scripts/data_browser.py

"""
pip install gradio    # proxy_on first
python vis_geochat_data.py
# browse data in http://127.0.0.1:10064
"""

import os
import io
import json
import copy
import time
import gradio as gr
import base64
from PIL import Image
from io import BytesIO
from argparse import Namespace
# from llava import conversation as conversation_lib
from typing import Sequence
from vlmeval import *
from vlmeval.dataset import SUPPORTED_DATASETS, build_dataset

SYS = "You are a helpful assistant. Your job is to faithfully translate all provided text into Chinese faithfully. "

# Translator = SiliconFlowAPI(model='Qwen/Qwen2.5-7B-Instruct', system_prompt=SYS)
Translator = OpenAIWrapper(model='gpt-4o-mini', system_prompt=SYS)


def image_to_mdstring(image):
    return f"![image](data:image/jpeg;base64,{image})"


def images_to_md(images):
    return '\n\n'.join([image_to_mdstring(image) for image in images])


def mmqa_display(question, target_size=2048):
    question = {k.lower() if len(k) > 1 else k: v for k, v in question.items()}
    keys = list(question.keys())
    keys = [k for k in keys if k not in ['index', 'image']]

    idx = question.pop('index', 'XXX')
    text = f'\n- INDEX: {idx}\n'

    if 'image' in question:
        images = question.pop('image')
        if images[0] == '[' and images[-1] == ']':
            images = eval(images)
        else:
            images = [images]
    else:
        images = question.pop('image_path')
        if images[0] == '[' and images[-1] == ']':
            images = eval(images)
        else:
            images = [images]
        images = [encode_image_file_to_base64(x) for x in images]
    
    qtext = question.pop('question', None)
    if qtext is not None:
        text += f'- QUESTION: {qtext}\n'

    if 'A' in question:
        text += f'- Choices: \n'
        for k in string.ascii_uppercase:
            if k in question:
                text += f'\t-{k}: {question.pop(k)}\n'
    answer = question.pop('answer', None)
    
    for k in question:
        if not pd.isna(question[k]):
            text += f'- {k.upper()}. {question[k]}\n'
    
    if answer is not None:
        text += f'- ANSWER: {answer}\n'

    image_md = images_to_md(images)

    return text, image_md


def parse_args():
    parser = argparse.ArgumentParser()
    # Essential Args, Setting the Names of Datasets and Models
    parser.add_argument('--port', type=int, default=7860)
    args = parser.parse_args()
    return args


def gradio_app_vis_dataset(port=7860):
    data, loaded_obj = None, {}

    def btn_submit_click(filename, ann_id):
        if filename not in loaded_obj:
            return filename_change(filename, ann_id)
        nonlocal data
        data_desc = gr.Markdown(f'Visualizing {filename}, {len(data)} samples in total. ')
        if ann_id < 0 or ann_id >= len(data):
            return filename, ann_id, data_desc, gr.Markdown('Invalid Index'), gr.Markdown(f'Index out of range [0, {len(data) - 1}]')
        item = data.iloc[ann_id]
        text, image_md = mmqa_display(item)
        return filename, ann_id, data_desc, image_md, text

    def btn_next_click(filename, ann_id):
        return btn_submit_click(filename, ann_id + 1)

    # def translate_click(anno_en):
    #     return gr.Markdown(Translator.generate(anno_en))

    def filename_change(filename, ann_id):
        nonlocal data, loaded_obj

        def legal_filename(filename):
            LMURoot = LMUDataRoot()
            if filename in SUPPORTED_DATASETS:
                return build_dataset(filename).data
            elif osp.exists(filename):
                data = load(filename)
                assert 'index' in data and 'image' in data
                image_map = {i: image for i, image in zip(data['index'], data['image'])}
                for k, v in image_map.items():
                    if (not isinstance(v, str) or len(v) < 64) and v in image_map:
                        image_map[k] = image_map[v]
                data['image'] = [image_map[k] for k in data['index']]
                return data
            elif osp.exists(osp.join(LMURoot, filename)):
                filename = osp.join(LMURoot, filename)
                return legal_filename(filename)
            else:
                return None

        data = legal_filename(filename)
        if data is None:
            return filename, 0, gr.Markdown(''), gr.Markdown("File not found"), gr.Markdown("File not found")
        
        loaded_obj[filename] = data
        return btn_submit_click(filename, 0)

    with gr.Blocks() as app:
        
        filename = gr.Textbox(
            value='Dataset Name (supported by VLMEvalKit) or TSV FileName (Relative under `LMURoot` or Real Path)', 
            label='Dataset', 
            interactive=True,
            visible=True)
            
        with gr.Row():
            ann_id = gr.Number(0, label='Sample Index (Press Enter)', interactive=True, visible=True)
            btn_next = gr.Button("Next")
            # btn_translate = gr.Button('CN Translate')

        with gr.Row():
            data_desc = gr.Markdown('Dataset Description', label='Dataset Description')
        
        with gr.Row():
            image_output = gr.Markdown('Image PlaceHolder', label='Image Visualization')
            anno_en = gr.Markdown('Image Annotation', label='Image Annotation')
            # anno_cn = gr.Markdown('Image Annotation (Chinese)', label='Image Annotation (Chinese)')

        input_components = [filename, ann_id]
        all_components = [filename, ann_id, data_desc, image_output, anno_en]

        filename.submit(filename_change, input_components, all_components)
        ann_id.submit(btn_submit_click, input_components, all_components)
        btn_next.click(btn_next_click, input_components, all_components)
        # btn_translate.click(translate_click, anno_en, anno_cn)

    # app.launch()
    app.launch(server_name='0.0.0.0', debug=True, show_error=True, server_port=port)


if __name__ == "__main__":
    args = parse_args()
    gradio_app_vis_dataset(port=args.port)