tf.keras.models.load_model() restores a complete Keras model, architecture, weights, and optimizer state, from a file previously created with model.save().

Syntax

tf.keras.models.load_model(filepath)

Deep Dive Course

load_model() reverses model.save(), reconstructing the full model from the saved file — its architecture, its learned weights, and its compiled optimizer along with its saved state — ready to immediately call predict(), evaluate(), or continue training with fit() exactly as if the original model object had never gone away. Unlike loading only weights with load_weights(), this requires no matching architecture already defined in code, since the architecture itself is reconstructed directly from the saved file.

1Understanding tf.keras.models.load_model()

💡

A model restored with load_model() is immediately ready to continue training with fit(), since the optimizer's saved state, not just its type, is restored too — training can genuinely resume rather than restart from a freshly initialized optimizer.

editor.html

import tensorflow as tf
from tensorflow.keras import layers, Sequential

model = Sequential([layers.Dense(2, activation='relu', input_shape=(3,))])
model.compile(optimizer='adam', loss='mse')
model.save('my_model.keras')

restored = tf.keras.models.load_model('my_model.keras')
print(restored.output_shape)

localhost:3000

2Practical Example

Here is a real-world application of tf.keras.models.load_model() showing how it is used in production TensorFlow code.

editor.html

import tensorflow as tf
from tensorflow.keras import layers, Sequential

model = Sequential([layers.Dense(2, input_shape=(3,))])
model.compile(optimizer='adam', loss='mse')
model.save('my_model.keras')

restored = tf.keras.models.load_model('my_model.keras')
print(len(restored.layers))

localhost:3000

3Best Practices

Follow these guidelines when working with tf.keras.models.load_model():

1. Use load_model() when you need a fully working model restored with no architecture code required beforehand, unlike load_weights() which needs a matching architecture already built

2. Verify the restored model's architecture with model.summary() after loading, especially when loading a model saved by a different version of the code

3. Continue training a restored model directly with fit() when needed, since its optimizer state is fully restored, not reset

⚠️

Tip: A model restored with load_model() is immediately ready to continue training with fit(), since the optimizer's saved state, not just its type, is restored too — training can genuinely resume rather than restart from a freshly initialized optimizer.

editor.html

import tensorflow as tf
from tensorflow.keras import layers, Sequential

model = Sequential([layers.Dense(2, activation='relu', input_shape=(3,))])
model.compile(optimizer='adam', loss='mse')
model.save('my_model.keras')

restored = tf.keras.models.load_model('my_model.keras')
print(restored.output_shape)

localhost:3000

Examples

Example 01Basic Usage

import tensorflow as tf
from tensorflow.keras import layers, Sequential

model = Sequential([layers.Dense(2, activation='relu', input_shape=(3,))])
model.compile(optimizer='adam', loss='mse')
model.save('my_model.keras')

restored = tf.keras.models.load_model('my_model.keras')
print(restored.output_shape)

Example 02Advanced Example

import tensorflow as tf
from tensorflow.keras import layers, Sequential

model = Sequential([layers.Dense(2, input_shape=(3,))])
model.compile(optimizer='adam', loss='mse')
model.save('my_model.keras')

restored = tf.keras.models.load_model('my_model.keras')
print(len(restored.layers))

Best Practices

Use load_model() when you need a fully working model restored with no architecture code required beforehand, unlike load_weights() which needs a matching architecture already built
Verify the restored model's architecture with model.summary() after loading, especially when loading a model saved by a different version of the code
Continue training a restored model directly with fit() when needed, since its optimizer state is fully restored, not reset

Interview Question

Why doesn't load_model() require you to redefine the model's architecture in code beforehand, unlike load_weights()?▼

Hint: Think about what information model.save() actually stores in the saved file, compared to save_weights().

model.save() stores the model's complete architecture as part of the saved file itself, a serialized description of every layer, its configuration, and how they're connected, alongside the numeric weight values — load_model() reads that architecture description directly from the file and reconstructs the model object from it, which is exactly why no matching architecture code needs to exist beforehand. save_weights(), in contrast, stores only the raw numeric parameter values with no architecture information at all, so load_weights() has nothing to reconstruct a model from — it can only load those saved numbers into the weight slots of a model whose architecture you've already built yourself in code, and that architecture must match the saved weights' shapes exactly for loading to succeed.

Exercises

MediumPractice using tf.keras.models.load_model() in a real scenario.

View Solution

import tensorflow as tf
from tensorflow.keras import layers, Sequential

model = Sequential([layers.Dense(2, activation='relu', input_shape=(3,))])
model.compile(optimizer='adam', loss='mse')
model.save('my_model.keras')

restored = tf.keras.models.load_model('my_model.keras')
print(restored.output_shape)

Frequently Asked Questions

Why doesn't load_model() require you to redefine the model's architecture in code beforehand, unlike load_weights()?▼

Related Functions

model-savemodel-load-weightsmodel-fit