How to set a random seed in TensorFlow

Random choices in TensorFlow affect layer initialization, shuffled input batches, and sampled data augmentation. Setting a seed near the start of the program keeps those choices on a repeatable pseudorandom path during debugging, smoke tests, and small comparison runs.

The direct seed entry point for mixed TensorFlow and Keras scripts is tf.keras.utils.set_random_seed(). It sets Python's random seed, NumPy's global random seed, and TensorFlow's global random seed with one integer, which matches training code that also uses tf.random, tf.data, and ordinary NumPy calls.

A seed does not make every TensorFlow operation deterministic by itself. Separate generators such as np.random.default_rng() still need their own seed, and identical outputs on the same hardware and TensorFlow stack may need deterministic TensorFlow operations in addition to a fixed seed.

1. Open a terminal in a Python environment where TensorFlow imports successfully.

   $ python3 -c "import tensorflow as tf; print(tf.__version__)"
   2.21.0

   Related: How to create a virtual environment for TensorFlow
   Related: How to install TensorFlow with pip

2. Save the seed check script as

   seed-demo.py

   .

   seed-demo.py

   import random
    
   import numpy as np
   import tensorflow as tf
    
   SEED = 7
    
   tf.keras.utils.set_random_seed(SEED)
   generator = np.random.default_rng(SEED)
    
   print("Python random:", random.randint(0, 999))
   print("NumPy global:", int(np.random.randint(0, 1000)))
   print("NumPy Generator:", int(generator.integers(0, 1000)))
   print(
       "TensorFlow:",
       tf.random.uniform((3,), minval=0, maxval=10, dtype=tf.int32).numpy().tolist(),
   )

   tf.keras.utils.set_random_seed() does not seed a newly created np.random.default_rng() object, so pass SEED when the program creates one.

3. Run the script once to capture the seeded values.

   $ python3 seed-demo.py
   Python random: 331
   NumPy global: 175
   NumPy Generator: 944
   TensorFlow: [4, 5, 7]

4. Run the same script again from a fresh Python process.

   $ python3 seed-demo.py
   Python random: 331
   NumPy global: 175
   NumPy Generator: 944
   TensorFlow: [4, 5, 7]

   Matching values mean Python, NumPy's global generator, the explicitly seeded NumPy generator, and TensorFlow all started from the same seed state again.

5. Place the seed call before random-dependent TensorFlow work in the real program.

   SEED = 7
    
   tf.keras.utils.set_random_seed(SEED)
    
   train_dataset = train_dataset.shuffle(
       1000,
       seed=SEED,
       reshuffle_each_iteration=False,
   )
    
   model = build_model()
   model.fit(train_dataset, epochs=5)

   Call the seed function before layer creation, dataset shuffling, augmentation, or training. Calling it later does not rewind random work that already happened.

6. Remove the temporary seed check script after copying the seed pattern into the real program.

   $ rm seed-demo.py