I am implementing a custom Keras Layer where the number of internal sublayers depends on a user-defined parameter, so I don’t know the number of layers in advance.

To handle this, I create the sublayers in a loop, assign each one as an attribute using setattr, and also keep them in a Python list for easy iteration:

@keras.saving.register_keras_serializable()
class MyBlock(tf.keras.layers.Layer):
    def __init__(self, n_layers):
        super().__init__()
        self.n_layers = n_layers
        self.blocks = []

        for i in range(n_layers):
            layer = tf.keras.layers.Dense(8)
            setattr(self, f"block_{i}", layer)
            self.blocks.append(getattr(self, f"block_{i}"))

    def call(self, x):
        for layer in self.blocks:
            x = layer(x)
        return x

    def get_config(self):
        return {"n_layers": self.n_layers}

My understanding is that:

• Keras only tracks layers assigned as attributes

• Lists alone are not tracked

• Using setattr ensures each sublayer is registered and serialized correctly

Is this approach correct and safe for model saving/loading, or is there a better Keras-recommended pattern for variable numbers of sublayers?