForSeq2SeqLM.from_pretrained("NoaiGPT/777").to(device)

splitter = SentenceSplitter(language='en')

@spaces.GPU
def process_and_generate(text):
    def generate_title(sentence):
        input_ids = tokenizer(f'paraphraser: {sentence}', return_tensors="pt", padding="longest", truncation=True, max_length=64).input_ids.to(device)
        outputs = model.generate(
            input_ids,
            num_beams=8,
            num_beam_groups=4,
            num_return_sequences=6,
            repetition_penalty=12.0,
            diversity_penalty=4.0,
            no_repeat_ngram_size=3,
            temperature=1.1,
            top_k=50,
            top_p=0.95,
            max_length=64
        )
        return tokenizer.batch_decode(outputs, skip_special_tokens=True)

    paragraphs = text.split('\n\n')
    detailed_results = []
    all_paraphrases = []

    for paragraph in paragraphs:
        sentences = splitter.split(paragraph)
        paragraph_results = []
        paragraph_paraphrases = []

        for sentence in sentences:
            titles = generate_title(sentence)
            paragraph_results.append(f"Original: {sentence}\nParaphrases:\n" + "\n".join(titles))
            paragraph_paraphrases.append(titles)

        detailed_results.append("\n\n".join(paragraph_results))
        all_paraphrases.append(paragraph_paraphrases)

    detailed_output = "\n\n---\n\n".join(detailed_results)

    # Generate all possible combinations of paraphrases for each paragraph
    all_paragraph_combinations = []
    for paragraph_paraphrases in all_paraphrases:
        paragraph_combinations = list(itertools.product(*paragraph_paraphrases))
        all_paragraph_combinations.append([" ".join(combo) for combo in paragraph_combinations])

    # Join the paragraphs to create full text variations
    all_text_variations = ["\n\n".join(variation) for variation in itertools.product(*all_paragraph_combinations)]

    # Limit the number of variations to prevent overwhelming output
    max_variations = 100
    all_text_variations = all_text_variations[:max_variations]

    final_output = "All Paraphrase Combinations:\n\n" + "\n\n---\n\n".join(all_text_variations)

    return detailed_output, final_output

iface = gr.Interface(
    fn=process_and_generate,
    inputs=gr.Textbox(lines=10, label="Input Text"),
    outputs=[
        gr.Textbox(lines=20, label="Detailed Paraphrases"),
        gr.Textbox(lines=20, label="All Paraphrase Combinations")
    ],
    title="Diverse Paraphrase Generator",
    description="Generate multiple diverse paraphrases for each sentence in the input text using NoaiGPT/777 model."
)

iface.launch()