dataset.py

import os
import numpy as np
import datasets

class MedMNIST(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            description="OrganCMNIST dataset from MedMNIST collection",
            features=datasets.Features(
                {
                    "image": datasets.Array3D(dtype="uint8", shape=(28, 28, 3)),
                    "label": datasets.ClassLabel(num_classes=11),
                }
            ),
        )

    def _split_generators(self, dl_manager):
        data_dir = dl_manager.manual_dir
        npz_path = os.path.join(data_dir, "organcmnist.npz")
        return [
            datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"npz_path": npz_path, "split": "train"}),
            datasets.SplitGenerator(name=datasets.Split.VALIDATION, gen_kwargs={"npz_path": npz_path, "split": "val"}),
            datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"npz_path": npz_path, "split": "test"}),
        ]

    def _generate_examples(self, npz_path, split):
        data = np.load(npz_path)
        images = data[f"{split}_images.npy"]
        labels = data[f"{split}_labels.npy"].squeeze()

        for idx in range(len(images)):
            yield idx, {"image": images[idx], "label": int(labels[idx])}