Tensorflow's random.truncated_normal returns different results with the same seed
Question:
The following lines are supposed to get the same result:
print (tf.random.truncated_normal(shape=[2],seed=1234))
print (tf.random.truncated_normal(shape=[2],seed=1234))
But I got:
tf.Tensor([-0.12297685 -0.76935077], shape=(2,), dtype=float32)
tf.Tensor([0.37034193 1.3367208 ], shape=(2,), dtype=float32)
Why?
Answers:
This seems to be intentional, see the docs here. Specifically the "Examples" section.
What you need is stateless_truncated_normal:
print(tf.random.stateless_truncated_normal(shape=[2],seed=[1234, 1]))
print(tf.random.stateless_truncated_normal(shape=[2],seed=[1234, 1]))
Gives me
tf.Tensor([1.0721238 0.10303579], shape=(2,), dtype=float32)
tf.Tensor([1.0721238 0.10303579], shape=(2,), dtype=float32)
Note: The seed needs to be two numbers here, I honestly don’t know why (the docs don’t say).
Tensorflow has two types of seeds the global and the operational – this is also why you need to pass two numbers to stateless_truncated_normal as xdurch0 describes in his answer. Tensorflow combines these two seeds to generate a new one.
tf.random.truncated_normal(shape=[2],seed=1234) # global seed #1 & operational 1234 -> Seed A
tf.random.truncated_normal(shape=[2],seed=1234) # global seed #2 & operational 1234 -> Seed B
There are multiple ways to tackle your problem. Set the global seed as well beforehand twice.
Work inside @tf.functions these kindof reset the global seed and have their own operational counters.
Or use stateless_truncated_normal as written in the other answer.
As already linked, in the documentation it is described as well.
The following lines are supposed to get the same result:
print (tf.random.truncated_normal(shape=[2],seed=1234))
print (tf.random.truncated_normal(shape=[2],seed=1234))
But I got:
tf.Tensor([-0.12297685 -0.76935077], shape=(2,), dtype=float32)
tf.Tensor([0.37034193 1.3367208 ], shape=(2,), dtype=float32)
Why?
This seems to be intentional, see the docs here. Specifically the "Examples" section.
What you need is stateless_truncated_normal:
print(tf.random.stateless_truncated_normal(shape=[2],seed=[1234, 1]))
print(tf.random.stateless_truncated_normal(shape=[2],seed=[1234, 1]))
Gives me
tf.Tensor([1.0721238 0.10303579], shape=(2,), dtype=float32)
tf.Tensor([1.0721238 0.10303579], shape=(2,), dtype=float32)
Note: The seed needs to be two numbers here, I honestly don’t know why (the docs don’t say).
Tensorflow has two types of seeds the global and the operational – this is also why you need to pass two numbers to stateless_truncated_normal as xdurch0 describes in his answer. Tensorflow combines these two seeds to generate a new one.
tf.random.truncated_normal(shape=[2],seed=1234) # global seed #1 & operational 1234 -> Seed A
tf.random.truncated_normal(shape=[2],seed=1234) # global seed #2 & operational 1234 -> Seed B
There are multiple ways to tackle your problem. Set the global seed as well beforehand twice.
Work inside @tf.functions these kindof reset the global seed and have their own operational counters.
Or use stateless_truncated_normal as written in the other answer.
As already linked, in the documentation it is described as well.