【求助】本地加载fashion-mnist数据集，利用frombuffer读取时为什么需要offset=16与offset=8

由于使用tensorflow.keras时加载fashion-mnist数据集都要在线下载，因此选择加载本地已经下载好的.gz格式的数据集，如图所示

加载数据集的程序如下：
其中注释行不太理解，gzip解压后，使用numpy.frombuffer读入时，为什么需要设置读取样本时的offset=16，读取标签时的offset=8？

def get_data():
	x_train_path = r"F:/Data/fashion_minist/train-images-idx3-ubyte.gz"
	y_train_path = r"F:/Data/fashion_minist/train-labels-idx1-ubyte.gz"
	x_test_path = r"F:/Data/fashion_minist/t10k-images-idx3-ubyte.gz"
	y_test_path = r"F:/Data/fashion_minist/t10k-labels-idx1-ubyte.gz"
	with gzip.open(x_train_path, "rb") as data:
    	x_train = np.frombuffer(data.read(), np.uint8, offset=16).reshape(-1, 28, 28)  # 有疑问处！
	with gzip.open(y_train_path, "rb") as data:
    	y_train = np.frombuffer(data.read(), np.uint8, offset=8)  # 有疑问处！
	with gzip.open(x_test_path, "rb") as data:
    	x_test = np.frombuffer(data.read(), np.uint8, offset=16).reshape(-1, 28, 28)  # 有疑问处！
	with gzip.open(y_test_path, "rb") as data:
    	y_test = np.frombuffer(data.read(), np.uint8, offset=8)  # 有疑问处！    
	return (x_train, y_train), (x_test, y_test)