Source code for hfai.datasets.imagenet21k

from typing import Callable, Optional

import pickle
from PIL import Image
from io import BytesIO
from ffrecord import FileReader
from .base import (
    BaseDataset,
    get_data_dir,
    register_dataset
)


"""
Expected file organization:

    [data_dir]
        train.ffr
            PART_00000.ffr
            PART_00001.ffr
            ...
        val.ffr
            PART_00000.ffr
            PART_00001.ffr
            ...
"""


@register_dataset
class ImageNet21K(BaseDataset):
    """
    这是一个图像识别数据集

    更多信息参考：https://github.com/Alibaba-MIIL/ImageNet21K

    Args:
        split (str): 数据集划分形式，包括：完整集（``full``）、训练集（``train``）或者验证集（``val``）
        transform (Callable): transform 函数，对图片进行 transfrom，接受一张图片作为输入，输出 transform 之后的图片
        check_data (bool): 是否对每一条样本检验校验和（默认为 ``True``）
        miniset (bool): 是否使用 mini 集合（默认为 ``False``）

    Returns:
        pic, name, label (PIL.Image.Image, str, int): 返回的每个样本是一个元组，包括一个RGB格式的图片，图片名，以及代表这张图片类别的标签

    Examples:

    .. code-block:: python

        from hfai.datasets import ImageNet21K
        from torchvision import transforms

        transform = transforms.Compose([
            transforms.Resize(224),
            transforms.ToTensor(),
            transforms.Normalize(mean=mean, std=std),
        ])
        dataset = ImageNet21K(split, transform)
        loader = dataset.loader(batch_size=64, num_workers=4)

        for pic, name, label in loader:
            # training model

    """

    def __init__(
        self,
        split: str,
        transform: Optional[Callable] = None,
        check_data: bool = True,
        miniset: bool = False
    ) -> None:
        super(ImageNet21K, self).__init__()

        assert split in ["full", "train", "val"]
        self.split = split
        self.transform = transform
        data_dir = get_data_dir()
        if miniset:
            data_dir = data_dir / "mini"
        self.data_dir = data_dir / "ImageNet21K"
        self.fname = self.data_dir / f"{split}.ffr"
        self.reader = FileReader(self.fname, check_data)


    def __len__(self):
        return self.reader.n

    def __getitem__(self, indices):

        imgs_bytes = self.reader.read(indices)
        samples = []
        for i, bytes_ in enumerate(imgs_bytes):
            img_bytes, name, label = pickle.loads(bytes_)
            try:
                img = Image.open(BytesIO(img_bytes)).convert("RGB")
            except Exception as e:
                continue

            if self.transform:
                img = self.transform(img)
            samples.append((img, name, int(label)))

        return samples