[tensorflow] reproduce in tensorflow: tf.set_random_seed() Tensorflow中复现实验

原创于 2018-11-14 16:14:32 发布 · 318 阅读

0 ·

CC 4.0 BY-SA版权

坑专栏收录该内容

18 篇文章

订阅专栏

本文深入探讨了TensorFlow中两种随机种子的使用：图级种子和操作级种子。图级种子优先于操作级种子，确保不同会话运行相同图结构时产生一致的随机行为。而操作级种子则控制特定操作的随机一致性。当在内部函数或lambda函数中使用随机函数时，需手动设置操作级种子，因图级种子不会被继承。

Reproducing experiments means to restore the random seed.
There are two types of seeding in tensorflow:

graph-level
typically by calling tf.set_random_seed(seed)
op-level
by setting the seed= param in the tf.random functions.

What are the differences?

graph-level seed has priority over op-level
when graph-level seed set, different sessions running identical graphs with the same graph-level seed will produce the same random behavior.
Note: graphs do not need to have same references but same structure and seed, i.e., set by tf.set_random_seed(seed)
op-level seed controls the specific op to keep its random behavior identical across different sessions.
Note: graphs do not need to have same references and even the same structure, but there should be two identical op with the same seed.

Warning
When tf.random functions are used in interior functions or lambda functions, It seems that the graph is not inherited, which means the graph-level seed is not inherited. It follows that in such cases, we have to manually set the op-level seed, if we want to reproduce the experiment.