Which is preferable from a computational graph perspective — dynamic or static?

The choice between a dynamic and static computation graph depends on the task and requirements for flexibility and performance.

• Static graph (e.g., TensorFlow 1.x) is built once before execution. This allows for optimizing computations, planning them in advance, and efficiently utilizing resources. Static graphs are well-suited for production environments and tasks with a fixed model structure.

• Dynamic graph (e.g., PyTorch, TensorFlow Eager) is built "on the fly" during code execution. This provides greater flexibility, ease of debugging, and the ability to use conditional operators and loops naturally.

The preference depends on the situation:

• For research and prototyping, a dynamic graph is often more convenient.
• For industrial use and optimization, a static graph is preferable.

Thus, there is no universal answer, but dynamic graphs offer flexibility, while static graphs provide performance and optimization.