pybullet_img2img.py

import os
from xml.etree import ElementTree as ET
import datasets

_DESCRIPTION = """\
The dataset img2img data.
"""
_NAME = "pybullet_img2img"

_HOMEPAGE = f"https://huggingface.co/datasets/pranjalipathre/{_NAME}"

_LICENSE = ""

_DATA = f"https://huggingface.co/datasets/pranjalipathre/{_NAME}/resolve/main/data/"
# _DATA = f"/home/pranjali/Documents/Research/pybullet/img2img/{_NAME}/data/"

class i2iDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        # sapien-doors sequentail dataset. - 5.6k 
        datasets.BuilderConfig(name="video_00", data_dir=f"{_DATA}video_00.zip"),
        # sapien-doors dataset with domain randomization. - 26.4k 
        datasets.BuilderConfig(name="video_01", data_dir=f"{_DATA}video_01.zip"),
        # sapien-all_objects dataset. - 155k 
        datasets.BuilderConfig(name="video_02", data_dir=f"{_DATA}video_02.zip"),
        # sapien-bottles dataset. - 12.3k 
        datasets.BuilderConfig(name="video_03", data_dir=f"{_DATA}video_03.zip"),
        # sapien dataset. - 2.6k 
        datasets.BuilderConfig(name="video_04", data_dir=f"{_DATA}video_04.zip"),
        # sapien dataset. - 0.8k 
        datasets.BuilderConfig(name="video_05", data_dir=f"{_DATA}video_05.zip"),
        # sapien door (open and closed) trajectories of a scene. - 22.0k 
        datasets.BuilderConfig(name="video_06", data_dir=f"{_DATA}video_06.zip"),
        # sapien door (open) trajectories of a scene. - 17.7k 
        datasets.BuilderConfig(name="video_07", data_dir=f"{_DATA}video_07.zip"),
        # rlbench trajectories of a scene. - 7.637k 
        datasets.BuilderConfig(name="video_08", data_dir=f"{_DATA}video_08.zip"),
        # habitat replicad door (open) trajectories of a scene. - 8.2k 
        datasets.BuilderConfig(name="video_09", data_dir=f"{_DATA}video_09.zip"),
        # rlebnch place sorter trajectories of a scene. - 19k 
        datasets.BuilderConfig(name="video_10", data_dir=f"{_DATA}video_10.zip"),
        # rlebnch window sorter trajectories of a scene. - 5k 
        datasets.BuilderConfig(name="video_11", data_dir=f"{_DATA}video_11.zip"),
        # habitat mp3d door trajectories of a scene. - 9k 
        datasets.BuilderConfig(name="video_12", data_dir=f"{_DATA}video_12.zip"),
        # rlbench charger trajectories of a scene. - 4k 
        datasets.BuilderConfig(name="video_13", data_dir=f"{_DATA}video_13.zip"),
        # rlbench door trajectories of a scene. - 4k 
        datasets.BuilderConfig(name="video_14", data_dir=f"{_DATA}video_14.zip"),
        # real-world cirle sorter trajectories - 4k 
        datasets.BuilderConfig(name="video_15", data_dir=f"{_DATA}video_15.zip"),
        # real-world hexagon sorter trajectories - 4k  
        datasets.BuilderConfig(name="video_16", data_dir=f"{_DATA}video_16.zip"),
        # real-world square sorter trajectories - 4k  
        datasets.BuilderConfig(name="video_17", data_dir=f"{_DATA}video_17.zip"),
        # real-world shape stacker trajectories - 900  
        datasets.BuilderConfig(name="video_18", data_dir=f"{_DATA}video_18.zip"),
        # unity warehouse trajectories - 110  
        datasets.BuilderConfig(name="video_19", data_dir=f"{_DATA}video_19.zip"),
    ]

    DEFAULT_CONFIG_NAME = "video_13"

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "original_image": datasets.Image(),
                    "edit_prompt": datasets.Value("string"),
                    "edited_image": datasets.Image(),
                }
            ),
            supervised_keys=None,
            homepage=_HOMEPAGE,
            citation=None,
        )

    def _split_generators(self, dl_manager):
        data = dl_manager.download_and_extract(self.config.data_dir)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "data": data,
                },
            ),
        ]

    @staticmethod
    def parse_text(root: ET.Element, file: str, index: int) -> dict:
        idx = str(index).zfill(6)
        ele = root.find(f".//*[@frame='{idx}']")
        dt = {
            "text": ele.get("text")
        }
        return dt

    def _generate_examples(self, data):
        treePath = os.path.join(data, "annotations.xml")

        tree = ET.parse(treePath)
        root = tree.getroot()

        for idx, file in enumerate(sorted(os.listdir(os.path.join(data, "original_images")))):
            dat = self.parse_text(root, file, idx)
            txt = dat["text"]

            yield idx, {
                "original_image": os.path.join(data, "original_images", file),
                "edit_prompt": txt,
                "edited_image": os.path.join(data, "edited_images", file),
            }