Datasets:

shijli
/

davis2017

# coding=utf-8
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

"""DAVIS 2017 (semi-supervised video object segmentation), packaged as a Hugging Face dataset.

One row per frame, so a split streams frame by frame and a single frame is one lookup; the rows of
a video are contiguous and in frame order, and carry the video name, so a video is a groupby:

    load_dataset('<user>/davis2017', 'semi-supervised')    # 60 train / 30 val videos, 480p

Images are the official 480p JPEGs and masks the official annotation PNGs, both copied through
byte for byte. Masks are palette (mode P) PNGs whose pixel value is the object id of the video
(0 = background, 1..N = objects, 255 = void), which is the encoding DAVIS ships and the one its
evaluation reads. Object ids are per video, not classes: object 1 of one video has nothing to do
with object 1 of another.

Only the trainval set is packed. The test-dev and test-challenge sets ship first-frame masks only,
so they cannot be scored offline.

Source data comes from an unpacked DAVIS folder when one is given, and is downloaded from the
official host otherwise:

    python create_dataset.py --repo <user>/davis2017 --davis-root ~/Data/datasets/DAVIS --private
    DAVIS_ROOT=~/Data/datasets/DAVIS python create_dataset.py --repo <user>/davis2017

Build without uploading anything, to check the result first:

    python create_dataset.py --dry-run --limit 2
"""

import argparse
import glob
import os

import datasets
import numpy as np
from PIL import Image

_VERSION = "1.0.0"
_HOMEPAGE_URL = "https://davischallenge.org/davis2017/code.html"
_DOWNLOAD_URL = "https://data.vision.ee.ethz.ch/csergi/share/davis/DAVIS-2017-trainval-480p.zip"

_CITATION = """\
@article{Pont-Tuset_arXiv_2017,
    author = {Jordi Pont-Tuset and Federico Perazzi and Sergi Caelles and Pablo Arbel\\'aez and
              Alexander Sorkine-Hornung and Luc {Van Gool}},
    title = {The 2017 DAVIS Challenge on Video Object Segmentation},
    journal = {arXiv:1704.00675},
    year = {2017}
}
@inproceedings{Perazzi_CVPR_2016,
    author = {Federico Perazzi and Jordi Pont-Tuset and Brian McWilliams and Luc {Van Gool} and
              Markus Gross and Alexander Sorkine-Hornung},
    title = {A Benchmark Dataset and Evaluation Methodology for Video Object Segmentation},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2016}
}
"""

_LICENSE = """\
DAVIS 2017 is released under CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/). The
videos come from online sources listed in SOURCES.md, which keep their own terms of use. This
packaging adds no rights: use it under those same terms and cite the original work.
"""

_DESCRIPTION = """\
DAVIS 2017 semi-supervised video object segmentation at 480p, one row per frame: the 60 train and
30 val videos of the official trainval set with their multi-object masks. Masks are palette PNGs
whose pixel value is the object id within the video (0 background, 255 void).
"""

SEMI_SUPERVISED = "semi-supervised"

# Videos and frames of the official 2017 split lists. Checked against what was built before
# anything is pushed: a stale cache or a half-unpacked root shows up first as wrong counts.
EXPECTED_SIZES = {SEMI_SUPERVISED: {"train": 4209, "validation": 1999}}
EXPECTED_VIDEOS = {SEMI_SUPERVISED: {"train": 60, "validation": 30}}
_VOID_INDEX = 255


def read_split(root, split):
    """The video names of one split, read from the ImageSets list that names it."""
    with open(os.path.join(root, "ImageSets", "2017", split + ".txt")) as f:
        return [line.strip() for line in f if line.strip()]


def first_frame_objects(mask_path):
    """Number of objects of a video: the largest id of its first-frame mask, void excluded."""
    mask = np.array(Image.open(mask_path))
    ids = mask[mask != _VOID_INDEX]
    return int(ids.max()) if ids.size else 0


class DAVIS2017Config(datasets.BuilderConfig):
    """One task view over the DAVIS videos."""

    def __init__(self, resolution="480p", davis_root=None, limit=None, **kwargs):
        super().__init__(version=datasets.Version(_VERSION, ""), **kwargs)
        self.resolution = resolution
        # davis_root and limit are config fields so that they reach the builder and enter the
        # config id: a truncated build lands in its own cache directory and is never served back
        # to a full one.
        self.davis_root = davis_root
        self.limit = limit


class DAVIS2017(datasets.GeneratorBasedBuilder):
    """DAVIS 2017, one row per frame."""

    BUILDER_CONFIGS = [
        DAVIS2017Config(
            name=SEMI_SUPERVISED,
            resolution="480p",
            description=(
                "Semi-supervised video object segmentation on the official trainval set at 480p, "
                "the resolution of the official evaluation: 60 train and 30 val videos, every frame "
                "with its multi-object mask. The task gives the first-frame mask and scores the rest."
            ),
        )
    ]
    DEFAULT_CONFIG_NAME = SEMI_SUPERVISED

    def _info(self):
        features = {
            "id": datasets.Value("string"),  # <video>/<frame file stem>, e.g. bear/00000
            "video": datasets.Value("string"),
            "frame": datasets.Value("int32"),  # 0-based index in the video
            "image": datasets.Image(),
            # pixel value = object id in the video, 0 background, 255 void. Palette for viewing only.
            "mask": datasets.Image(),
            "num_objects": datasets.Value("int32"),  # objects of the video (its first-frame mask)
        }
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(features),
            supervised_keys=None,
            license=_LICENSE,
            homepage=_HOMEPAGE_URL,
            citation=_CITATION,
        )

    def _davis_root(self, dl_manager):
        """The unpacked DAVIS folder given on the command line or in the environment, or the official
        archive downloaded and extracted on demand."""
        root = self.config.davis_root or os.environ.get("DAVIS_ROOT")
        if root:
            root = os.path.expanduser(root)
            if not os.path.isdir(os.path.join(root, "JPEGImages", self.config.resolution)):
                raise FileNotFoundError(
                    "no JPEGImages/%s under the given DAVIS root: %s" % (self.config.resolution, root)
                )
            return root
        return os.path.join(dl_manager.download_and_extract(_DOWNLOAD_URL), "DAVIS")

    def _split_generators(self, dl_manager):
        root = self._davis_root(dl_manager)
        limit = self.config.limit
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN, gen_kwargs={"root": root, "videos": read_split(root, "train")[:limit]}
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION, gen_kwargs={"root": root, "videos": read_split(root, "val")[:limit]}
            ),
        ]

    def _generate_examples(self, root, videos):
        res = self.config.resolution
        for video in videos:
            frames = sorted(glob.glob(os.path.join(root, "JPEGImages", res, video, "*.jpg")))
            masks = sorted(glob.glob(os.path.join(root, "Annotations", res, video, "*.png")))
            stems = [os.path.splitext(os.path.basename(p))[0] for p in frames]
            if not frames or stems != [os.path.splitext(os.path.basename(p))[0] for p in masks]:
                raise FileNotFoundError("%s: frames and masks do not match one to one" % video)
            num_objects = first_frame_objects(masks[0])
            for index, (stem, image, mask) in enumerate(zip(stems, frames, masks, strict=True)):
                key = "%s/%s" % (video, stem)
                # paths, not decoded images: the Image feature then stores the file bytes unchanged
                yield (
                    key,
                    {
                        "id": key,
                        "video": video,
                        "frame": index,
                        "image": image,
                        "mask": mask,
                        "num_objects": num_objects,
                    },
                )


##############################


def check_dataset(dataset, root, samples=8):
    """Confirm the masks survived packing as labels: mode P, ids within the video's objects (or void),
    paired with the right image, and byte-identical to the source files under ``root`` (when given);
    and that videos are whole, contiguous and in frame order."""
    for split, split_ds in dataset.items():
        videos = split_ds["video"]
        frames = split_ds["frame"]
        seen = []
        for i, (video, frame) in enumerate(zip(videos, frames, strict=True)):
            if not seen or seen[-1] != video:
                assert video not in seen, "%s: video %s is not contiguous" % (split, video)
                assert frame == 0, "%s: video %s does not start at frame 0" % (split, video)
                seen.append(video)
            else:
                assert frame == frames[i - 1] + 1, "%s: %s is out of frame order at row %d" % (split, video, i)
        picks = np.linspace(0, len(split_ds) - 1, min(samples, len(split_ds))).astype(int).tolist()
        raw = split_ds.cast_column("mask", datasets.Image(decode=False)).cast_column(
            "image", datasets.Image(decode=False)
        )
        for i in picks:
            record, stored = split_ds[i], raw[i]
            mask = record["mask"]
            values = np.unique(np.array(mask))
            assert mask.mode == "P", "%s[%d]: mask is mode %s, expected P" % (split, i, mask.mode)
            bad = [v for v in values if v > record["num_objects"] and v != _VOID_INDEX]
            assert not bad, "%s[%d]: mask holds ids beyond the video's %d objects: %s" % (
                split,
                i,
                record["num_objects"],
                bad,
            )
            assert mask.size == record["image"].size, "%s[%d]: mask and image differ in size" % (split, i)
            if root is None:  # built from the downloaded archive: no local files to compare with
                continue
            video, stem = record["id"].split("/")
            for column, sub, ext in (("image", "JPEGImages", ".jpg"), ("mask", "Annotations", ".png")):
                # the local build keeps file references; push_to_hub embeds those files' bytes as they are
                value = stored[column]
                data = value["bytes"] if value["bytes"] is not None else open(value["path"], "rb").read()
                with open(os.path.join(root, sub, "480p", video, stem + ext), "rb") as f:
                    assert data == f.read(), "%s[%d]: %s bytes differ from the source" % (split, i, column)
        print("  %s: %d frames of %d videos, masks and bytes check out" % (split, len(split_ds), len(seen)))


def parse_args():
    parser = argparse.ArgumentParser("Pack DAVIS 2017 as a Hugging Face dataset")
    parser.add_argument("--repo", type=str, default=None, help="dataset repo to push to, e.g. <user>/davis2017")
    parser.add_argument(
        "--davis-root",
        type=str,
        default=None,
        help="unpacked DAVIS folder (the one holding JPEGImages). Downloads the official archive if unset",
    )
    parser.add_argument("--no-push-code", action="store_true", help="skip uploading this script alongside the data")
    parser.add_argument("--private", action="store_true", help="create the repo as private")
    parser.add_argument("--dry-run", action="store_true", help="build and verify locally, push nothing")
    parser.add_argument(
        "--limit",
        type=int,
        default=None,
        help="build only the first N videos of each split, for a check. Needs --dry-run",
    )
    parser.add_argument("--cache-dir", type=str, default=None, help="where to build, defaults to the datasets cache")
    args = parser.parse_args()
    if args.repo is None and not args.dry_run:
        parser.error("give --repo to upload, or --dry-run to build locally")
    if args.limit is not None and not args.dry_run:
        parser.error("--limit builds a truncated dataset, which is only useful with --dry-run")
    return args


def main(args):
    config_name = SEMI_SUPERVISED
    print("\n===== %s =====" % config_name)
    builder = DAVIS2017(
        dataset_name="davis2017",
        config_name=config_name,
        cache_dir=args.cache_dir,
        davis_root=args.davis_root,
        limit=args.limit,
    )
    builder.download_and_prepare()
    dataset = builder.as_dataset()
    print(dataset)

    actual = {split: len(rows) for split, rows in dataset.items()}
    videos = {split: len(set(rows["video"])) for split, rows in dataset.items()}
    if args.limit is None and (actual != EXPECTED_SIZES[config_name] or videos != EXPECTED_VIDEOS[config_name]):
        raise SystemExit(
            "refusing to push: expected %s frames / %s videos, built %s / %s"
            % (EXPECTED_SIZES[config_name], EXPECTED_VIDEOS[config_name], actual, videos)
        )
    root = args.davis_root or os.environ.get("DAVIS_ROOT")
    check_dataset(dataset, os.path.expanduser(root) if root else None)

    if args.dry_run:
        print("dry run: not pushing")
        return

    dataset.push_to_hub(args.repo, config_name=config_name, set_default=True, private=args.private)
    print("pushed %s to %s" % (config_name, args.repo))

    from huggingface_hub import HfApi

    api = HfApi()
    if not args.no_push_code:
        # the script that produced the parquet, kept in the repo so the packing is reproducible
        api.upload_file(
            path_or_fileobj=os.path.abspath(__file__),
            path_in_repo="create_dataset.py",
            repo_id=args.repo,
            repo_type="dataset",
            commit_message="Add the script that packs this dataset",
        )
        print("pushed create_dataset.py to %s" % args.repo)
    sources = os.path.join(os.path.expanduser(root), "SOURCES.md") if root else None
    if sources and os.path.exists(sources):
        # where every video comes from: their terms of use apply alongside the DAVIS license
        api.upload_file(
            path_or_fileobj=sources,
            path_in_repo="SOURCES.md",
            repo_id=args.repo,
            repo_type="dataset",
            commit_message="Add the online sources of the DAVIS videos",
        )
        print("pushed SOURCES.md to %s" % args.repo)


if __name__ == "__main__":
    main(parse_args())

	# coding=utf-8
	#
	# Licensed under the Apache License, Version 2.0 (the "License");
	# you may not use this file except in compliance with the License.
	# You may obtain a copy of the License at
	#
	# http://www.apache.org/licenses/LICENSE-2.0
	#
	# Unless required by applicable law or agreed to in writing, software
	# distributed under the License is distributed on an "AS IS" BASIS,
	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

	"""DAVIS 2017 (semi-supervised video object segmentation), packaged as a Hugging Face dataset.

	One row per frame, so a split streams frame by frame and a single frame is one lookup; the rows of
	a video are contiguous and in frame order, and carry the video name, so a video is a groupby:

	load_dataset('<user>/davis2017', 'semi-supervised') # 60 train / 30 val videos, 480p

	Images are the official 480p JPEGs and masks the official annotation PNGs, both copied through
	byte for byte. Masks are palette (mode P) PNGs whose pixel value is the object id of the video
	(0 = background, 1..N = objects, 255 = void), which is the encoding DAVIS ships and the one its
	evaluation reads. Object ids are per video, not classes: object 1 of one video has nothing to do
	with object 1 of another.

	Only the trainval set is packed. The test-dev and test-challenge sets ship first-frame masks only,
	so they cannot be scored offline.

	Source data comes from an unpacked DAVIS folder when one is given, and is downloaded from the
	official host otherwise:

	python create_dataset.py --repo <user>/davis2017 --davis-root ~/Data/datasets/DAVIS --private
	DAVIS_ROOT=~/Data/datasets/DAVIS python create_dataset.py --repo <user>/davis2017

	Build without uploading anything, to check the result first:

	python create_dataset.py --dry-run --limit 2
	"""

	import argparse
	import glob
	import os

	import datasets
	import numpy as np
	from PIL import Image

	_VERSION = "1.0.0"
	_HOMEPAGE_URL = "https://davischallenge.org/davis2017/code.html"
	_DOWNLOAD_URL = "https://data.vision.ee.ethz.ch/csergi/share/davis/DAVIS-2017-trainval-480p.zip"

	_CITATION = """\
	@article{Pont-Tuset_arXiv_2017,
	author = {Jordi Pont-Tuset and Federico Perazzi and Sergi Caelles and Pablo Arbel\\'aez and
	Alexander Sorkine-Hornung and Luc {Van Gool}},
	title = {The 2017 DAVIS Challenge on Video Object Segmentation},
	journal = {arXiv:1704.00675},
	year = {2017}
	}
	@inproceedings{Perazzi_CVPR_2016,
	author = {Federico Perazzi and Jordi Pont-Tuset and Brian McWilliams and Luc {Van Gool} and
	Markus Gross and Alexander Sorkine-Hornung},
	title = {A Benchmark Dataset and Evaluation Methodology for Video Object Segmentation},
	booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	year = {2016}
	}
	"""

	_LICENSE = """\
	DAVIS 2017 is released under CC BY-NC 4.0 (https://creativecommons.org/licenses/by-nc/4.0/). The
	videos come from online sources listed in SOURCES.md, which keep their own terms of use. This
	packaging adds no rights: use it under those same terms and cite the original work.
	"""

	_DESCRIPTION = """\
	DAVIS 2017 semi-supervised video object segmentation at 480p, one row per frame: the 60 train and
	30 val videos of the official trainval set with their multi-object masks. Masks are palette PNGs
	whose pixel value is the object id within the video (0 background, 255 void).
	"""

	SEMI_SUPERVISED = "semi-supervised"

	# Videos and frames of the official 2017 split lists. Checked against what was built before
	# anything is pushed: a stale cache or a half-unpacked root shows up first as wrong counts.
	EXPECTED_SIZES = {SEMI_SUPERVISED: {"train": 4209, "validation": 1999}}
	EXPECTED_VIDEOS = {SEMI_SUPERVISED: {"train": 60, "validation": 30}}
	_VOID_INDEX = 255


	def read_split(root, split):
	"""The video names of one split, read from the ImageSets list that names it."""
	with open(os.path.join(root, "ImageSets", "2017", split + ".txt")) as f:
	return [line.strip() for line in f if line.strip()]


	def first_frame_objects(mask_path):
	"""Number of objects of a video: the largest id of its first-frame mask, void excluded."""
	mask = np.array(Image.open(mask_path))
	ids = mask[mask != _VOID_INDEX]
	return int(ids.max()) if ids.size else 0


	class DAVIS2017Config(datasets.BuilderConfig):
	"""One task view over the DAVIS videos."""

	def __init__(self, resolution="480p", davis_root=None, limit=None, **kwargs):
	super().__init__(version=datasets.Version(_VERSION, ""), **kwargs)
	self.resolution = resolution
	# davis_root and limit are config fields so that they reach the builder and enter the
	# config id: a truncated build lands in its own cache directory and is never served back
	# to a full one.
	self.davis_root = davis_root
	self.limit = limit


	class DAVIS2017(datasets.GeneratorBasedBuilder):
	"""DAVIS 2017, one row per frame."""

	BUILDER_CONFIGS = [
	DAVIS2017Config(
	name=SEMI_SUPERVISED,
	resolution="480p",
	description=(
	"Semi-supervised video object segmentation on the official trainval set at 480p, "
	"the resolution of the official evaluation: 60 train and 30 val videos, every frame "
	"with its multi-object mask. The task gives the first-frame mask and scores the rest."
	),
	)
	]
	DEFAULT_CONFIG_NAME = SEMI_SUPERVISED

	def _info(self):
	features = {
	"id": datasets.Value("string"), # <video>/<frame file stem>, e.g. bear/00000
	"video": datasets.Value("string"),
	"frame": datasets.Value("int32"), # 0-based index in the video
	"image": datasets.Image(),
	# pixel value = object id in the video, 0 background, 255 void. Palette for viewing only.
	"mask": datasets.Image(),
	"num_objects": datasets.Value("int32"), # objects of the video (its first-frame mask)
	}
	return datasets.DatasetInfo(
	description=_DESCRIPTION,
	features=datasets.Features(features),
	supervised_keys=None,
	license=_LICENSE,
	homepage=_HOMEPAGE_URL,
	citation=_CITATION,
	)

	def _davis_root(self, dl_manager):
	"""The unpacked DAVIS folder given on the command line or in the environment, or the official
	archive downloaded and extracted on demand."""
	root = self.config.davis_root or os.environ.get("DAVIS_ROOT")
	if root:
	root = os.path.expanduser(root)
	if not os.path.isdir(os.path.join(root, "JPEGImages", self.config.resolution)):
	raise FileNotFoundError(
	"no JPEGImages/%s under the given DAVIS root: %s" % (self.config.resolution, root)
	)
	return root
	return os.path.join(dl_manager.download_and_extract(_DOWNLOAD_URL), "DAVIS")

	def _split_generators(self, dl_manager):
	root = self._davis_root(dl_manager)
	limit = self.config.limit
	return [
	datasets.SplitGenerator(
	name=datasets.Split.TRAIN, gen_kwargs={"root": root, "videos": read_split(root, "train")[:limit]}
	),
	datasets.SplitGenerator(
	name=datasets.Split.VALIDATION, gen_kwargs={"root": root, "videos": read_split(root, "val")[:limit]}
	),
	]

	def _generate_examples(self, root, videos):
	res = self.config.resolution
	for video in videos:
	frames = sorted(glob.glob(os.path.join(root, "JPEGImages", res, video, "*.jpg")))
	masks = sorted(glob.glob(os.path.join(root, "Annotations", res, video, "*.png")))
	stems = [os.path.splitext(os.path.basename(p))[0] for p in frames]
	if not frames or stems != [os.path.splitext(os.path.basename(p))[0] for p in masks]:
	raise FileNotFoundError("%s: frames and masks do not match one to one" % video)
	num_objects = first_frame_objects(masks[0])
	for index, (stem, image, mask) in enumerate(zip(stems, frames, masks, strict=True)):
	key = "%s/%s" % (video, stem)
	# paths, not decoded images: the Image feature then stores the file bytes unchanged
	yield (
	key,
	{
	"id": key,
	"video": video,
	"frame": index,
	"image": image,
	"mask": mask,
	"num_objects": num_objects,
	},
	)


	##############################


	def check_dataset(dataset, root, samples=8):
	"""Confirm the masks survived packing as labels: mode P, ids within the video's objects (or void),
	paired with the right image, and byte-identical to the source files under ``root`` (when given);
	and that videos are whole, contiguous and in frame order."""
	for split, split_ds in dataset.items():
	videos = split_ds["video"]
	frames = split_ds["frame"]
	seen = []
	for i, (video, frame) in enumerate(zip(videos, frames, strict=True)):
	if not seen or seen[-1] != video:
	assert video not in seen, "%s: video %s is not contiguous" % (split, video)
	assert frame == 0, "%s: video %s does not start at frame 0" % (split, video)
	seen.append(video)
	else:
	assert frame == frames[i - 1] + 1, "%s: %s is out of frame order at row %d" % (split, video, i)
	picks = np.linspace(0, len(split_ds) - 1, min(samples, len(split_ds))).astype(int).tolist()
	raw = split_ds.cast_column("mask", datasets.Image(decode=False)).cast_column(
	"image", datasets.Image(decode=False)
	)
	for i in picks:
	record, stored = split_ds[i], raw[i]
	mask = record["mask"]
	values = np.unique(np.array(mask))
	assert mask.mode == "P", "%s[%d]: mask is mode %s, expected P" % (split, i, mask.mode)
	bad = [v for v in values if v > record["num_objects"] and v != _VOID_INDEX]
	assert not bad, "%s[%d]: mask holds ids beyond the video's %d objects: %s" % (
	split,
	i,
	record["num_objects"],
	bad,
	)
	assert mask.size == record["image"].size, "%s[%d]: mask and image differ in size" % (split, i)
	if root is None: # built from the downloaded archive: no local files to compare with
	continue
	video, stem = record["id"].split("/")
	for column, sub, ext in (("image", "JPEGImages", ".jpg"), ("mask", "Annotations", ".png")):
	# the local build keeps file references; push_to_hub embeds those files' bytes as they are
	value = stored[column]
	data = value["bytes"] if value["bytes"] is not None else open(value["path"], "rb").read()
	with open(os.path.join(root, sub, "480p", video, stem + ext), "rb") as f:
	assert data == f.read(), "%s[%d]: %s bytes differ from the source" % (split, i, column)
	print(" %s: %d frames of %d videos, masks and bytes check out" % (split, len(split_ds), len(seen)))


	def parse_args():
	parser = argparse.ArgumentParser("Pack DAVIS 2017 as a Hugging Face dataset")
	parser.add_argument("--repo", type=str, default=None, help="dataset repo to push to, e.g. <user>/davis2017")
	parser.add_argument(
	"--davis-root",
	type=str,
	default=None,
	help="unpacked DAVIS folder (the one holding JPEGImages). Downloads the official archive if unset",
	)
	parser.add_argument("--no-push-code", action="store_true", help="skip uploading this script alongside the data")
	parser.add_argument("--private", action="store_true", help="create the repo as private")
	parser.add_argument("--dry-run", action="store_true", help="build and verify locally, push nothing")
	parser.add_argument(
	"--limit",
	type=int,
	default=None,
	help="build only the first N videos of each split, for a check. Needs --dry-run",
	)
	parser.add_argument("--cache-dir", type=str, default=None, help="where to build, defaults to the datasets cache")
	args = parser.parse_args()
	if args.repo is None and not args.dry_run:
	parser.error("give --repo to upload, or --dry-run to build locally")
	if args.limit is not None and not args.dry_run:
	parser.error("--limit builds a truncated dataset, which is only useful with --dry-run")
	return args


	def main(args):
	config_name = SEMI_SUPERVISED
	print("\n===== %s =====" % config_name)
	builder = DAVIS2017(
	dataset_name="davis2017",
	config_name=config_name,
	cache_dir=args.cache_dir,
	davis_root=args.davis_root,
	limit=args.limit,
	)
	builder.download_and_prepare()
	dataset = builder.as_dataset()
	print(dataset)

	actual = {split: len(rows) for split, rows in dataset.items()}
	videos = {split: len(set(rows["video"])) for split, rows in dataset.items()}
	if args.limit is None and (actual != EXPECTED_SIZES[config_name] or videos != EXPECTED_VIDEOS[config_name]):
	raise SystemExit(
	"refusing to push: expected %s frames / %s videos, built %s / %s"
	% (EXPECTED_SIZES[config_name], EXPECTED_VIDEOS[config_name], actual, videos)
	)
	root = args.davis_root or os.environ.get("DAVIS_ROOT")
	check_dataset(dataset, os.path.expanduser(root) if root else None)

	if args.dry_run:
	print("dry run: not pushing")
	return

	dataset.push_to_hub(args.repo, config_name=config_name, set_default=True, private=args.private)
	print("pushed %s to %s" % (config_name, args.repo))

	from huggingface_hub import HfApi

	api = HfApi()
	if not args.no_push_code:
	# the script that produced the parquet, kept in the repo so the packing is reproducible
	api.upload_file(
	path_or_fileobj=os.path.abspath(__file__),
	path_in_repo="create_dataset.py",
	repo_id=args.repo,
	repo_type="dataset",
	commit_message="Add the script that packs this dataset",
	)
	print("pushed create_dataset.py to %s" % args.repo)
	sources = os.path.join(os.path.expanduser(root), "SOURCES.md") if root else None
	if sources and os.path.exists(sources):
	# where every video comes from: their terms of use apply alongside the DAVIS license
	api.upload_file(
	path_or_fileobj=sources,
	path_in_repo="SOURCES.md",
	repo_id=args.repo,
	repo_type="dataset",
	commit_message="Add the online sources of the DAVIS videos",
	)
	print("pushed SOURCES.md to %s" % args.repo)


	if __name__ == "__main__":
	main(parse_args())