physicsgen.py

import os
import datasets
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

# Define configurations for each flavor.
BUILDER_CONFIGS = [
    datasets.BuilderConfig(
        name="sound_baseline",
        description="Physical dataset: baseline variant",
        data_dir="sound_baseline"
    ),
    datasets.BuilderConfig(
        name="sound_reflection",
        description="Physical dataset: reflection variant",
        data_dir="sound_reflection"
    ),
    datasets.BuilderConfig(
        name="sound_diffraction",
        description="Physical dataset: diffraction variant",
        data_dir="sound_diffraction"
    ),
    datasets.BuilderConfig(
        name="sound_combined",
        description="Physical dataset: combined variant",
        data_dir="sound_combined"
    ),
    datasets.BuilderConfig(
        name="lens_p1",
        description="Distortion dataset variant",
        data_dir="lens_p1"
    ),
    datasets.BuilderConfig(
        name="lens_p2",
        description="Distortion dataset variant",
        data_dir="lens_p2"
    ),
    datasets.BuilderConfig(
        name="ball_roll",
        description="Double image dataset variant",
        data_dir="ball_roll"
    ),
    datasets.BuilderConfig(
        name="ball_bounce",
        description="Double image dataset variant",
        data_dir="ball_bounce"
    ),
]

class MyPhysicalDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = BUILDER_CONFIGS
    VERSION = datasets.Version("1.1.0")

    def _info(self):
        if self.config.name in ["sound_baseline", "sound_reflection", "sound_diffraction", "sound_combined"]:
            features = datasets.Features({
                "lat": datasets.Value("float"),
                "long": datasets.Value("float"),
                "db": datasets.Value("string"),
                "soundmap": datasets.Image(),    # Expects a dict: {"bytes": ...}
                "osm": datasets.Image(),
                "temperature": datasets.Value("int32"),
                "humidity": datasets.Value("int32"),
                "yaw": datasets.Value("float"),
                "sample_id": datasets.Value("int32"),
                "soundmap_512": datasets.Image(),
            })
        elif self.config.name in ["lens_p1", "lens_p2"]:
            features = datasets.Features({
                "label_path": datasets.Value("string"),
                "fx": datasets.Value("float"),
                "k1": datasets.Value("float"),
                "k2": datasets.Value("float"),
                "k3": datasets.Value("float"),
                "p1": datasets.Value("float"),
                "p2": datasets.Value("float"),
                "cx": datasets.Value("float"),
            })
        elif self.config.name in ["ball_roll", "ball_bounce"]:
            features = datasets.Features({
                "ImgName": datasets.Value("string"),
                "StartHeight": datasets.Value("int32"),
                "GroundIncli": datasets.Value("float"),
                "InputTime": datasets.Value("int32"),
                "TargetTime": datasets.Value("int32"),
                "input_image": datasets.Image(),   # Expects {"bytes": ...}
                "target_image": datasets.Image(),
            })
        else:
            raise ValueError(f"Unknown config name: {self.config.name}")
        return datasets.DatasetInfo(
            description="Multiple variant physical tasks dataset stored as parquet files.",
            features=features,
        )

    def _split_generators(self, dl_manager):
        # Use hf_hub_download to fetch the parquet files directly from the Hub.
        repo_id = "mspitzna/physicsgen" 
        train_file = hf_hub_download(repo_id=repo_id, filename=f"{self.config.data_dir}/train.parquet", repo_type="dataset")
        test_file = hf_hub_download(repo_id=repo_id, filename=f"{self.config.data_dir}/test.parquet", repo_type="dataset")
        eval_file = hf_hub_download(repo_id=repo_id, filename=f"{self.config.data_dir}/eval.parquet", repo_type="dataset")
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"parquet_file": train_file},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"parquet_file": test_file},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"parquet_file": eval_file},
            ),
        ]

    def _generate_examples(self, parquet_file):
        table = pq.read_table(parquet_file)
        examples = table.to_pylist()

        # Wrap image bytes into the format expected by datasets.Image.
        if self.config.name in ["sound_baseline", "sound_reflection", "sound_diffraction", "sound_combined"]:
            for example in examples:
                for key in ["soundmap", "osm", "soundmap_512"]:
                    if example.get(key) is not None and isinstance(example[key], bytes):
                        example[key] = {"bytes": example[key]}
        elif self.config.name in ["ball_roll", "ball_bounce"]:
            for example in examples:
                for key in ["input_image", "target_image"]:
                    if example.get(key) is not None and isinstance(example[key], bytes):
                        example[key] = {"bytes": example[key]}
        for idx, row in enumerate(examples):
            yield idx, row