app/app.py

import os
from typing import List

import streamlit as st
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

os.environ["TOKENIZERS_PARALLELISM"] = "false"


@st.cache(allow_output_mutation=True)
def load_model(model_name):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
    return model, tokenizer


def paraphrase(model, encoding, top_k=120, top_p=0.95, max_len=120) -> List[str]:
    outputs = model.generate(
        input_ids=encoding["input_ids"],
        attention_mask=encoding["attention_mask"],
        do_sample=True,
        top_k=top_k,
        top_p=top_p,
        max_length=max_len,
        early_stopping=True,
        num_return_sequences=5,
    )
    return [
        tokenizer.decode(
            output, skip_special_tokens=True, clean_up_tokenization_spaces=True
        )
        for output in outputs
    ]


if __name__ == "__main__":
    st.header("Indonesian Paraphrase Generation")
    user_input = st.text_area("Original Sentence", "", height=30)

    # Slider for max_len
    st.sidebar.header("Decoding Settings")
    max_len = st.sidebar.slider("Max-Length", 0, 512, 256)
    top_k = st.sidebar.slider("Top-K", 0, 512, 200)
    top_p = st.sidebar.slider("Top-P", 0.0, 1.0, 0.95)

    if st.button("Paraphrase") or user_input:
        with st.spinner("T5 is processing your text..."):
            model, tokenizer = load_model("Wikidepia/IndoT5-base-paraphrase")
            text = "paraphrase: " + user_input + " </s>"
            encode_id = tokenizer(text, return_tensors="pt")
            outputs = paraphrase(
                model, encode_id, top_k=top_k, top_p=top_p, max_len=max_len
            )
            st.markdown("### Hasil Parafrase")
            for i, output in enumerate(outputs):
                st.markdown(f"- {output}")