How to set random seed in Keras

Repeatable Keras experiments need the same random state before model initialization, data shuffling, augmentation, and backend operations run. keras.utils.set_random_seed() sets Python, NumPy, and backend framework seeds from one call.

Call the seed utility before creating layers, initializers, datasets, or random tensors. Reusing the same seed at the start of the same code path should reproduce initial weights and predictions for deterministic operations.

Random seeds do not guarantee full hardware determinism in every distributed, accelerator, or cuDNN-backed path. Treat this as the starting point for repeatability, then add backend-specific determinism settings only when the target runtime needs them and supports them.

1. Install Keras and a backend.

   $ python -m pip install --upgrade keras jax

2. Create set_random_seed.py.

   set_random_seed.py

   import os
    
   os.environ["KERAS_BACKEND"] = "jax"
    
   import keras
   import numpy as np
    
    
   def seeded_prediction(seed):
       keras.utils.set_random_seed(seed)
       model = keras.Sequential(
           [
               keras.Input(shape=(3,)),
               keras.layers.Dense(5, activation="relu"),
               keras.layers.Dense(1, activation="sigmoid"),
           ]
       )
       sample = np.array([[0.25, 0.50, 0.75]], dtype="float32")
       return model.predict(sample, verbose=0), model.get_weights()[0].copy()
    
    
   prediction_a, kernel_a = seeded_prediction(123)
   prediction_b, kernel_b = seeded_prediction(123)
   prediction_c, kernel_c = seeded_prediction(124)
    
   print(f"backend: {keras.backend.backend()}")
   print("seed used for first two runs: 123")
   print(f"matching initial kernel: {np.allclose(kernel_a, kernel_b)}")
   print(f"matching prediction: {np.allclose(prediction_a, prediction_b)}")
   print(f"different seed changes kernel: {not np.allclose(kernel_a, kernel_c)}")
   print(f"prediction with seed 123: {prediction_a[0, 0]:.6f}")
   print(f"prediction with seed 124: {prediction_c[0, 0]:.6f}")

   The proof compares both initial weights and a prediction result because layer initializers are the common source of surprising non-repeatability.

3. Run the script.

   $ python set_random_seed.py
   backend: jax
   seed used for first two runs: 123
   matching initial kernel: True
   matching prediction: True
   different seed changes kernel: True
   prediction with seed 123: 0.683945
   prediction with seed 124: 0.357631

4. Confirm that the repeated seed produced matching output.

   matching initial kernel: True and matching prediction: True prove that the same seeded code path reproduced the initializer and prediction result.

5. Put the seed call near the top of real training scripts.

   import keras
    
   keras.utils.set_random_seed(123)

   Set the seed before creating models, loading randomized datasets, or constructing augmentation layers.