app.py

import json
import os
import random
import re
import sys
from datetime import datetime
from pathlib import Path
from typing import List, Optional
from uuid import uuid4

import gradio as gr
import numpy as np
import requests
from datasets import load_dataset
from huggingface_hub import (
    CommitScheduler,
    HfApi,
    InferenceClient,
    login,
    snapshot_download,
)
from PIL import Image
from glob import glob


session_token = os.environ.get("SessionToken")
login(token=session_token, add_to_git_credential=True)

DEFAILT_USERNAME_MESSAGE = "You must be logged in before you start labeling images."
REPO_URL = "glitchbench/GlitchBenchReviewData"
DATASET_URL = "glitchbench/GlitchBench"

SUBMIT_MESSAGE = "✅  Submit the description"
IDUNNU_MESSAGE = "🤷 I don't know what to do with this image"
SKIP_MESSAGE = "⏭️ Skip, for now"

glitchbench_dataset = load_dataset(DATASET_URL)["validation"]
dataset_size = len(glitchbench_dataset)

# map id to index:
id_to_index = {x["id"]: i for i, x in enumerate(glitchbench_dataset)}

JSON_DATASET_DIR = Path("local_dataset")
JSON_DATASET_DATA_DIR = JSON_DATASET_DIR / "data"
JSON_DATASET_PATH = JSON_DATASET_DATA_DIR / f"labels-{uuid4()}.json"


INSTRUCTIONS = """
            <div>
                <section>
                    <h2>Introduction</h2>
                    <p>Welcome to the GlitchBench Dataset Labeling Tool! Your contribution is vital in helping us clean and label the dataset accurately. This tool is designed to be intuitive and user-friendly, allowing you to view images from the dataset and provide descriptions.</p>
                </section>

                <section>
                    <h2>How to Label Images</h2>
                    <ol>
                        <li><strong>Login:</strong> Use the login button at the top to login with your Hugging Face Account.</li>
                        <li><strong>Start Labeling:</strong> Click the 'Start Labeling' button to begin.</li>
                        <li><strong>View Image:</strong> An image from the GlitchBench dataset will be displayed.</li>
                        <li><strong>Briefly describe in 1-2 sentences what you find <strong>wrong or unusual</strong> about this game scene. <strong>There is at least one thing wrong/unusual </strong>.</li>
                        <li><strong>Submit:</strong> Click 'Submit' to save your description. You can also use 'I Don't Know' if you can't spot the bug or unusual part of the image, or 'Skip' if you don't want to provide a description for the given image.</li>
                        <li><strong>Continue:</strong> Move on to the next image and repeat the process.</li>
                    </ol>
                </section>


                <section>
                    <h2>Examples</h2>
                    <table style="width:100%">
                        <tr>
                            <th>Image</th>
                            <th>Sample Description</th>
                            <th>Image</th>
                            <th>Sample Description</th>
                        </tr>
                        <tr>
                            <td><img src="https://glitchbench.github.io/static/instructions/flying-horse.jpeg" alt="flying horse" style="width:100%; max-width:300px;"></td>
                            <td>"A horse floating in the air", or "A person riding a horse flying in the air"</td>
                            <td><img src="https://glitchbench.github.io/static/instructions/t-pose2.jpeg" alt="t-pose" style="width:100%; max-width:300px;"></td>
                            <td>"A character performing a T-pose"</td>
                        </tr>
                        <tr>
                            <td><img src="https://glitchbench.github.io/static/instructions/a-dog-clipping-with-door.webp" alt="dog in the door" style="width:100%; max-width:300px;"></td>
                            <td>"A dog is clipping through the door"</td>
                            <td><img src="https://glitchbench.github.io/static/instructions/low-poly-face.avif" alt="low-poly face" style="width:100%; max-width:300px;"></td>
                            <td>"A person with a low-poly face"</td>
                        </tr>
                        <tr>
                            <td><img src="https://glitchbench.github.io/static/instructions/unnatural-hand-position.jpeg" alt="unnatural hands" style="width:100%; max-width:300px;"></td>
                            <td>"A person with unnatural hand positions."</td>
                            <td><img src="https://glitchbench.github.io/static/instructions/stretched_neck.jpeg" alt="unnatural neck" style="width:100%; max-width:300px;"></td>
                            <td>"A person with a stretched neck."</td>
                        </tr>
                        <tr>
                            <td><img src="https://glitchbench.github.io/static/instructions/placeholder.png" alt="unnatural hands" style="width:100%; max-width:300px;"></td>
                            <td>"A tree has a placeholder texture."</td>
                            <td><img src="https://glitchbench.github.io/static/instructions/low-res-texture.jpeg" alt="unnatural neck" style="width:100%; max-width:300px;"></td>
                            <td>"The ground has a low-resolution texture."</td>
                        </tr>
                    </table>
                </section>


            </div>
        """


if not JSON_DATASET_DIR.exists():
    JSON_DATASET_DIR.mkdir()

if not JSON_DATASET_DATA_DIR.exists():
    JSON_DATASET_DATA_DIR.mkdir()

print("Downloading the dataset")
print(REPO_URL)

snapshot_download(
    repo_id=REPO_URL,
    allow_patterns="*.json",
    local_dir=JSON_DATASET_DIR,
    use_auth_token=session_token,
    repo_type="dataset",
)

scheduler = CommitScheduler(
    repo_id=REPO_URL,
    repo_type="dataset",
    folder_path=JSON_DATASET_DIR,
    path_in_repo="./",
    every=1,
    private=True,
)


def save_json(image_id: str, provided_description: str, username: str) -> None:
    with scheduler.lock:
        with JSON_DATASET_PATH.open("a") as f:
            json.dump(
                {
                    "username": username,
                    "image_id": image_id,
                    "user_description": provided_description,
                    "datetime": datetime.now().isoformat(),
                },
                f,
            )
            f.write("\n")


def set_username(profile: Optional[gr.OAuthProfile]) -> str:
    if profile is None:
        return DEFAILT_USERNAME_MESSAGE
    return profile["preferred_username"]


def start_labeling(username_label):
    if username_label == DEFAILT_USERNAME_MESSAGE:
        raise gr.Error("Please login first, then click start labeling")

    all_json_files = glob(str(JSON_DATASET_DATA_DIR / "*.json"))
    # read json files and keep records related to the current user
    all_user_records = []

    for json_file in all_json_files:
        with open(json_file) as f:
            for line in f:
                record = json.loads(line)
                if record["username"] == username_label:
                    all_user_records.append(record["image_id"])

    print(f"Found {len(all_user_records)} records for user {username_label}")

    # go throught all images in the dataset and exlcude those that are already labeled by the user
    remaining_indicies = set(range(dataset_size))
    solved_indices = [id_to_index[x] for x in all_user_records]
    remaining_indicies = remaining_indicies - set(solved_indices)

    print(f"Found {len(remaining_indicies)} remaining images for user {username_label}")

    return list(remaining_indicies), gr.Button(interactive=False)


def show_random_sample(username_label, remaining_batch):
    if remaining_batch is None or not remaining_batch:
        raise gr.Error("Make sure you are logged in.")
    
    rindex = random.choice(remaining_batch)
    remaining_batch.remove(rindex)

    # get the image
    image = glitchbench_dataset[rindex]["image"]
    image_id = glitchbench_dataset[rindex]["id"]

    return image, image_id, "", remaining_batch


def write_user_description(username_label, image_id, user_description, skip_or_submit):
    if skip_or_submit == SKIP_MESSAGE:
        return

    if skip_or_submit == IDUNNU_MESSAGE:
        provided_description = "N/A"
    else:
        provided_description = user_description

    save_json(image_id, provided_description, username_label)


with gr.Blocks() as demo:
    gr.Markdown("## GlitchBench Dataset Labeling Tool")
    gr.Markdown("Help us to improve our GlitchBench dataset.")

    with gr.Accordion("Instructions"):
        gr.HTML(INSTRUCTIONS)

    with gr.Row():
        username_label = gr.Text(label="Username", interactive=False)
        gr.LoginButton()
        gr.LogoutButton()

    start_button = gr.Button("Start Labeling")

    gr.HTML(
        "<strong>Briefly describe in 1-2 sentences what you find <strong>wrong or unusual</strong> about this game scene. <strong>There is at least one thing wrong/unusual </strong>."
    )

    username_label.attach_load_event(set_username, None)

    with gr.Row():
        with gr.Column(scale=5, min_width=500):
            glitch_image = gr.Image(label="Image")
            glitch_image_id = gr.Textbox(label="Image ID", visible=False)
        with gr.Column(scale=3, min_width=200):
            user_description = gr.Textbox(lines=5, label="Description")
            submit_btn = gr.Button(SUBMIT_MESSAGE)
            idunnu_btn = gr.Button(IDUNNU_MESSAGE)
            skip_btn = gr.Button(SKIP_MESSAGE)

    remaining_batch = gr.State()

    start_button.click(
        start_labeling, inputs=[username_label], outputs=[remaining_batch, start_button]
    ).then(
        show_random_sample,
        inputs=[username_label, remaining_batch],
        outputs=[glitch_image, glitch_image_id, user_description],
    )

    submit_btn.click(
        write_user_description,
        inputs=[username_label, glitch_image_id, user_description, submit_btn],
        outputs=[],
    ).then(
        show_random_sample,
        inputs=[username_label, remaining_batch],
        outputs=[glitch_image, glitch_image_id, user_description],
    )

    idunnu_btn.click(
        write_user_description,
        inputs=[username_label, glitch_image_id, user_description, idunnu_btn],
        outputs=[],
    ).then(
        show_random_sample,
        inputs=[username_label, remaining_batch],
        outputs=[glitch_image, glitch_image_id, user_description],
    )

    skip_btn.click(
        write_user_description,
        inputs=[username_label, glitch_image_id, user_description, skip_btn],
        outputs=[],
    ).then(
        show_random_sample,
        inputs=[username_label, remaining_batch],
        outputs=[glitch_image, glitch_image_id, user_description],
    )

demo.launch()