Deserialization failure during load_model with unregistered custom objects

Keras models utilizing custom layers or functions rely on the get_config and from_config methods for serialization. However, these models are not fully self-contained; the corresponding custom classes or functions must be available in the environment during deserialization.

When a model is saved using the .keras format in Keras 3, the serialization captures the configuration but not the underlying Python code. If a model containing custom objects is loaded in a fresh process where those objects have not been registered via keras.utils.get_custom_objects or passed explicitly to keras.models.load_model, the process fails.

This behavior creates uncertainty when sharing models across different environments or when using classes defined in interactive sessions. Specifically, it is unclear how to ensure portability for models containing lambda layers or symbols defined in the __main__ scope without manual registration.

• What is the recommended strategy for automating the resolution of custom objects in Keras 3 to avoid manual registration during load_model?
• How does the deserialization mechanism handle symbols defined in an interactive session compared to those imported from a module?