How to save and load a Keras model

A trained Keras model often needs to move from a notebook or training script into another Python process before it can be scored, evaluated, or shared with a teammate. Saving the complete model to the native .keras format keeps the architecture and learned weights together instead of relying on a separate model-building script.

The native Keras file is the standard whole-model artifact for Keras 3. It can store the model configuration, layer weights, and compile information, including optimizer state when the optimizer has been built during training.

Use native .keras saving when the next consumer is another Keras process that should reload the same model for prediction, evaluation, or fine-tuning. Use a SavedModel export when the target is TensorFlow serving tooling, and handle custom-layer registration separately when the model contains custom Python objects.

1. Install Keras and the backend package used by the model.

   $ python -m pip install --upgrade keras jax

   The sample script uses JAX. Use the same backend package and runtime that the saved model will run with.

2. Select the backend before importing Keras.

   $ export KERAS_BACKEND=jax

   Keras reads KERAS_BACKEND during import, so set it before opening a Python shell, notebook kernel, or script process.
   Related: How to set the Keras backend

3. Create the save and load test script.

   save_load_model.py

   import os
   from pathlib import Path
    
   os.environ.setdefault("KERAS_BACKEND", "jax")
    
   import keras
   import numpy as np
    
    
   keras.utils.set_random_seed(42)
    
   model_path = Path("credit-risk-score.keras")
   if model_path.exists():
       model_path.unlink()
    
   x_train = np.linspace(0.05, 0.95, 32, dtype="float32").reshape(8, 4)
   y_train = ((x_train[:, 0] + x_train[:, 1]) > 0.9).astype("float32").reshape(-1, 1)
    
   model = keras.Sequential(
       [
           keras.Input(shape=(4,), name="features"),
           keras.layers.Dense(6, activation="relu", name="hidden"),
           keras.layers.Dense(1, activation="sigmoid", name="risk_score"),
       ]
   )
   model.compile(
       optimizer=keras.optimizers.Adam(learning_rate=0.05),
       loss=keras.losses.BinaryCrossentropy(),
       metrics=[keras.metrics.BinaryAccuracy(name="accuracy")],
   )
   model.fit(x_train, y_train, epochs=8, batch_size=4, verbose=0)
    
   sample = np.array([[0.25, 0.60, 0.35, 0.45]], dtype="float32")
   before_save = model.predict(sample, verbose=0)
    
   model.save(model_path)
   loaded_model = keras.models.load_model(model_path)
   after_load = loaded_model.predict(sample, verbose=0)
   loaded_loss, loaded_accuracy = loaded_model.evaluate(x_train, y_train, verbose=0)
    
   print(f"Backend: {keras.backend.backend()}")
   print(f"Saved model: {model_path}")
   print(f"Saved file exists: {model_path.exists()}")
   print(f"Loaded model type: {loaded_model.__class__.__name__}")
   print(f"Loaded optimizer: {loaded_model.optimizer.__class__.__name__}")
   print(f"Reloaded accuracy: {loaded_accuracy:.4f}")
   print(f"Prediction before save: {before_save[0, 0]:.6f}")
   print(f"Prediction after load: {after_load[0, 0]:.6f}")
   print(f"Predictions match: {np.allclose(before_save, after_load, atol=1e-6)}")

   Replace the demo training data and model definition with the trained model that should become the portable .keras artifact.

4. Run the script.

   $ python save_load_model.py
   Backend: jax
   Saved model: credit-risk-score.keras
   Saved file exists: True
   Loaded model type: Sequential
   Loaded optimizer: Adam
   Reloaded accuracy: 0.6250
   Prediction before save: 0.595262
   Prediction after load: 0.595262
   Predictions match: True

   Predictions match: True confirms that keras.models.load_model() restored a model that returns the same prediction for the sample input.

5. Remove the demo script and saved model if the run was only a save/load test.

   $ rm -f save_load_model.py credit-risk-score.keras

   Do not delete a production .keras artifact until it has been copied to the intended model registry, release directory, or deployment handoff location.