Understanding how TensorFlow executes operations is fundamental to writing efficient and scalable code. While TensorFlow 2 defaults to eager execution, running operations imperatively like Python, the creation and optimization of static computational graphs remain central for achieving high performance and enabling deployment. This chapter focuses on this interplay.

You will learn about:

• The differences and trade-offs between eager execution and graph execution.
• How tf.function uses AutoGraph to convert Python code into optimized TensorFlow graphs.
• The mechanics of function tracing and how TensorFlow represents computations internally.
• Implementing control flow (e.g., conditionals, loops) within graphs.
• Using tf.GradientTape for automatic differentiation, the engine behind gradient-based training like $L = \sum (y_{pred} - y_{true})^2$.
• Basic resource management considerations and debugging techniques for TensorFlow programs.

Mastering these execution details provides the foundation needed for the performance tuning, distributed training, and custom component development covered subsequently.