Serialization with activation layer does not work

[MathiesW]

Most Keras layer support an activation function. While it is possible to use string identifiers like "relu", we can also use actual activation layers like layers.ReLU().

When using a layer, the deserialization is broken, since the activations.deserialize(activation) in the from_config() method does not support an instance of class layers.Layer. This significantly reduces the flexibility, because using, e.g., LeakyReLU with a negative slope of 0.1 is not possible when you rely on loading the trained model later on.

An easy fix would be to use saving.serialize_keras_object(self.activation) in get_config() and saving.deserialize_keras_object(activation_cfg) in the from_config() method.

from keras import layers

layer = layers.Conv1D(filters=1, kernel_size=1, activation=layers.ReLU())  # works flawless
layer_from_config = layers.Conv1D.from_config(layer.get_config())

The last line throws the following exception (tested on Keras 3.6.0 and 3.9.0)
"Exception encountered: Could not interpret activation function identifier: {'module': 'keras.layers', 'class_name': 'ReLU', 'config': {'name': 're_lu', 'trainable': True, 'dtype': {'module': 'keras', 'class_name': 'DTypePolicy', 'config': {'name': 'float32'}, 'registered_name': None}, 'max_value': None, 'negative_slope': 0.0, 'threshold': 0.0}, 'registered_name': None}"

[MathiesW]

I wanted to add that the following code runs without problems and allows serialization / deserialization with a layer class as activation. The simple fix, as I hinted in my initial post, is to use saving.(de)serialize_keras_object() in the get_config() and from_config() methods of BaseConv. Since the __init__() of BaseConv calls self.activation = activations.get(activation), the following code is still compatible with using string identifiers such as "relu" as activation function.

from keras import saving, layers
from keras.src.layers.convolutional.base_conv import BaseConv

class MyBaseConv(BaseConv):
    def get_config(self):
        config: dict = super().get_config()
        config.update({"activation": saving.serialize_keras_object(self.activation)})

        return config
    
    @classmethod
    def from_config(cls, config: dict):
        activation_cfg = config.pop("activation")
        config.update({"activation": saving.deserialize_keras_object(activation_cfg)})

        return cls(**config)


class MyConv3D(MyBaseConv):
    def __init__(
        self,
        filters,
        kernel_size,
        strides=(1, 1, 1),
        padding="valid",
        data_format=None,
        dilation_rate=(1, 1, 1),
        groups=1,
        activation=None,
        use_bias=True,
        kernel_initializer="glorot_uniform",
        bias_initializer="zeros",
        kernel_regularizer=None,
        bias_regularizer=None,
        activity_regularizer=None,
        kernel_constraint=None,
        bias_constraint=None,
        **kwargs
    ):
        super().__init__(
            rank=3,
            filters=filters,
            kernel_size=kernel_size,
            strides=strides,
            padding=padding,
            data_format=data_format,
            dilation_rate=dilation_rate,
            groups=groups,
            activation=activation,
            use_bias=use_bias,
            kernel_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            kernel_regularizer=kernel_regularizer,
            bias_regularizer=bias_regularizer,
            activity_regularizer=activity_regularizer,
            kernel_constraint=kernel_constraint,
            bias_constraint=bias_constraint,
            **kwargs
        )


layer = MyConv3D(filters=1, kernel_size=1, activation=layers.ReLU(negative_slope=0.1))
layer_from_config = MyConv3D.from_config(layer.get_config())

[dhantule]

Hi @MathiesW, thanks for reporting this.
I have tested your code with Keras 3.9.0 and I'm facing the same error in this gist, we'll look into this issue and update you.