5 Best Ways to Save and Serialize Models with Keras in Python

Rate this post

πŸ’‘ Problem Formulation: When developing machine learning models with Keras, a Python deep learning library, it’s crucial for practitioners to know how to save and serialize their models. This preserves models’ states, allowing for resume training, model sharing, and deployment without the need to retrain. An example of input could be a fully trained Keras model while the desired output is a saved file that contains all the essential information to recreate the model, including its architecture, weights, and training configuration.

Method 1: Save the Entire Model to a HDF5 file

The save() method in Keras allows you to save an entire model into a single HDF5 file which contains the model’s architecture, weights, and training configuration. This method is very convenient because it bundles everything into one neat file, which can be loaded later without requiring the original code used to create the model.

Here’s an example:

from keras.models import Sequential
from keras.layers import Dense
model = Sequential([
    Dense(units=1, input_shape=[1])

Output: A single HDF5 file named ‘my_model.h5’.

This code first defines a simple Keras Sequential model with a single dense layer. After defining the model, it simply calls model.save(), specifying the filename to which the model should be saved. With this, the model can be later loaded using Keras’ load_model function.

Method 2: Saving the Model Weights and Architecture Separately

Sometimes, you may want to save the weights and architecture of the model separately. This can be done by using the model.save_weights() to save the weights and model.to_json() (or model.to_yaml()) to save the architecture. This method provides more flexibility but requires extra steps to recreate the model from the files.

Here’s an example:

json_model = model.to_json()
with open('model_architecture.json', 'w') as json_file:

Output: Two separate files – ‘model_architecture.json’ and ‘model_weights.h5’.

This snippet first converts the model architecture into a JSON formatted string using the to_json() method, then writes this string to a JSON file. The model’s weights are saved separately in an HDF5 file using save_weights(). Reconstituting this model requires loading both the architecture and the weights, then compiling the model.

Method 3: Saving Only the Model Weights

Saving only the model weights can be useful when you have the model architecture available in your code and you only need to preserve the trained weights. This can significantly save space for complex models. The weights are saved using the save_weights() method, which creates an HDF5 or TensorFlow Checkpoint file.

Here’s an example:

# To save as HDF5 format, specify the ''.h5' extension:
# model.save_weights('model_weights.h5')

Output: A TensorFlow Checkpoint folder or HDF5 file containing model weights.

This code shows how to save the trained weights of your Keras model to disk. With this method, you can later create a model with the same architecture and load these weights into it using the model.load_weights() function.

Method 4: Use Model Checkpointing

Model checkpointing is a way to automatically save the model during training, usually at the end of each epoch, and optionally save only the best model according to a certain metric. Keras provides the ModelCheckpoint callback for this purpose, which can be used to perform saves at desired points during the training process.

Here’s an example:

from keras.callbacks import ModelCheckpoint
checkpoint = ModelCheckpoint('best_model.h5', save_best_only=True, monitor='val_loss', mode='min')
model.fit(X_train, Y_train, validation_split=0.2, epochs=10, callbacks=[checkpoint])

Output: The best-performing model file ‘best_model.h5’ based on validation loss.

In this code, we first create a ModelCheckpoint callback specifying we only want to save the best model according to minimum validation loss. This callback is then passed to the model’s fit() method, which ensures the model is saved automatically during training.

Bonus One-Liner Method 5: Save Model as a TensorFlow SavedModel

TensorFlow’s SavedModel format allows you to store not only the model’s weights and architecture but also its training configuration and even the optimizer’s state, which is useful for resuming training exactly where you left off. This can be done through the model.save('my_saved_model') function provided by Keras, which uses the SavedModel format by default when no file extension is given.

Here’s an example:


Output: A SavedModel directory with the name ‘my_saved_model’.

This simple one-liner saves the entire model in TensorFlow’s SavedModel format into the specified directory. This saves the model architecture, weights, and training configuration all together in a format that’s also compatible with TensorFlow serving.


  • Method 1: Save the Entire Model. Strengths: Simplicity, convenience, and portability. Weaknesses: Potentially larger file size due to bundling all model information together.
  • Method 2: Save Weights and Architecture Separately. Strengths: Flexibility in modifying architecture and weights independently. Weaknesses: Requires additional code to reconstruct the model.
  • Method 3: Saving Only the Weights. Strengths: Smaller file size, quick loading if architecture is known. Weaknesses: Inapplicable if model architecture is not readily available.
  • Method 4: Model Checkpointing. Strengths: Useful in long training processes, can save only the best model according to performance on a metric. Weaknesses: Additional overhead in setting up callbacks.
  • Method 5: TensorFlow SavedModel. Strengths: Comprehensive save format including optimizer state, ideal for TensorFlow ecosystem. Weaknesses: Files can be large, tied to TensorFlow’s framework.