Saving Keras models with Custom Layers

Question

I am trying to save a Keras model in a H5 file. The Keras model has a custom layer . When I try to restore the model , I get the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-5-0fbff9b56a9d> in <module>()
      1 model.save('model.h5')
      2 del model
----> 3 model = tf.keras.models.load_model('model.h5')

8 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/utils/generic_utils.py in class_and_config_for_serialized_keras_object(config, module_objects, custom_objects, printable_module_name)
    319   cls = get_registered_object(class_name, custom_objects, module_objects)
    320   if cls is None:
--> 321     raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)
    322 
    323   cls_config = config['config']

ValueError: Unknown layer: CustomLayer

Could you please tell me how I am supposed to save and load weights of all the custom Keras layers too? (Also, there was no warning when saving, will it be possible to load models from H5 files which I have already saved but can't load back now?)

Here is the minimal working code sample (MCVE) for this error, as well as the full expanded message: Google Colab Notebook

Just for completeness, this is the code I used to make my custom layer. get_config and from_config are both working fine.

class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, name=None):
        super(CustomLayer, self).__init__(name=name)
        self.k = k

    def get_config(self):
        return {'k': self.k}

    def call(self, input):
        return tf.multiply(input, 2)

model = tf.keras.models.Sequential([
    tf.keras.Input(name='input_layer', shape=(10,)),
    CustomLayer(10, name='custom_layer'),
    tf.keras.layers.Dense(1, activation='sigmoid', name='output_layer')
])
model.save('model.h5')
model = tf.keras.models.load_model('model.h5')

Answer 1

Correction number 1 is to use Custom_Objects while loading the Saved Model ie, replace the code,

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

with

new_model = tf.keras.models.load_model('model.h5', custom_objects={'CustomLayer': CustomLayer})

Since we are using Custom Layers to build the Model and before Saving it, we should use Custom Objects while Loading it.

Correction number 2 is to add **kwargs in the __init__ function of the Custom Layer like

def __init__(self, k, name=None, **kwargs):
        super(CustomLayer, self).__init__(name=name)
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)

Complete working code is shown below:

import tensorflow as tf

class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, name=None, **kwargs):
        super(CustomLayer, self).__init__(name=name)
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)


    def get_config(self):
        config = super(CustomLayer, self).get_config()
        config.update({"k": self.k})
        return config

    def call(self, input):
        return tf.multiply(input, 2)

model = tf.keras.models.Sequential([
    tf.keras.Input(name='input_layer', shape=(10,)),
    CustomLayer(10, name='custom_layer'),
    tf.keras.layers.Dense(1, activation='sigmoid', name='output_layer')
])
tf.keras.models.save_model(model, 'model.h5')
new_model = tf.keras.models.load_model('model.h5', custom_objects={'CustomLayer': CustomLayer})

print(new_model.summary())

Output of the above code is shown below:

WARNING:tensorflow:No training configuration found in the save file, so the model was *not* compiled. Compile it manually.
Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
custom_layer_1 (CustomLayer) (None, 10)                0         
_________________________________________________________________
output_layer (Dense)         (None, 1)                 11        
=================================================================
Total params: 11
Trainable params: 11
Non-trainable params: 0

Hope this helps. Happy Learning!

Answer 2

You can provide manually the mapping custom_objects in the load_model method as mentioned in the answer https://stackoverflow.com/a/62326857/8056572 but it can be tedious when you have a lot of custom layers (or any custom callables defined. eg metrics, losses, optimizers, ...).

Tensorflow provides a utils function to do it automatically: tf.keras.utils.register_keras_serializable

You have to update your CustomLayer as follows:

import tensorflow as tf

@tf.keras.utils.register_keras_serializable()
class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, **kwargs):
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)

    def get_config(self):
        config = super().get_config()
        config["k"] = self.k
        return config

    def call(self, input):
        return tf.multiply(input, 2)

---

Here is the complete working code:

import tensorflow as tf


@tf.keras.utils.register_keras_serializable()
class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, **kwargs):
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)

    def get_config(self):
        config = super().get_config()
        config["k"] = self.k
        return config

    def call(self, input):
        return tf.multiply(input, 2)


def main():
    model = tf.keras.models.Sequential(
        [
            tf.keras.Input(name='input_layer', shape=(10,)),
            CustomLayer(10, name='custom_layer'),
            tf.keras.layers.Dense(1, activation='sigmoid', name='output_layer')
        ]
    )
    print("SUMMARY OF THE MODEL CREATED")
    print("-" * 60)
    print(model.summary())
    model.save('model.h5')

    del model

    print()
    print()

    model = tf.keras.models.load_model('model.h5')
    print("SUMMARY OF THE MODEL LOADED")
    print("-" * 60)
    print(model.summary())

if __name__ == "__main__":
    main()

And the corresponding output:

SUMMARY OF THE MODEL CREATED
------------------------------------------------------------
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
custom_layer (CustomLayer)   (None, 10)                0         
_________________________________________________________________
output_layer (Dense)         (None, 1)                 11        
=================================================================
Total params: 11
Trainable params: 11
Non-trainable params: 0
_________________________________________________________________
None


WARNING:tensorflow:No training configuration found in the save file, so the model was *not* compiled. Compile it manually.
SUMMARY OF THE MODEL LOADED
------------------------------------------------------------
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
custom_layer (CustomLayer)   (None, 10)                0         
_________________________________________________________________
output_layer (Dense)         (None, 1)                 11        
=================================================================
Total params: 11
Trainable params: 11
Non-trainable params: 0
_________________________________________________________________
None

Answer 3

The other answer provided here does not work in a real example. The code example is running in the same process, so the CustomLayer is available in-memory when loading the model. If you try to load the model in a new process it will fail and ask you to use the custom_objects argument. I am still looking for a clean way to export custom objects from the source code.