import datasets import pandas as pd import os from pathlib import Path from tqdm import tqdm _CITATION = """\ @article{liu2025causal3d, title={CAUSAL3D: A Comprehensive Benchmark for Causal Learning from Visual Data}, author={Liu, Disheng and Qiao, Yiran and Liu, Wuche and Lu, Yiren and Zhou, Yunlai and Liang, Tuo and Yin, Yu and Ma, Jing}, journal={arXiv preprint arXiv:2503.04852}, year={2025} } """ _DESCRIPTION = """\ Causal3D is a benchmark for evaluating causal reasoning in physical and hypothetical visual scenes. It includes both real-world recordings and rendered synthetic scenes demonstrating causal interactions. """ _HOMEPAGE = "https://huggingface.co/datasets/LLDDSS/Causal3D" _LICENSE = "CC-BY-4.0" class Causal3dDataset(datasets.GeneratorBasedBuilder): DEFAULT_CONFIG_NAME = "Real_Water_flow" BUILDER_CONFIGS = [ # hypothetical_scenes datasets.BuilderConfig(name="Hypothetical_V2_linear", version=datasets.Version("1.0.0"), description="Hypothetical_V2_linear scene"), datasets.BuilderConfig(name="Hypothetical_V2_nonlinear", version=datasets.Version("1.0.0"), description="Hypothetical_V2_nonlinear scene"), datasets.BuilderConfig(name="Hypothetical_V3_fully_connected_linear", version=datasets.Version("1.0.0"), description="Hypothetical_V3_fully_connected_linear scene"), datasets.BuilderConfig(name="Hypothetical_V3_v_structure_linear", version=datasets.Version("1.0.0"), description="Hypothetical_V3_v_structure_linear scene"), datasets.BuilderConfig(name="Hypothetical_V3_v_structure_nonlinear", version=datasets.Version("1.0.0"), description="Hypothetical_V3_v_structure_nonlinear scene"), datasets.BuilderConfig(name="Hypothetical_V4_linear", version=datasets.Version("1.0.0"), description="Hypothetical_V4_linear scene"), datasets.BuilderConfig(name="Hypothetical_V4_v_structure_nonlinear", version=datasets.Version("1.0.0"), description="Hypothetical_V4_v_structure_nonlinear scene"), datasets.BuilderConfig(name="Hypothetical_V4_v_structure_linear", version=datasets.Version("1.0.0"), description="Hypothetical_V4_v_structure_linear scene"), datasets.BuilderConfig(name="Hypothetical_V5_linear", version=datasets.Version("1.0.0"), description="Hypothetical_V5_linear scene"), datasets.BuilderConfig(name="Hypothetical_V5_v_structure_linear", version=datasets.Version("1.0.0"), description="Hypothetical_V5_v_structure_linear scene"), datasets.BuilderConfig(name="Hypothetical_V5_v_structure_nonlinear", version=datasets.Version("1.0.0"), description="Hypothetical_V5_v_structure_nonlinear scene"), # real_scenes datasets.BuilderConfig(name="Real_Parabola", version=datasets.Version("1.0.0"), description="Real_Parabola scene"), datasets.BuilderConfig(name="Real_Magnet", version=datasets.Version("1.0.0"), description="Real_Magnet scene"), datasets.BuilderConfig(name="Real_Spring", version=datasets.Version("1.0.0"), description="Real_Spring scene"), datasets.BuilderConfig(name="Real_Water_flow", version=datasets.Version("1.0.0"), description="Real_Water_flow scene"), datasets.BuilderConfig(name="Real_Seesaw", version=datasets.Version("1.0.0"), description="Real_Seesaw scene"), datasets.BuilderConfig(name="Real_Reflection", version=datasets.Version("1.0.0"), description="Real_Reflection scene"), datasets.BuilderConfig(name="Real_Pendulum", version=datasets.Version("1.0.0"), description="Real_Pendulum scene"), datasets.BuilderConfig(name="Real_Convex_len", version=datasets.Version("1.0.0"), description="Real_Convex_len scene"), # multi_view_scenes datasets.BuilderConfig(name="MV_Pendulum", version=datasets.Version("1.0.0"), description="Multi_View_Real_Pendulum scene"), datasets.BuilderConfig(name="MV_H3_v_structure_linear", version=datasets.Version("1.0.0"), description="MV_H3_v_structure_linear scene"), datasets.BuilderConfig(name="MV_H2_linear", version=datasets.Version("1.0.0"), description="MV_H2_linear scene"), datasets.BuilderConfig(name="MV_H2_nonlinear", version=datasets.Version("1.0.0"), description="MV_H2_nonlinear scene"), datasets.BuilderConfig(name="MV_H4_fully_connected_linear", version=datasets.Version("1.0.0"), description="MV_H4_fully_connected_linear scene"), datasets.BuilderConfig(name="MV_H4_v_structure_linear", version=datasets.Version("1.0.0"), description="MV_H4_v_structure_linear scene"), datasets.BuilderConfig(name="MV_H4_v_structure_nonlinear", version=datasets.Version("1.0.0"), description="MV_H4_v_structure_nonlinear scene"), datasets.BuilderConfig(name="MV_H5_fully_connected_linear", version=datasets.Version("1.0.0"), description="MV_H5_fully_connected_linear scene"), datasets.BuilderConfig(name="MV_H5_v_structure_linear", version=datasets.Version("1.0.0"), description="MV_H5_v_structure_linear scene"), datasets.BuilderConfig(name="MV_H5_v_structure_nonlinear", version=datasets.Version("1.0.0"), description="MV_H5_v_structure_nonlinear scene"), ] def _info(self): return datasets.DatasetInfo( description=_DESCRIPTION, features=datasets.Features({ "image": datasets.Image(), "file_name": datasets.Value("string"), "metadata": datasets.Value("string"), # optionally replace with structured fields }), homepage=_HOMEPAGE, license=_LICENSE, citation=_CITATION, ) def _split_generators(self, dl_manager): print(">>>>>>>>>>>>>>>>>>>>>>> Starting to load dataset <<<<<<<<<<<<<<<<<<<<<<<") parts = self.config.name.split("_", 1) # 🚩 Real_Parabola -> ["Real", "Parabola"] category = parts[0] scene = parts[1] local_scene_dir = os.path.join(category, scene) if os.path.exists(local_scene_dir): data_dir = local_scene_dir print(f"Using local folder: {data_dir}") else: zip_name = f"{self.config.name}.zip" archive_path = dl_manager.download_and_extract(zip_name) data_dir = archive_path print(f"Downloaded and extracted: {zip_name}") return [ datasets.SplitGenerator( name=datasets.Split.TRAIN, gen_kwargs={"data_dir": data_dir}, ) ] def _generate_examples(self, data_dir): image_files = {} for ext in ("*.png", "*.jpg", "*.jpeg"): for img_path in Path(data_dir).rglob(ext): relative = str(img_path.relative_to(data_dir)) image_files[relative] = str(img_path) csv_files = [f for f in Path(data_dir).rglob("*.csv") if not f.name.startswith("._")] df = pd.read_csv(csv_files[0]) if csv_files else None if df is not None and "imgs" in df.columns: images = df["imgs"].tolist() else: images = [] for idx, row in tqdm(df.iterrows(), total=len(df)) if df is not None else enumerate(image_files): if df is not None: fname = row["imgs"] if "imgs" in row else str(idx) # catch error if happen try: image_name = images[idx].split("/")[-1].split(".")[0] if images else "" record_img_path = next((key for key in image_files if image_name in key), None) except Exception as e: print(f"Error: {e} in row {idx}, using index as file name") print(images[idx]) record_img_path = None break # raise error if the path does not exist # check the path existance if not os.path.exists(image_files[record_img_path]) if record_img_path else None: raise FileNotFoundError(f"Image file not found: {image_files[record_img_path]}") yield idx, { "image": image_files[record_img_path] if record_img_path else None, "file_name": fname, "metadata": row.to_json(), } else: fname = Path(image_files[idx]).stem yield idx, { "image": image_files[idx], "file_name": fname, "metadata": None, }