Deep Learning Toolbox Converter for TensorFlow Models

Import and export TensorFlow models within MATLAB

The converter for TensorFlow models enables you to import a pretrained TensorFlow model and weights and export a MATLAB network or layergraph as a TensorFlow model.

Import:

To import a TensorFlow network in MATLAB, please refer to importNetworkFromTensorFlow.

Note: the model must be saved in TensorFlow in the SavedModel format before importing to MATLAB.

For TensorFlow 2 models up to version 2.15, save the model (myModelTF, expressed as a tf.Keras.Model object) with the following command in TensorFlow first:

model.save("myModelTF")

For TensorFlow 2 models with versions 2.16 or later, TensorFlow will be installed with Keras 3 instead of Keras 2. To ensure compatibility with importNetworkFromTensorFlow, please build and save the save the model using the Keras 2 API using the following Python commands.

First, install the Keras 2 (tf_keras) package in bash:

pip install tf_keras

Then, precede any import tensorflow statement by the following lines of code:

import os
os.environ["TF_USE_LEGACY_KERAS"] = "1"
import tensorflow

Next, instead of importing the keras (Keras 3) package directly, import tf_keras (Keras 2) as keras:

import tf_keras as keras
from tf_keras import layers

Finally, save the model:

model.save("myModelTF")

The TensorFlow Model Importer has been tested up to 2.15. SavedModels saved with versions 2.15 and higher may also work without issues. When using tf_keras for versions 2.15 and higher, Keras 3 specific features such as Keras Ops will not be supported.

Export:

To export a TensorFlow network from MATLAB, please refer to exportNetworkToTensorFlow.

Export supports:

• TensorFlow v2.0 or later, no support for Keras 3.0
• Python version 3.0 or later