app.py

"""
python interactive.py
"""
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import TextClassificationPipeline
import gradio as gr

# global var
MODEL_NAME = 'momo/KcELECTRA-base_Hate_speech_Privacy_Detection'
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSequenceClassification.from_pretrained(
    MODEL_NAME,
    num_labels= 15,
    problem_type="multi_label_classification"
)

MODEL_BUF = {
    "name": MODEL_NAME,
    "tokenizer": tokenizer,
    "model": model,
}

def change_model_name(name):
    MODEL_BUF["name"] = name
    MODEL_BUF["tokenizer"] = AutoTokenizer.from_pretrained(name)
    MODEL_BUF["model"] = AutoModelForSequenceClassification.from_pretrained(name)

def predict(model_name, text):
    if model_name != MODEL_BUF["name"]:
        change_model_name(model_name)
    
    tokenizer = MODEL_BUF["tokenizer"]
    model = MODEL_BUF["model"]

    unsmile_labels = ["여성/가족","남성","성소수자","인종/국적","연령","지역","종교","기타 혐오","악플/욕설", "clean", '이름', '전화번호', '주소', '계좌번호', '주민번호']
    num_labels = len(unsmile_labels)

    model.config.id2label = {i: label for i, label in zip(range(num_labels), unsmile_labels)}
    model.config.label2id = {label: i for i, label in zip(range(num_labels), unsmile_labels)}

    pipe = TextClassificationPipeline(
    model = model,
    tokenizer = tokenizer,
    return_all_scores=True,
    function_to_apply='sigmoid'
    )
    for result in pipe(text):
        print(result)
    return result

if __name__ == '__main__':
    exam1 = '경기도 성남시 수정구 태평3동은 우리 동네야!'
    exam2 = '내 핸드폰 번호는 010-3930-8237 이야!'
    exam3 = '아 젠장 너무 짜증난다'

    model_name_list = [
        'momo/KcELECTRA-base_Hate_speech_Privacy_Detection',
        "momo/KcBERT-base_Hate_speech_Privacy_Detection",
    ]

    #Create a gradio app with a button that calls predict()
    app = gr.Interface(
        fn=predict,
        inputs=[gr.inputs.Dropdown(model_name_list, label="Model Name"), 'text'],
        outputs='text', 
        examples = [
            [MODEL_BUF["name"], exam1], 
            [MODEL_BUF["name"], exam2],
            [MODEL_BUF["name"], exam3]
            ],
        title="한국어 혐오표현, 개인정보 판별기 (Korean Hate Speech and Privacy Detection)",
        description="Korean Hate Speech and Privacy Detection.\n 15개 label Detection: 여성/가족, 남성, 성소수자, 인종/국적, 연령, 지역, 종교, 기타 혐오, 악플/욕설, clean, '이름', '전화번호', '주소', '계좌번호', '주민번호' \n Current Supported Model:\n1. momo/KcELECTRA-base_Hate_speech_Privacy_Detection\n2. momo/KcBERT-base_Hate_speech_Privacy_Detection."
        )
    app.launch()