Stable results with TensorFlow

Question

What would be a good way to ensure TensorFlow results persist across different runs? I can't find much information on e.g. seeding random weight initialization, so my results tend to differ with every run.

Answer 1

The API method tf.set_random_seed() can be used to set a random seed that will be used in all TensorFlow random operations (including the usual random weight initializers and tf.RandomShuffleQueue).