app.py

import gdown
from git import Repo
import gradio as gr
from huggingface_hub import snapshot_download
import os
import penman
import sys
import time
import torch
from transformers import pipeline

if not os.path.exists("amr-tst-indo"):
    Repo.clone_from("https://github.com/AbdiHaryadi/amr-tst-indo.git", "amr-tst-indo")
sys.path.append("./amr-tst-indo")

from text_to_amr import TextToAMR
from style_detector import StyleDetector
from style_rewriting import StyleRewriting
from amr_to_text import AMRToTextWithTaufiqMethod

amr_parsing_model_name = "mbart-en-id-smaller-indo-amr-parsing-translated-nafkhan"
snapshot_download(
    repo_id=f"abdiharyadi/{amr_parsing_model_name}",
    local_dir=f"./amr-tst-indo/AMRBART-id/models/{amr_parsing_model_name}",
    ignore_patterns=[
        "*log*",
        "*checkpoint*",
    ]
)
t2a = TextToAMR(model_name=amr_parsing_model_name)

gdown.download(
    "https://drive.google.com/uc?id=1J_6PbYsQ6Kl4Qfs1wBVwd52_r9uTpIxx",
    "./model-best.pt"
)
sd = StyleDetector(
    config_path="./amr-tst-indo/indonesian-aste-generative/resources/exp-v2/exp-m0.yaml",
    model_path="./model-best.pt"
)

device_type = "cuda" if torch.cuda.is_available() else "cpu"
clf_pipeline = pipeline(
    "text-classification",
    model="abdiharyadi/roberta-base-indonesian-522M-with-sa-william-dataset",
    device=device_type
)
gdown.download(
    "https://drive.google.com/uc?id=15KctCcsHgTFMUh_tWNBNUiCyX56fq6p-",
    "./fasttext_skipgram_indo.bin"
)
sr = StyleRewriting(
    clf_pipeline=clf_pipeline,
    fasttext_model_path="./fasttext_skipgram_indo.bin",
    position_aware_concatenation=False,
    reset_sense_strategy=False,
    max_score_strategy=True,
    maximize_style_words_expansion=False
)

amr_gen_model_name = "taufiq-indo-amr-generation-gold-uncased"
model_path = f"./{amr_gen_model_name}"
snapshot_download(
    repo_id=f"abdiharyadi/{amr_gen_model_name}",
    local_dir=model_path,
    allow_patterns=[
        "*checkpoint-3*"
    ]
)
a2t = AMRToTextWithTaufiqMethod(
    model_path=os.path.join(model_path, "checkpoint-3"),
    lowercase=True,
)

def run(text, source_style):
    yield (
        "(Memproses ...)",
        "(Menunggu ...)",
        "(Menunggu ...)",
        "(Menunggu ...)",
        "(Menunggu ...)",
    )

    start_time = time.time()

    source_amr, *_ = t2a([text])
    source_amr.metadata = {}
    source_amr_display = penman.encode(source_amr)
    source_amr_display += f"\n\n({time.time() - start_time:.2f} s)"
    yield (
        source_amr_display,
        "(Memproses ...)",
        "(Menunggu ...)",
        "(Menunggu ...)",
        "(Menunggu ...)",
    )

    triplets = sd.get_triplets(text)
    triplets_display = "\n".join(f"({x[0]}, {x[1]}, {x[2]})" for x in triplets)
    triplets_display += f"\n\n({time.time() - start_time:.2f} s)"
    yield (
        source_amr_display,
        triplets_display,
        "(Memproses ...)",
        "(Menunggu ...)",
        "(Menunggu ...)",
    )

    style_words = sd.get_style_words_from_triplets(triplets)
    style_words_display = ", ".join(style_words)
    style_words_display += f"\n\n({time.time() - start_time:.2f} s)"
    yield (
        source_amr_display,
        triplets_display,
        style_words_display,
        "(Memproses ...)",
        "(Menunggu ...)",
    )

    target_amr = sr(text, source_amr, source_style, style_words)
    target_amr_display = penman.encode(target_amr)
    target_amr_display += f"\n\n({time.time() - start_time:.2f} s)"
    yield (
        source_amr_display,
        triplets_display,
        style_words_display,
        target_amr_display,
        "(Memproses ...)",
    )

    result, *_ = a2t([target_amr])
    result += f"\n\n({time.time() - start_time:.2f} s)"
    yield (
        source_amr_display,
        triplets_display,
        style_words_display,
        target_amr_display,
        result
    )

with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            gr.Interface(
                fn=run,
                inputs=[
                    gr.Textbox(label="Teks (Text)"),
                    gr.Radio(label="Gaya sumber (Source style)", choices=[
                        ("Positif (Positive)", "LABEL_1"),
                        ("Negatif (Negative)", "LABEL_0"),
                    ], value="LABEL_1"),
                ],
                outputs=[
                    gr.Textbox(label="Graf AMR sumber (Source AMR graph)"),
                    gr.Textbox(label="Triplet (Triplets)"),
                    gr.Textbox(label="Kata bergaya (Style words)"),
                    gr.Textbox(label="Graf AMR target (Target AMR graph)"),
                    gr.Textbox(label="Hasil (Result)"),
                ]
            )
        
        with gr.Column():
            gr.Markdown("""
            # Pengakuan
            Demo ini disiapkan untuk Program Penelitian dan Pengabdian Masyarakat STEI ITB 2024.

            **Tim Peneliti**:
            - Masayu Leylia Khodra ([email protected])
            - M. Abdi Haryadi. H ([email protected])
            """)

demo.launch()