Shuffling || Global or operational seed

[tarmas99]

Hello Sir,
in lecture 17, you described about global and operational seeds in connection with Shuffling the Tensors.
You shown that while we apply both the seeds together we get the same output as input. But when I tried that today, this happened.

not_shuffled = tf.constant([[1,2,3],
                            [4,5,6],
                            [7,8,9]])
not_shuffled.ndim
print("[INFO]--before shuffle")
print(not_shuffled)

print()
print("[INFO] after shuffling")
print(tf.random.shuffle(not_shuffled))

# Without any seed
for i in range(5):
    print()
    print("[i = " + str(i) + "] -- ")
    print(tf.random.shuffle(not_shuffled))    

Output:

# Operational Level seed
for i in range(5):
    print()
    print("[i = " + str(i) + "] -- ")
    print(tf.random.shuffle(not_shuffled , seed=42))

Output:

# Global level seed
tf.random.set_seed(42)
for i in range(5):
    print("[i = " + str(i) + "] -- ")
    print(tf.random.shuffle(not_shuffled))

Output:

# Global level seed and Operational Level seed
tf.random.set_seed(42)
for i in range(5):
    print("[i = " + str(i) + "] -- ")
    print(tf.random.shuffle(not_shuffled , seed = 42))

Output:

So, in none of the cases the output was same as expected.

[mrdbourke]

Hey Samrat,

Looks like the operation seed is getting reset every time you call the for loop.

I'd try running the different combinations of the seed you've run without the for loop.

For example:

# Cell 1
import tensorflow as tf
tf.random.set_seed(42)
A = tf.reshape(tf.range(9)+1, [3, 3])
tf.random.shuffle(A)

>>> <tf.Tensor: shape=(3, 3), dtype=int32, numpy=
array([[4, 5, 6],
       [7, 8, 9],
       [1, 2, 3]], dtype=int32)>

# Cell 2
import tensorflow as tf
tf.random.set_seed(42)
A = tf.reshape(tf.range(9)+1, [3, 3])
tf.random.shuffle(A)

<tf.Tensor: shape=(3, 3), dtype=int32, numpy=
array([[4, 5, 6],
       [7, 8, 9],
       [1, 2, 3]], dtype=int32)>

You should notice the same outputs (because the global seed has been set), give it a try with the operational level seed too.

It can be quite confusing at first but with some practice, you'll get used to it.

For further practice, I'd also try reproducing the code in the tf.random.set_seed guide: https://www.tensorflow.org/api_docs/python/tf/random/set_seed