从 .pb 文件加载 TensorFlow 模型

Question

Background: I have my TensorFlow model stored in a directory with the folders/files: assets, variables, and saved_model.pb. All I need to do with this model is make predictions (no training) so I do not need the assets and variables folders. The saved_model.pb file is much smaller than the variables folder and storage is extremely important to me.

Question: How could I load my TensorFlow model from the saved_model.pb file only (no variables or assets folders) in Tensorflow 2.3?

This is my file structure:

saved_models
├── model_1
|   ├── assets
|   |   └── (this is an empty folder)
|   ├── saved_model.pb
|   ├── variables
|   |   ├── variables.data-00000-of-00001
|   |   └── variables.index

Answer 1

The saved_model.pb file only contains a description of the structure of the network. You need the variables as they contain the weights for each of the network layers, without them you would be doing prediction with a blank network.

Answer 2

You could try saving the model in .h5 format. This just generates a single file . I am not sure if it is smaller than the .pb size but give it a try and find out. Documentation is here.