SBI-16-3D.py

import os
import random
from glob import glob
import json
from huggingface_hub import hf_hub_download


from astropy.io import fits
import datasets
from datasets import DownloadManager
from fsspec.core import url_to_fs

_DESCRIPTION = (
  """SBI-16-3D is a dataset which is part of the AstroCompress project. """
  """It contains data assembled from the James Webb Space Telescope (JWST). """
  """<TODO>Describe data format</TODO>"""
)

_HOMEPAGE = "https://google.github.io/AstroCompress"

_LICENSE = "CC BY 4.0"

_URL = "https://huggingface.co/datasets/AstroCompress/SBI-16-3D/resolve/main/"

_URLS = {
    "tiny": {
        "train": "./splits/tiny_train.jsonl",
        "test": "./splits/tiny_test.jsonl",
    },
    "full": {
        "train": "./splits/full_train.jsonl",
        "test": "./splits/full_test.jsonl",
    }
}

_REPO_ID = "AstroCompress/SBI-16-3D"

class GBI_16_4D(datasets.GeneratorBasedBuilder):
    """GBI-16-4D Dataset"""

    VERSION = datasets.Version("1.0.0")

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="tiny",
            version=VERSION,
            description="A small subset of the data, to test downsteam workflows.",
        ),
        datasets.BuilderConfig(
            name="full",
            version=VERSION,
            description="The full dataset",
        ),
    ]

    DEFAULT_CONFIG_NAME = "tiny"

    def __init__(self, **kwargs):
        super().__init__(version=self.VERSION, **kwargs)

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "image": datasets.Array3D(shape=(None, 2048, 2048), dtype="uint16"),
                    "ra": datasets.Value("float64"),
                    "dec": datasets.Value("float64"),
                    "pixscale": datasets.Value("float64"),
                    "ntimes": datasets.Value("int64"),
                    "image_id": datasets.Value("string"),
                }
            ),
            supervised_keys=None,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation="TBD",
        )

    def _split_generators(self, dl_manager: DownloadManager):

        ret = []
        base_path = dl_manager._base_path
        locally_run = not base_path.startswith(datasets.config.HF_ENDPOINT) 
        _, path = url_to_fs(base_path)

        for split in ["train", "test"]:
            if locally_run:
                split_file_location = os.path.normpath(os.path.join(path, _URLS[self.config.name][split]))
                split_file = dl_manager.download_and_extract(split_file_location)
            else:
                split_file = hf_hub_download(repo_id=_REPO_ID, filename=_URLS[self.config.name][split], repo_type="dataset")
            with open(split_file, encoding="utf-8") as f:
                data_filenames = []
                data_metadata = []
                for line in f:
                    item = json.loads(line)
                    data_filenames.append(item["image"])
                    data_metadata.append({"ra": item["ra"], 
                                          "dec": item["dec"], 
                                          "pixscale": item["pixscale"], 
                                          "ntimes": item["ntimes"], 
                                          "image_id": item["image_id"]})
                if locally_run:
                    data_urls = [os.path.normpath(os.path.join(path,data_filename)) for data_filename in data_filenames]
                    data_files = [dl_manager.download(data_url) for data_url in data_urls]
                else:
                    data_urls = data_filenames
                    data_files = [hf_hub_download(repo_id=_REPO_ID, filename=data_url, repo_type="dataset") for data_url in data_urls]
            ret.append(
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN if split == "train" else datasets.Split.TEST,
                    gen_kwargs={"filepaths": data_files,
                                "split_file": split_file,
                                "split": split, 
                                "data_metadata": data_metadata},
                ),
            )
        return ret

    def _generate_examples(self, filepaths, split_file, split, data_metadata):
        """Generate GBI-16-4D examples"""

        for idx, (filepath, item) in enumerate(zip(filepaths, data_metadata)):
            task_instance_key = f"{self.config.name}-{split}-{idx}"
            with fits.open(filepath, memmap=False) as hdul:
                # the first axis is length one, so we take the first element
                # the second axis is the time axis and varies between images
                image_data = hdul["SCI"].data[0,:,:,:].tolist()
                yield task_instance_key, {**{"image": image_data}, **item}

def make_split_jsonl_files(config_type="tiny", data_dir="./data", 
                        outdir="./splits", seed=42):
    """
    Create jsonl files for the SBI-16-3D dataset.

    config_type: str, default="tiny"
        The type of split to create. Options are "tiny" and "full".
    data_dir: str, default="./data"
        The directory where the FITS files are located.
    outdir: str, default="./splits"
        The directory where the jsonl files will be created.
    seed: int, default=42
        The seed for the random split.
    """
    random.seed(seed)
    os.makedirs(outdir, exist_ok=True)

    fits_files = glob(os.path.join(data_dir, "*.fits"))
    random.shuffle(fits_files)
    if config_type == "tiny":
        train_files = fits_files[:2]
        test_files = fits_files[2:3]
    elif config_type == "full":
        split_idx = int(0.8 * len(fits_files))
        train_files = fits_files[:split_idx]
        test_files = fits_files[split_idx:]
    else:
        raise ValueError("Unsupported config_type. Use 'tiny' or 'full'.")

    def create_jsonl(files, split_name):
        output_file = os.path.join(outdir, f"{config_type}_{split_name}.jsonl")
        with open(output_file, "w") as out_f:
            for file in files:
                print(file, flush=True, end="...")
                with fits.open(file, memmap=False) as hdul:
                    image_id = os.path.basename(file).split(".fits")[0]
                    ra = hdul["SCI"].header.get('CRVAL1', 0)
                    dec = hdul["SCI"].header.get('CRVAL2', 0)
                    pixscale = hdul["SCI"].header.get('CD1_2', 0.396)
                    ntimes = hdul["SCI"].data.shape[0]
                    item = {"image_id": image_id, "image": file, "ra": ra, "dec": dec, 
                            "pixscale": pixscale, "ntimes": ntimes}
                    out_f.write(json.dumps(item) + "\n")

    create_jsonl(train_files, "train")
    create_jsonl(test_files, "test")