TensorFlow Network Execution Process

The following figure shows the interaction process of each component during TensorFlow network execution.

Figure 1 Interaction process

1. After you execute the training code, the TensorFlow frontend generates a trained model based on your training script and reads the checkpoint file from the specified path to initialize model weights (or performs random initialization if no checkpoint is provided).
2. The frontend calls the GE initialization API through TF Adapter to open devices, initialize compute engines, and initialize operator information libraries. After that, it converts the trained model into an IR model and triggers graph compilation and execution.
3. Once graph compilation and optimization are complete, GE calls the Runtime API to allocate runtime resources such as memory, streams, and events. These resources are then managed by the Runtime component.