How to use ModelCheckpoint in Keras

Long Keras training runs need a checkpoint file before the final epoch finishes. ModelCheckpoint saves a model or weight file during model.fit(), which protects progress and preserves the best validation result from a run that may take minutes, hours, or longer.

The callback watches a metric from the training logs, such as val_loss or val_accuracy. With save_best_only=True, it overwrites the target checkpoint only when the monitored metric improves, so the path keeps the strongest model seen so far instead of the last epoch by default.

A whole-model checkpoint in native .keras format stores the architecture, weights, compile information, and optimizer state that Keras can reload with keras.models.load_model(). Use the weights-only mode only when the architecture will be recreated separately and the checkpoint filename ends in .weights.h5.

1. Create checkpoint_best_model.py with a ModelCheckpoint callback.

   checkpoint_best_model.py

   import keras
   import numpy as np
    
    
   keras.utils.set_random_seed(42)
    
   x_train = np.linspace(-1.0, 1.0, 160, dtype="float32").reshape(80, 2)
   y_train = (0.75 * x_train[:, :1]) - (0.25 * x_train[:, 1:2])
   x_val = np.linspace(-0.8, 0.8, 40, dtype="float32").reshape(20, 2)
   y_val = (0.75 * x_val[:, :1]) - (0.25 * x_val[:, 1:2])
    
   model = keras.Sequential(
       [
           keras.layers.Input(shape=(2,)),
           keras.layers.Dense(8, activation="relu"),
           keras.layers.Dense(1),
       ]
   )
   model.compile(
       optimizer=keras.optimizers.Adam(learning_rate=0.05),
       loss="mse",
       metrics=["mae"],
   )
    
   checkpoint_path = "checkpoints/best.keras"
   checkpoint = keras.callbacks.ModelCheckpoint(
       filepath=checkpoint_path,
       monitor="val_loss",
       mode="min",
       save_best_only=True,
       verbose=1,
   )
    
   history = model.fit(
       x_train,
       y_train,
       validation_data=(x_val, y_val),
       epochs=4,
       batch_size=16,
       callbacks=[checkpoint],
       verbose=0,
   )
    
   loaded = keras.models.load_model(checkpoint_path)
   loaded_loss, loaded_mae = loaded.evaluate(x_val, y_val, verbose=0)
    
   print(f"Saved checkpoint: {checkpoint_path}")
   print(f"Best val_loss: {min(history.history['val_loss']):.4f}")
   print(f"Loaded val_loss: {loaded_loss:.4f}")
   print(f"Loaded val_mae: {loaded_mae:.4f}")

   Use a .keras suffix for a full-model checkpoint. Use .weights.h5 only with save_weights_only=True, and compile the model before loading weights when optimizer state needs to be restored.

2. Run the training script in the Python environment where Keras is installed.

   $ python checkpoint_best_model.py
   Epoch 1: val_loss improved from None to 0.05945, saving model to checkpoints/best.keras
   ##### snipped #####
   Saved checkpoint: checkpoints/best.keras
   Best val_loss: 0.0006
   Loaded val_loss: 0.0006
   Loaded val_mae: 0.0218

   Loaded val_loss appears only after keras.models.load_model() opens the saved .keras checkpoint and evaluates it on the validation data.

3. Remove the demo script and checkpoint directory when the run was only a callback test.

   $ rm -rf checkpoint_best_model.py checkpoints

   Do not remove a real training checkpoint directory until the checkpoint has been copied, promoted, or replaced by a newer saved model.