I. Dynamic Graph vs Static Graph (动态图 vs 静态图)#

1) Definition (定义)#

A Dynamic Computation Graph (动态图计算图) is built during runtime (运行时构建).

👉 The graph changes as the program executes (执行时动态变化).

2) Characteristics (特点)#

• Defined on-the-fly (即时定义)
• Flexible control flow (灵活控制流)
• Easy debugging (易调试)

3) Example (示例)#

1
import torch
2

3
x = torch.tensor(2.0, requires_grad=True)
4
y = x * x + 3
5

6
y.backward()
7
print(x.grad)  # dy/dx = 2x = 4

👉 The graph is created dynamically when y is computed.

1) Definition (定义)#

A Static Computation Graph (静态计算图) is defined before execution (运行前定义).

👉 The graph structure does not change during runtime (运行时不变).

2) Characteristics (特点)#

• Predefined graph (预先定义)
• Optimized before execution (执行前优化)
• Better performance (更高性能)

3) Conceptual Example (概念示例)#

1
# Pseudo-code (伪代码)
2
x = placeholder()
3
y = x * x + 3
4

5
# Build graph first (先构建图)
6
graph = build_graph(x, y)
7

8
# Execute later (再执行)
9
run(graph, feed_dict={x: 2})

1) Dynamic Graph#

• Research (科研)
• Prototyping (快速实验)

2) Static Graph#

• Production systems (生产环境)
• Performance-critical tasks (高性能场景)