python - Tensorflow `set_random_seed` not working

17
votes

Calling tf.set_random_seed(SEED) has no effect that I can tell...

For example, running the code below several times inside an IPython notebook produces different output each time:

import tensorflow as tf
tf.set_random_seed(42)
sess = tf.InteractiveSession()
a = tf.constant([1, 2, 3, 4, 5])
tf.initialize_all_variables().run()
a_shuf = tf.random_shuffle(a)
print(a.eval())
print(a_shuf.eval())
sess.close()

If I set the seed explicitly: a_shuf = tf.random_shuffle(a, seed=42), the output is the same after each run. But why do I need to set the seed if I already call tf.set_random_seed(42)?

The equivalent code using numpy just works:

import numpy as np
np.random.seed(42)
a = [1,2,3,4,5]
np.random.shuffle(a)
print(a)
1
pythontensorflow
Your code snippet only shows one evaluation of the random shuffle. How do you run it a second time to see a different result? (If I copy your snippet to a file seed.py and run python seed.py repeatedly, I do get the same results, and tf.set_random_seed(42) is working as intended.) - mrry
@mrry I have the code above inside a cell of an IPython notebook. Running the cell multiple times generates a different result, even though I would expect it to be the same... - ostrokach
It's also strange that if I run !python seed.py multiple times from inside the IPython notebook I get the same result, but if I run %run seed.py multiple times, I get a different result... - ostrokach
That's because you are appending ops to existing graph each time you do %run seed.py, try print [n.name for n in tf.get_default_graph().as_graph_def().node] to see what's in your graph - Yaroslav Bulatov
@STJ This question was asked two years earlier and has more upvotes. Why is it the duplicate and not the other way around? - ostrokach

1 Answers

13
votes

That only sets the graph-level random seed. If you execute this snippet several times in a row, the graph will change, and two shuffle statements will get different operation-level seeds. The details are described in the doc string for set_random_seed

To get deterministic a_shuf you can either

  1. Call tf.reset_default_graph() between invocations or
  2. Set operation-level seed for shuffle: a_shuf = tf.random_shuffle(a, seed=42)