转换Cifar10数据集

大家好，又见面了，我是你们的朋友全栈君。

Cifar10数据集不讲了吧，入门必备，下载地址: https://www.cs.toronto.edu/~kriz/cifar.html
官方提供三种形式的下载:

可以看出是不提供图片形式的下载的，需要进行数据转换，虽然可以直接读成ndarray，但是对于初学者可能读图更直观点

自己写了个转换程序(将bytes形式的文件转换为图片并分类存储):

def recover_cifar10(cifar10_dir):
    """Save cifar 10 data(only training data) to files. Args: cifar10_dir: cifar 10 dataset path(python version). Returns: """
    save_dir = './data/cifar10'

    def save_batch(path):
        with open(path, 'rb') as fo:
            batch_data = pickle.load(fo, encoding='bytes')

        X, Y, N = batch_data[b'data'], batch_data[b'labels'], batch_data[b'filenames']
        batch_size = X.shape[0]
        for x in range(0, batch_size):
            sample = X[x].reshape((3, 32, 32))
            r, g, b = sample[0], sample[1], sample[2]
            r0, g0, b0 = Image.fromarray(r), Image.fromarray(g), Image.fromarray(b)
            sample_rgb = Image.merge('RGB', (r0, g0, b0))
            sample_label = bytes.decode(label_names[Y[x]])  # image label
            sample_name = bytes.decode(N[x])

            cat_dir = join(save_dir, sample_label)
            if not os.path.exists(cat_dir):
                os.makedirs(cat_dir)

            sample_save_path = join(cat_dir, sample_name)
            sample_rgb.save(sample_save_path)

    with open(join(cifar10_dir, 'batches.meta'), 'rb') as fo:
        meta_data = pickle.load(fo, encoding='bytes')
    label_names = meta_data[b'label_names']
    for x in range(1, 6):
        path = join(cifar10_dir, 'data_batch_%s' % str(x))
        save_batch(path)

有需要的拿走，可以直接用