Absolutely! You’re looking for Keras' ModelCheckpoint callback—it’s tailor-made to solve exactly this problem: saving the optimal version of your model as it trains, even when your validation loss has those annoying peaks and valleys.

Here’s how to implement it step by step:

Step 1: Import the Required Callbacks #

First, make sure you import both EarlyStopping (which you’re already using) and ModelCheckpoint:

from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.models import Sequential
from keras.layers import Dense

Step 2: Configure Your Callbacks #

Define your EarlyStopping monitor as before, then set up ModelCheckpoint to track your validation loss and save only the best model:

# Define early stopping (you likely already have this)
early_stopping_monitor = EarlyStopping(patience=5, monitor='val_loss')

# Configure ModelCheckpoint to save the best model
model_checkpoint = ModelCheckpoint(
    filepath='best_model.h5',  # Path where the best model will be saved
    monitor='val_loss',        # Metric to monitor (use val_accuracy if you care more about accuracy)
    save_best_only=True,       # Only save the model when the monitored metric improves
    mode='min',                # 'min' because we want to minimize val_loss; use 'max' for accuracy
    verbose=1                  # Print a message when a better model is saved
)

Step 3: Update Your Training Code #

Fix the small typo in your original code (model_2.compile should be model.compile) and add both callbacks to the fit method:

input_shape = # Your input shape here
X, y = # Your training data here
X_test = # Your test data here

model = Sequential()
model.add(Dense(100, activation='relu', input_shape=input_shape))
model.add(Dense(1))

# Compile the model (fixed the typo here)
model.compile(optimizer='adam', loss='mean_squared_error', metrics=['accuracy'])

# Train with both callbacks
model.fit(
    X, y,
    epochs=15,
    validation_split=0.4,
    callbacks=[early_stopping_monitor, model_checkpoint],
    verbose=False
)

# Later, when you want to use the best model
best_model = keras.models.load_model('best_model.h5')
predictions = best_model.predict(X_test)

How It Works #

• The ModelCheckpoint callback will evaluate your model’s performance on the validation set after every epoch.
• If the val_loss (or whatever metric you’re monitoring) is better than the previous best, it will overwrite the saved model file with this new, better version.
• Once training finishes, best_model.h5 will contain the exact version of your model that achieved the lowest validation loss during training—no more worrying about those late-epoch performance drops!

Notes: #

• If you care more about validation accuracy than loss, just change monitor='val_accuracy' and mode='max' in the ModelCheckpoint config.
• Make sure the filepath you choose is accessible (you can use an absolute path if needed) to avoid permission issues.

内容的提问来源于stack exchange，提问作者dJOKER_dUMMY