Spatial457.py

import os
import json
import datasets


class Spatial457(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(name="L1_single"),
        datasets.BuilderConfig(name="L2_objects"),
        datasets.BuilderConfig(name="L3_2d_spatial"),
        datasets.BuilderConfig(name="L4_occ"),
        datasets.BuilderConfig(name="L4_pose"),
        datasets.BuilderConfig(name="L5_6d_spatial"),
        datasets.BuilderConfig(name="L5_collision"),
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description="Spatial457: A multi-task spatial visual question answering dataset.",
            features=datasets.Features({
                "image": datasets.Image(),  # 自动加载图片
                "image_filename": datasets.Value("string"),
                "question": datasets.Value("string"),
                "answer": datasets.Value("string"),  # 会自动处理 True / False / str
                "question_index": datasets.Value("int32"),
                "program": datasets.Sequence(
                    {
                        "type": datasets.Value("string"),
                        "inputs": datasets.Sequence(datasets.Value("int32")),
                        "_output": datasets.Value("string"),
                        "value_inputs": datasets.Sequence(datasets.Value("string")),
                    }
                ),
            }),
            supervised_keys=None,
        )

    def _split_generators(self, dl_manager):
        
        base_url = "https://huggingface.co/datasets/RyanWW/Spatial457/resolve/main"
    
        json_url = f"{base_url}/questions/{self.config.name}.json"
        # 下载 json 文件
        task_json = dl_manager.download(json_url)
        
        # 读取 JSON 获取所有图片文件名
        with open(task_json, "r", encoding="utf-8") as f:
            all_data = json.load(f)["questions"]
        
        # 构建所有图片的 URL
        image_urls = {
            q["image_filename"]: f"{base_url}/images/{q['image_filename']}"
            for q in all_data
        }
        
        # 批量下载所有图片
        downloaded_images = dl_manager.download(image_urls)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"json_file": task_json, "downloaded_images": downloaded_images}
            )
        ]

    def _generate_examples(self, json_file, downloaded_images):
        with open(json_file, "r", encoding="utf-8") as f:
            all_data = json.load(f)["questions"]

        for idx, q in enumerate(all_data):
            img_filename = q["image_filename"]
            img_path = downloaded_images[img_filename]
            yield idx, {
                "image": img_path,
                "image_filename": img_filename,
                "question": q["question"],
                "answer": str(q["answer"]),
                "question_index": q["question_index"],
                "program": q["program"],
            }