If you would like to process CIFAR dataset like the ImageNet-1K, i.e. split the train and test dataset and every category as an individual subdirectory, try this gist!

cifar_pre_process.py

import cv2
import numpy as np
import pickle
import os
import shutil
from torchvision import datasets


# %%
def unpickle(file):
    with open(file, 'rb') as fo:
        dict = pickle.load(fo, encoding='bytes')
    return dict

def cifar10():

    _ = datasets.CIFAR10(root='./data', train=True, download=True)
    root = os.getcwd()
    os.chdir("./data/cifar-10-batches-py")


    def process(filenames, folder_name):
        os.makedirs(folder_name, exist_ok=True)

        # make directory for each class
        meta = unpickle('batches.meta')
        label_names = meta[b'label_names']
        for label in label_names:
            os.makedirs(folder_name + "/" + label.decode('utf-8'), exist_ok=True)

        for index, batch in enumerate(filenames):
            data = unpickle(batch)
            images = data[b'data']
            size = images.shape[0]
            images = images.reshape(size, 3, 32, 32).transpose(0,2,3,1).astype("uint8")

            label = data[b'labels']
                
            # resize dataset to 224x224
            images224 = np.zeros((size, 224, 224, 3), dtype=np.uint8)
            for i in range(size):
                images224[i] = cv2.cvtColor((cv2.resize(images[i], (224, 224), interpolation=cv2.INTER_CUBIC)), cv2.COLOR_BGR2RGB)  
            
            # save images to each directory
            for i in range(size):
                cv2.imwrite(folder_name + "/" + label_names[label[i]].decode('utf-8') + '/batch_' + str(index) + "_" + str(i) + '.jpg', images224[i])  


    process(['data_batch_1', 'data_batch_2', 'data_batch_3', 'data_batch_4', 'data_batch_5'], "train")
    process(['test_batch'], "test")
    os.chdir(root)

# %%
def cifar100():

    ## download training dataset
    _ = datasets.CIFAR100(root='./data', train=False, download=True)
    root = os.getcwd()
    os.chdir("./data/cifar-100-python")
    os.rename("test", "test.pkl")
    os.rename("train", "train.pkl")
    os.rename("meta", "meta.pkl")

    def process(pkl_file, folder_name):
        images = pkl_file[b'data']

        size = images.shape[0]

        # %%
        images = images.reshape(size, 3, 32, 32).transpose(0,2,3,1).astype("uint8")

        # %%
        # resize dataset to 224x224
        images224 = np.zeros((size, 224, 224, 3), dtype=np.uint8)
        for i in range(size):
            images224[i] = cv2.cvtColor((cv2.resize(images[i], (224, 224), interpolation=cv2.INTER_CUBIC)), cv2.COLOR_BGR2RGB)

        os.makedirs(folder_name, exist_ok=True)

        # save dataset to jpeg files
        for i in range(size):
            cv2.imwrite(folder_name + '/' + str(i) + '.jpg', images224[i])

        # %%
        # save the 224x224 images into each class folder

        # create a folder for each class
        labels = pkl_file[b'fine_labels']
        class_names = unpickle("meta.pkl")[b'fine_label_names']
        for i in range(100):
            # os.makedirs("train/" + class_names[i].decode("utf-8"))
            os.makedirs(folder_name + "/" + class_names[i].decode("utf-8"), exist_ok=True)

        for i in range(size):
            shutil.move(folder_name + "/" + str(i) + ".jpg", folder_name + "/" + class_names[labels[i]].decode("utf-8") + "/" + str(i) + ".png")
    
    process(unpickle("train.pkl"), "train")
    process(unpickle("test.pkl"), "test")
    os.chdir(root)

# %%
if __name__ == "__main__":
    cifar10()
    cifar100()

The final directory tree looks like this: