app.py

import gradio as gr
from transformers import AutoProcessor, Pix2StructForConditionalGeneration, T5Tokenizer, T5ForConditionalGeneration, Pix2StructProcessor, BartConfig,ViTConfig,VisionEncoderDecoderConfig, DonutProcessor, VisionEncoderDecoderModel, AutoTokenizer, AutoModel
from PIL import Image  
import torch
import warnings
import re
import json
import os
import numpy as np
import pandas as pd
from tqdm import tqdm
import argparse
from scipy import optimize
from typing import Optional
import dataclasses
import editdistance
import itertools
import sys
import time
import logging
import subprocess
import spaces
import openai
import base64
from io import StringIO
from huggingface_hub import hf_hub_download

#Git LFS pull 명령어 실행
#result = subprocess.run(['git', 'lfs', 'pull'], capture_output=True, text=True)

#명령어 실행 결과 출력 (선택 사항)
#if result.returncode == 0:
#    print("LFS 파일이 성공적으로 다운로드되었습니다.")
#else:
#    print(f"오류 발생: {result.stderr}")

#logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
#logger = logging.getLogger()

#warnings.filterwarnings('ignore')
#MAX_PATCHES = 512
# Load the models and processor
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Paths to the models
#ko_deplot_model_path = './deplot_model_ver_24.11.21_korean_only(exclude NUUA)_epoch1.bin'

file_path = hf_hub_download(
    repo_id="Sangjun2/skku-Deplot", 
    filename="deplot_model_ver_24.11.21_korean_only(exclude NUUA)_epoch3.bin"
)

# Load first model ko-deplot
def load_model1():
    processor1 = Pix2StructProcessor.from_pretrained('nuua/ko-deplot')
    model1 = Pix2StructForConditionalGeneration.from_pretrained('nuua/ko-deplot')
    model1.load_state_dict(torch.load(file_path, map_location="cpu"))
    model1.to(torch.device("cuda"))
    return processor1, model1

processor1, model1 = load_model1()

# Function to format output
def format_output(prediction):
    return prediction.replace('<0x0A>', '\n')

# First model prediction: ko-deplot
@spaces.GPU(enable_queue=True,duration=100)
def predict_model1(image):
    images = [image]
    inputs = processor1(images=images, text="Generate underlying data table of the figure below:", return_tensors="pt", padding=True)
    inputs = {k: v.to(device) for k, v in inputs.items()}  # Move to GPU

    model1.eval()
    with torch.no_grad():
        predictions = model1.generate(**inputs, max_new_tokens=4096)
    outputs = [processor1.decode(pred, skip_special_tokens=True) for pred in predictions]

    formatted_output = format_output(outputs[0])
    return formatted_output

# Set your OpenAI API key
openai.api_key = os.getenv('gpt_api_key')

# Function to encode the image as base64
def encode_image(image_path):
    with open(image_path, "rb") as image_file:
        return base64.b64encode(image_file.read()).decode("utf-8")

# Second model prediction: gpt-4o-mini
def predict_model2(image):
    # Encode the uploaded image to base64
    image_data = encode_image(image)

    # Prepare the request content
    response = openai.ChatCompletion.create(
        model="gpt-4o-mini",
        messages=[
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": "please extract chart title and chart data manually and present them as a table. you should only provide title and table without adding any additional comments such as **Chart Title:** ."
                    },
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": f"data:image/jpeg;base64,{image_data}"
                        }
                    }
                ]
            }
        ]
    )

    # Return the table data from the response
    return response.choices[0]["message"]["content"]

def ko_deplot_convert_to_dataframe(label_table_str): #function that converts text generated by ko-deplot to pandas dataframe
    lines = label_table_str.strip().split("\n")
    data=[]
    title= lines[0].split(" | ")[1]

    if(len(lines[1].split("|")) == len(lines[2].split("|"))):
      headers=lines[1].split("|")
      for line in lines[2:]:
          data.append(line.split("|"))
      df = pd.DataFrame(data, columns=headers)
      return df, title
    else:
      legend_row=lines[1].split("|")
      legend_row.insert(0," ")
      for line in lines[2:]:
          data.append(line.split("|"))
      df = pd.DataFrame(data, columns=legend_row)
      return df, title

def gpt_convert_to_dataframe(table_text): #function that converts text generated by gpt to pandas dataframe
    try:
        # Split the text into lines
        lines = table_text.strip().split("\n")
        title=lines[0]
        lines.pop(1)
        lines.pop(2)
        # Process the remaining lines to create the DataFrame
        data = [line.split("|")[1:-1] for line in lines[1:]]  # Split by | and remove empty first/last items
        dataframe = pd.DataFrame(data[1:], columns=[col.strip() for col in data[0]])  # Use the first row as headers
        
        return dataframe, title
    except Exception as e:
        return f"Error converting table to DataFrame: {e}"

def real_time_check(image_file):
    image = Image.open(image_file)
    ko_deplot_generated_txt = predict_model1(image)
    parts=ko_deplot_generated_txt.split("\n")
    del parts[-1]
    ko_deplot_generated_txt="\n".join(parts)
    gpt_generated_txt=predict_model2(image_file)
    try:
        ko_deplot_generated_df, ko_deplot_generated_title=ko_deplot_convert_to_dataframe(ko_deplot_generated_txt)
        gpt_generated_df, gpt_generated_title=gpt_convert_to_dataframe(gpt_generated_txt)
        return ko_deplot_generated_df, ko_deplot_generated_title, gpt_generated_df, gpt_generated_title, None,None,0
    except Exception as e:
        return None,None,None,None,ko_deplot_generated_txt,gpt_generated_txt,1
    
flag = 0 #flag to check whether exception happens or not. if flag is 1, it means that exception(generated txt cannot be converted to pandas dataframe) happens. 
def inference(image_uploader,mode_selector):
    if(mode_selector=="파일 업로드"):
        ko_deplot_generated_df, ko_deplot_generated_title, gpt_generated_df, gpt_generated_title, ko_deplot_generated_txt, gpt_generated_txt, flag= real_time_check(image_uploader)
        if flag==1:
            return gr.update(visible=False), gr.update(visible=False), gr.Text(ko_deplot_generated_txt,visible=True),gr.Text(gpt_generated_txt,visible=True)
        else:
            return gr.DataFrame(ko_deplot_generated_df, label = ko_deplot_generated_title, visible=True), gr.DataFrame(gpt_generated_df, label = gpt_generated_title, visible=True), gr.update(visible=False),gr.update(visible=False)
    else:
        ko_deplot_generated_df, ko_deplot_generated_title, gpt_generated_df, gpt_generated_title, ko_deplot_generated_txt, gpt_generated_txt, flag= real_time_check(image_files[current_image_index])
        if flag==1:
            return gr.update(visible=False), gr.update(visible=False), gr.Text(ko_deplot_generated_txt,visible=True),gr.Text(gpt_generated_txt,visible=True)
        else:
            return gr.DataFrame(ko_deplot_generated_df, label = ko_deplot_generated_title, visible=True), gr.DataFrame(gpt_generated_df, label = gpt_generated_title, visible=True), gr.update(visible=False),gr.update(visible=False)
        
def toggle_model(selected_models,flag):
    # Create a visibility list initialized to False for all components
    visibility = [False] * 6
    # Update visibility based on the selected models
    if "VAIV_DePlot" in selected_models:
        visibility[4]= True
        if flag:
            visibility[2]=  True
        else:
            visibility[0]= True
    if "gpt-4o-mini" in selected_models:
        visibility[5]= True
        if flag:
            visibility[3]=  True
        else:
            visibility[1]= True
    if "all" in selected_models:
        visibility[4]=True
        visibility[5]=True
        if flag:
            visibility[2]=  True
            visibility[3]= True
        else:
            visibility[0]= True
            visibility[1]= True
    # Return gr.update for each component with the corresponding visibility status
    return tuple(gr.update(visible=v) for v in visibility)

def toggle_mode(mode):
        if mode == "파일 업로드":
            return gr.update(visible=True), gr.update(visible=False), gr.update(visible=False), gr.update(visible=False) , gr.update(visible=False)
        else:
            return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)

def display_image(image_file):
    image=Image.open(image_file)
    return image, os.path.basename(image_file)

# Function to display the images in the folder sequentially
image_files = []
current_image_index = 0
image_files_cnt=0

def display_folder_images(image_file_path_list):
    global image_files, current_image_index,image_files_cnt
    image_files = image_file_path_list
    image_files_cnt=len(image_files)
    current_image_index = 0
    if image_files:
        return Image.open(image_files[current_image_index]), os.path.basename(image_files[current_image_index]), gr.update(interactive=False), gr.update(interactive=True), gr.update(visible = True), gr.update(visible = False)
    return None, "No images found"
    

def next_image():
    global current_image_index
    if image_files:
        current_image_index = (current_image_index + 1)
        prev_disabled = current_image_index == 0
        next_disabled = current_image_index == (len(image_files) - 1)
        return Image.open(image_files[current_image_index]), os.path.basename(image_files[current_image_index]), gr.update(interactive=not prev_disabled), gr.update(interactive= not next_disabled)
    return None, "No images found"

def prev_image():
    global current_image_index
    if image_files:
        current_image_index = (current_image_index - 1)
        prev_disabled = current_image_index == 0
        next_disabled = current_image_index == (len(image_files) - 1)
        return Image.open(image_files[current_image_index]), os.path.basename(image_files[current_image_index]), gr.update(interactive=not prev_disabled), gr.update(interactive= not next_disabled)
    return None, "No images found"

def folder_reupload():
    return gr.update(visible=False), gr.update(visible=True)

css = """
.dataframe-class {
    overflow-y: auto !important; /* 스크롤을 가능하게 */
    height: 250px
}
"""

with gr.Blocks(css=css) as iface:
    with gr.Row():
        gr.Markdown("<h1 style='text-align: center;'>SKKU-VAIV Automatic chart understanding evaluation tool</h1>")
    gr.Markdown("<hr style='border: 1px solid #ddd;' />")
    with gr.Row():
        with gr.Column():
            mode_selector = gr.Radio(["파일 업로드", "폴더 업로드"], label="Upload Mode", value="파일 업로드")
            image_uploader = gr.File(file_count="single", file_types=["image"], visible=True)
            folder_uploader = gr.File(file_count="directory", file_types=["image"], visible=False, height=50)
            folder_reupload_button = gr.Button("폴더 업로드", visible=False)
            model_type=gr.Dropdown(["VAIV_DePlot","gpt-4o-mini","all"],value="all",label="model",multiselect=True)
            image_displayer = gr.Image(visible=True)
            image_name = gr.Text("", visible=True)
            with gr.Row():
                prev_button = gr.Button("이전", visible=False, interactive=False)
                next_button = gr.Button("다음", visible=False, interactive=False)
            inference_button = gr.Button("추론")
        with gr.Column():
            md1 = gr.Markdown("# VAIV_DePlot Inference Result")
            ko_deplot_generated_df = gr.DataFrame(visible=True, elem_classes="dataframe-class")
            ko_deplot_generated_txt = gr.Text(visible=False)
        with gr.Column():    
            md2 = gr.Markdown("# gpt-4o-mini Inference Result", visible=True)
            gpt_generated_df = gr.DataFrame(visible=True, elem_classes="dataframe-class")
            gpt_generated_txt = gr.Text(visible=False)
            #label_df = gr.DataFrame(visible=False, label="Ground Truth Table", elem_classes="dataframe-class",scale=1)

    model_type.change(
                        toggle_model,
                        inputs=[model_type, gr.State(flag)],
                        outputs=[ko_deplot_generated_df,gpt_generated_df,ko_deplot_generated_txt,gpt_generated_txt,md1,md2]
                        )

    mode_selector.change(
        toggle_mode,
        inputs=[mode_selector],
        outputs=[image_uploader, folder_uploader, prev_button, next_button, folder_reupload_button]
    )

    image_uploader.upload(display_image,inputs=[image_uploader],outputs=[image_displayer,image_name])
    folder_uploader.upload(display_folder_images, inputs=[folder_uploader], outputs=[image_displayer, image_name, prev_button, next_button, folder_reupload_button, folder_uploader])
    prev_button.click(prev_image, outputs=[image_displayer, image_name, prev_button, next_button])
    next_button.click(next_image, outputs=[image_displayer, image_name, prev_button, next_button])
    folder_reupload_button.click(folder_reupload, outputs =[folder_reupload_button, folder_uploader])
    inference_button.click(inference,inputs=[image_uploader,mode_selector],outputs=[ko_deplot_generated_df, gpt_generated_df, ko_deplot_generated_txt, gpt_generated_txt])

    if __name__ == "__main__":
        iface.launch(share=True)