app.py

import gradio as gr
from prompt_prefix import prompt_prefix
import re
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Modell und Tokenizer laden
# distilgpt2 is only 80MB -> no inference model, thus add prompt_prefix or train
tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
model = AutoModelForCausalLM.from_pretrained("distilgpt2")

# conversion method
def text_to_emoji(input_text):
    # Eingabetext bereinigen (optional)
    cleaned_text = re.sub(r"[.,!?;:]", "", input_text)

    # Pattern-based prompt with data from prompt_prefix
    prompt = "(\n" + "".join(f'"{line}\\n"\n' for line in prompt_prefix) + f"\"{input_text} →\"" + ")"
    
    # Tokenisierung und Generation
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs,
        max_new_tokens=10,
        do_sample=True,
        temperature=0.9,
        top_k=50,
        pad_token_id=tokenizer.eos_token_id  # Prevents warning
    )

    # Decodieren
    generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)

    # Nur den generierten Teil nach dem letzten "→"
    emoji_part = generated_text.split("→")[-1].strip().split("\n")[0]

    return emoji_part
    
# Gradio UI
iface = gr.Interface(
    fn=text_to_emoji,
    inputs=gr.Textbox(lines=2, placeholder="Enter a sentence..."),
    outputs="text",
    title="AI-Powered Emoji Translator",
    description="Enter a sentence, and the AI will transform it into an emoji-version 🥳"
)

iface.launch()