CLEVR-Sudoku.py

import os
import json

import datasets

_DESCRIPTION = """\
CLEVR-Sudoku is a dataset for the task of Sudoku puzzle solving. It is a synthetic dataset generated using the CLEVR engine. The dataset consists of 3x3 Sudoku puzzles with varying levels of difficulty. The dataset is divided into three categories based on the number of known cells in the puzzle: Easy (K10), Medium (K30), and Hard (K50). Each puzzle is accompanied by a set of 10 possible solutions. The dataset is generated using the CLEVR engine and is available in the form of images and JSON files. The images are 256x256 pixels in size and are stored in the PNG format. The JSON files contain the puzzle, the solution, and the possible solutions in the form of a dictionary. The dataset is available for download in the form of a zip file. The dataset is intended for use in the development of machine learning models for the task of Sudoku puzzle solving.
"""

_CITATION = """\
@article{stammer2024neural,
  title={Neural Concept Binder},
  author={Stammer, Wolfgang and W{\"u}st, Antonia and Steinmann, David and Kersting, Kristian},
  journal={Advances in Neural Information Processing Systems},
  year={2024}
}"""

_HOME_PAGE = "https://ml-research.github.io/NeuralConceptBinder/"
_IMAGES_URL = "https://huggingface.co/datasets/AIML-TUDA/CLEVR-Sudoku/resolve/main"
_LICENSE = "cc-by-4.0"
_DIR = _IMAGES_URL

_URL_DATA = {
    "CLEVR-Easy-K10": [f"{_DIR}/CLEVR-Easy-Sudokus-K10.zip", f"{_DIR}/CLEVR-Easy-1.zip"],
    "CLEVR-Easy-K30": [f"{_DIR}/CLEVR-Easy-Sudokus-K30.zip", f"{_DIR}/CLEVR-Easy-1.zip"],
    "CLEVR-Easy-K50": [f"{_DIR}/CLEVR-Easy-Sudokus-K50.zip", f"{_DIR}/CLEVR-Easy-1.zip"],
    "CLEVR-4-K10": [f"{_DIR}/CLEVR-4-Sudokus-K10.zip", f"{_DIR}/sudoku.zip"],
    "CLEVR-4-K30": [f"{_DIR}/CLEVR-4-Sudokus-K30.zip", f"{_DIR}/sudoku.zip"],
    "CLEVR-4-K50": [f"{_DIR}/CLEVR-4-Sudokus-K50.zip", f"{_DIR}/sudoku.zip"],
}


class CLEVRSudokuConfig(datasets.BuilderConfig):
    """Builder Config for CLEVR-Sudoku."""

    def __init__(self, data_url, image_url, **kwargs):
        """Builder Config for CLEVR-Sudoku.
        Args:
          metadata_urls: dictionary with keys 'train' and 'validation' containing the archive metadata URLs
          **kwargs: keyword arguments forwarded to super.
        """
        super(CLEVRSudokuConfig, self).__init__(
            version=datasets.Version("1.0.0"), **kwargs
        )
        # if isinstance(data_url, dict):
        #     self.metadata_urls = data_url
        # else:
        self.data_url = data_url
        self.image_url = image_url
        self.metadata_urls = {"train": data_url, "test": None}


class CLEVRSudoku(datasets.GeneratorBasedBuilder):

    BUILDER_CONFIGS = [
        CLEVRSudokuConfig(
            name=name, description=name, data_url=urls[0], image_url=urls[1]
        )
        for name, urls in _URL_DATA.items()
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "sudoku": datasets.Sequence(datasets.Sequence(datasets.Image())),
                    "options": datasets.Sequence(datasets.Sequence(datasets.Image())),
                    # attributes as dict of features
                    "attributes": {
                        "1": datasets.Sequence(datasets.Value("string")),
                        "2": datasets.Sequence(datasets.Value("string")),
                        "3": datasets.Sequence(datasets.Value("string")),
                        "4": datasets.Sequence(datasets.Value("string")),
                        "5": datasets.Sequence(datasets.Value("string")),
                        "6": datasets.Sequence(datasets.Value("string")),
                        "7": datasets.Sequence(datasets.Value("string")),
                        "8": datasets.Sequence(datasets.Value("string")),
                        "9": datasets.Sequence(datasets.Value("string")),
                    },
                    "id": datasets.Value("int32"),
                    "solution": datasets.Sequence(datasets.Sequence(datasets.Value("int32"))),
                }
            ),
            supervised_keys=None,
            homepage=_HOME_PAGE,
            citation=_CITATION,
            license=_LICENSE,
        )

    
    def _split_generators(self, dl_manager):

        # Download and extract images
        image_path = dl_manager.download_and_extract(self.config.image_url)

        # Download and extract the dataset archive
        archive_path = dl_manager.download_and_extract(self.config.data_url)

        # Define the dataset splits
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"archive_path": archive_path, "image_path": image_path, "dl_manager": dl_manager}
            )
        ]
        

    def _generate_examples(self, archive_path, image_path, dl_manager):
        """Yields examples from the archive, assuming JSON files inside a 'json' folder."""

        # list items in archive path
        print(os.listdir(archive_path))

        dirs = [d for d in os.listdir(archive_path) if d != "__MACOSX"]
        top_dir = dirs[0]

        json_dir = os.path.join(archive_path, top_dir, "json")
        json_files = os.listdir(json_dir)

        for i, file_name in enumerate(json_files):

            print(i, file_name)

            with open(os.path.join(archive_path, top_dir, "json", file_name), 'r') as f:
                # Read and parse the JSON content
                json_content = json.load(f)

            # sudoku image paths
            sudoku_images = json_content.get("images")
            sudoku_images = [os.path.join(image_path, x) if x is not None else None for row in sudoku_images for x in row]
            sudoku_images = [sudoku_images[i:i+9] for i in range(0, len(sudoku_images), 9)]

            # option image paths
            option_images = json_content.get("options")
            max_options = len(option_images[0])
            option_images = [os.path.join(image_path, x) if x is not None else None for row in option_images for x in row]
            option_images = [option_images[i:i+max_options] for i in range(0, len(option_images), max_options)]

            # Extract the specific fields from the JSON
            extracted_data = {
                "sudoku": sudoku_images,
                "options": option_images,
                "attributes": json_content.get("map_number_to_attributes"),
                "file_name": file_name,        # The name of the file inside the archive
                "id": i,  # Extract the 'id' field
                "name": json_content.get("name"),  # Extract the 'name' field
                "solution": json_content.get("solution")  # Extract the 'puzzle' field
            }
            
            # Yield the extracted data
            yield i, extracted_data