keras-team/keras · error · ValueError

The endpoint '{call_endpoint}' is neither an attribute of th

Error message

The endpoint '{call_endpoint}' is neither an attribute of the reloaded SavedModel, nor an entry in the `signatures` field of the reloaded SavedModel. Select another endpoint via the `call_endpoint` argument. Available endpoints for this SavedModel: {list(self._reloaded_obj.signatures.keys())}

What it means

TFSMLayer resolves call_endpoint by first checking attributes of the reloaded SavedModel, then its signatures dict; if neither contains the requested name it raises this ValueError listing the actually available endpoints. The endpoint name must exactly match an attribute of the loaded object or a key in reloaded.signatures. Typos, wrong casing, and endpoints that were never exported all fail here.

Source

Thrown at keras/src/export/tfsm_layer.py:81

        # Initialize an empty layer, then add_weight() etc. as needed.
        super().__init__(trainable=trainable, name=name, dtype=dtype)

        self._reloaded_obj = tf.saved_model.load(filepath)

        self.filepath = filepath
        self.call_endpoint = call_endpoint
        self.call_training_endpoint = call_training_endpoint

        # Resolve the call function.
        if hasattr(self._reloaded_obj, call_endpoint):
            # Case 1: it's set as an attribute.
            self.call_endpoint_fn = getattr(self._reloaded_obj, call_endpoint)
        elif call_endpoint in self._reloaded_obj.signatures:
            # Case 2: it's listed in the `signatures` field.
            self.call_endpoint_fn = self._reloaded_obj.signatures[call_endpoint]
        else:
            raise ValueError(
                f"The endpoint '{call_endpoint}' "
                "is neither an attribute of the reloaded SavedModel, "
                "nor an entry in the `signatures` field of "
                "the reloaded SavedModel. Select another endpoint via "
                "the `call_endpoint` argument. Available endpoints for "
                "this SavedModel: "
                f"{list(self._reloaded_obj.signatures.keys())}"
            )

        # Resolving the training function.
        if call_training_endpoint:
            if hasattr(self._reloaded_obj, call_training_endpoint):
                self.call_training_endpoint_fn = getattr(
                    self._reloaded_obj, call_training_endpoint
                )
            elif call_training_endpoint in self._reloaded_obj.signatures:
                self.call_training_endpoint_fn = self._reloaded_obj.signatures[
                    call_training_endpoint

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Read the error message: it lists the valid endpoints; use one of those names for call_endpoint.
2. Inspect before loading: m = tf.saved_model.load(path); print(list(m.signatures.keys())) and dir(m) for attribute endpoints.
3. Re-export with an explicit signature name and use that same name on both sides.

Example fix

# before
layer = TFSMLayer('saved_model/', call_endpoint='predict')  # -> ValueError

# after
import tensorflow as tf
print(list(tf.saved_model.load('saved_model/').signatures.keys()))  # e.g. ['serving_default']
layer = TFSMLayer('saved_model/', call_endpoint='serving_default')

import tensorflow as tf
loaded = tf.saved_model.load(path)
valid = set(loaded.signatures.keys()) | {a for a in dir(loaded) if not a.startswith('_')}
assert call_endpoint in valid, f'use one of {sorted(valid)}'

try:
    layer = TFSMLayer(path, call_endpoint=name)
except ValueError as e:
    # the message lists available endpoints; parse it or surface it to the user
    raise

When it happens

Trigger: keras.layers.TFSMLayer(path, call_endpoint='serve') when the model only exported 'serving_default'; using a custom signature name that was not passed to tf.saved_model.save(..., signatures={...}); passing the endpoint name of a different SavedModel version.

Common situations: Mismatch between the signature name used at export time (export_archive.add_endpoint(..., name='x')) and at load time; TF-Hub models whose only endpoint is 'serving_default'; renamed endpoints after retraining or re-export.

Related errors

• The endpoint '{call_training_endpoint}' is neither an attrib
• If using `weights="imagenet"` as true, `classes` should be 1
• The number of repeats in `EfficientNet` must be > 0. Receive
• If using `weights="imagenet"` as true, `classes` should be 1
• The number of repeats in `EfficientNetV2` must be > 0. Recei

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/4c4cb5ec9a791d02. Report an issue: GitHub.