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 training code, TensorFlow generates a trained model based on your training script and reads the checkpoint file in the specified path to complete model weight initialization or random initialization.
2. TensorFlow calls the GE's initialization API through TF Adapter to start devices, initialize compute engines, and initialize operator information libraries. Then, TF Adapter converts the trained model into an IR model, and compiles and runs the model.
3. After the compilation and optimization of graphs are complete, GE calls the Runtime API to allocate runtime resources, including the memory, streams, and events. Allocated resources are managed by Runtime.