Scenario Description

You need to define placeholders, variables, and operators to form a computational graph. Then, use the new session instance to start the model running. The session instance executes the graph in a distributed manner, inputs data, updates variables based on the optimization algorithm, and returns the execution result (tensor instance). Model training using tf.Session is more customizable. You can modify the mode based on the actual training method.

Training Process

Figure 1 tf.Session training flowchart