Keras custom layers and SavedModel export: which pieces survive the round trip without custom_objects?

Answer to the three core questions

1. Inference via tf.saved_model.load()
Yes. A tf.keras model saved with model.save() writes a full TensorFlow graph (the concrete function) and all variable weights. The resulting SavedModel can be loaded with tf.saved_model.load() and its serving signatures can be called without any knowledge of the Python layer classes. The graph is static; any Python‑side logic that was not captured in the traced graph is lost.

2. Registering with @tf.keras.utils.register_keras_serializable()
Registration only makes the class discoverable by tf.keras.models.load_model() if the module that applied the decorator is imported in the loading process. The registry entry does not survive across processes automatically; you still need to import the module or supply custom_objects if the class is not importable.

3. get_config() alone is not enough for all layers
Layers that receive non‑JSON‑serializable constructor arguments (e.g., lambda functions, tf.Variable, or arbitrary Python objects) will raise a serialization error when load_model() tries to rebuild the architecture. In such cases from_config() must handle deserialization, or the constructor arguments must be made serializable.

Practical steps for a deployment pipeline

1. Ensure every custom layer implements get_config() returning only JSON‑serializable values.
2. Before calling model.save(), run a dummy forward pass or call model.build(input_shape) so all variables are created.
3. Export with model.save('path', save_format='tf').
4. For inference-only consumers, load with tf.saved_model.load('path') and invoke the exported signature.
5. For consumers that need to retrain, either:
   • Import the module defining the custom layers (or apply @tf.keras.utils.register_keras_serializable() in that module) before calling tf.keras.models.load_model('path'), or
   • Supply a custom_objects dictionary mapping layer names to the classes.
6. Verify that the loaded model produces the same outputs as the original by feeding the same input tensors.

What to check if you run into errors

When tf.keras.models.load_model() fails, confirm whether the module containing the custom layer was imported in the loading process. If it was not, the registry entry will be missing and custom_objects becomes mandatory.

Key take‑aways

• SavedModel is fully usable for inference without Python classes.
• Keras reload requires the class to be importable or registered.
• Non‑serializable constructor arguments break the config‑only path.