app.py

from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig
import gradio as gr
from torch.nn import functional as F
import seaborn

import matplotlib
import platform

from transformers.file_utils import ModelOutput

if platform.system() == "Darwin":
    print("MacOS")
    matplotlib.use('Agg')
import matplotlib.pyplot as plt
import io
from PIL import Image

import matplotlib.font_manager as fm
import util


# global var
MODEL_NAME = 'jason9693/SoongsilBERT-beep-base'
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)
config = AutoConfig.from_pretrained(MODEL_NAME)

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


font_dir = ['./']
for font in fm.findSystemFonts(font_dir):
    print(font)
    fm.fontManager.addfont(font)
plt.rcParams["font.family"] = 'NanumGothicCoding'


def visualize_attention(sent, attention_matrix, n_words=10):
    def draw(data, x, y, ax):

        seaborn.heatmap(data, 
                        xticklabels=x, square=True, yticklabels=y, vmin=0.0, vmax=1.0, 
                        cbar=False, ax=ax)
        
    # make plt figure with 1x6 subplots
    fig = plt.figure(figsize=(16, 8))
    # fig.subplots_adjust(hspace=0.7, wspace=0.2)
    for i, layer in enumerate(range(1, 12, 2)):
        ax = fig.add_subplot(2, 3, i+1)
        ax.set_title("Layer {}".format(layer))
        draw(attention_matrix[layer], sent if layer > 6 else [], sent if layer in [1,7] else [], ax=ax)
 
    fig.tight_layout()
    plt.close()
    return fig


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

    tokenized_text = tokenizer([text], return_tensors='pt')

    input_tokens = tokenizer.convert_ids_to_tokens(tokenized_text.input_ids[0])
    try:
        input_tokens = util.bytetokens_to_unicdode(input_tokens) if config.model_type in ['roberta', 'gpt', 'gpt2'] else input_tokens
    except KeyError:
        input_tokens = input_tokens

    model.eval()
    output, attention = model(**tokenized_text, output_attentions=True, return_dict=False)
    output = F.softmax(output, dim=-1)
    result = {}
    
    for idx, label in enumerate(output[0].detach().numpy()):
        result[config.id2label[idx]] = float(label)

    fig = visualize_attention(input_tokens, attention[0][0].detach().numpy())
    return result, fig#.logits.detach()#.numpy()#, output.attentions.detach().numpy()


if __name__ == '__main__':
    text = '읿딴걸 홍볿글 읿랉곭 쌑젩낄고 앉앟있냩'

    model_name_list = [
        'jason9693/SoongsilBERT-beep-base',
        "beomi/beep-klue-roberta-base-hate",
        "beomi/beep-koelectra-base-v3-discriminator-hate",
        "beomi/beep-KcELECTRA-base-hate"
    ]

    #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=['label', 'plot'], 
        examples = [[MODEL_BUF["name"], text], [MODEL_BUF["name"], "4=🦀 4≠🦀"]],
        title="한국어 혐오성 발화 분류기 (Korean Hate Speech Classifier)",
        description="Korean Hate Speech Classifier with Several Pretrained LM\nCurrent Supported Model:\n1. SoongsilBERT\n2. KcBERT(+KLUE)\n3. KcELECTRA\n4.KoELECTRA."
        )
    app.launch(inline=False)