Chapter 19: Saving, Loading, and Callbacks

“Training is hard-earned. Never lose progress. Save smart, load fast, and train with foresight.”

---

This chapter focuses on one of the most practical and essential aspects of any machine learning workflow: saving and restoring models and automating training behavior with callbacks.

By the end, you'll know how to:

• Save your model weights or full architecture
• Load and resume training or inference
• Use callbacks like EarlyStopping, ModelCheckpoint, and custom logic
• Ensure your work survives interruptions and scales reliably

Saving Your Model

There are two main formats in TensorFlow:

1. SavedModel Format (Recommended)

model.save('saved_model/my_model')

To load:

new_model = tf.keras.models.load_model('saved_model/my_model')

✅ Includes:

• Architecture
• Weights
• Training config
• Optimizer state

2. HDF5 Format (.h5)

model.save('my_model.h5')
new_model = tf.keras.models.load_model('my_model.h5')

This format is widely used in other frameworks but has limitations with newer custom layers and optimizers.

---

Saving Only Weights

If you just want to save parameters:

model.save_weights('weights.ckpt')

To load them into the same architecture:

model.load_weights('weights.ckpt')

Useful when:

• You want to share weights but not the model code
• You're using subclassed models

---

Using Callbacks

Callbacks are hooks into the training loop that let you intervene at key moments: epoch end, batch end, etc.

🔹 Common Callbacks

1. Early Stopping Stop training if validation loss doesn’t improve.

tf.keras.callbacks.EarlyStopping(
    monitor='val_loss',
    patience=3,
    restore_best_weights=True
)

2. Model Checkpoint Save model during training.

tf.keras.callbacks.ModelCheckpoint(
    filepath='best_model.h5',
    save_best_only=True,
    monitor='val_loss',
    mode='min'
)

3. Learning Rate Scheduler Adjust learning rate during training.

tf.keras.callbacks.LearningRateScheduler(lambda epoch: 1e-3 * 0.95 ** epoch)

---

Using Callbacks in fit()

callbacks = [
    tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=3),
    tf.keras.callbacks.ModelCheckpoint(filepath='best_model.keras', save_best_only=True)
]

model.fit(
    x_train, y_train,
    validation_split=0.2,
    epochs=20,
    callbacks=callbacks
)

---

Custom Callback (Advanced)

You can define your own logic:

class PrintEpoch(tf.keras.callbacks.Callback):
    def on_epoch_end(self, epoch, logs=None):
        print(f"Epoch {epoch+1} finished. Accuracy: {logs['accuracy']:.4f}")

Use with:

model.fit(..., callbacks=[PrintEpoch()])

---

Summary

In this chapter, you:

• Learned to save and load models in multiple formats
• Used callbacks to control training and automate decisions
• Understood when to save weights only vs full model
• Built your own custom callback for fine-grained control

These tools are critical for long training runs, collaborative projects, and production deployment. Always train as if your power might go out. ⚡

---