Tensorflow 基于TF2KerasAPI和TF数据集的图像特征提取

我应该如何使用tf数据集来运行model.predict（数据）并访问tf数据集的其他功能

例如：我的tf数据集具有以下格式： （张量，张量）->将图像预处理为tf.float32，将图像的UUID预处理为tf.string

如果我像这样提取特征向量：

for image_data, uuids in ds.batch(100):
    features = model.predict(data[0]) -> I get an array of features.

    At this moment features is an array of (100, 2048) and uuids is a tensor of (100,) tf.string
    How can I combine them in order to write the feature vectors to disk?
    From my understanding, I need to have both of them in the same format, either both tensors so I can continue using tf code and save the feature vector as a tfrecord, either to get the uuid as a string from the uuid tensor so I can use python code and save the array in the file using numpy.tofile.

    So my questions are:
    - How can I make the features to be a tensor?
    - Or can I get the string value from the tensor uuid?
    - Does anything sounds wrong in what I try to do? Is there a more optimal way to create the input pipeline? Or did I misunderstood the usage of Keras API and tf dataset? 

如果我使用python管道，我可以成功地将数组保存在文件中。但我想使用tf数据集，因为它的并行映射功能，批处理和自动调整并行调用，我认为它会更快、更优化