keras-team/keras · error · ValueError

Feature '{name}' has `output_mode='one_hot'`. However it isn

Error message

Feature '{name}' has `output_mode='one_hot'`. However it isn't a standard feature and the dimensionality of its output space is not known, thus it cannot be one-hot encoded. Try using `output_mode='int'`.

What it means

To one-hot encode a feature, FeatureSpace must know the output dimensionality (cardinality). It can infer num_bins only from standard preprocessors (IntegerHashed/StringHashed/CategoryCrossing/Hashing, etc.). A custom or non-standard preprocessor gives no cardinality, so one_hot is impossible.

Source

Thrown at keras/src/layers/preprocessing/feature_space.py:704

                        f"Feature '{name}' has `output_mode='one_hot'`. "
                        "Thus its preprocessor should return an integer dtype. "
                        f"Instead it returns a {dtype} dtype."
                    )

                if isinstance(
                    preprocessor, (layers.IntegerLookup, layers.StringLookup)
                ):
                    cardinality = preprocessor.vocabulary_size()
                elif isinstance(preprocessor, layers.CategoryEncoding):
                    cardinality = preprocessor.num_tokens
                elif isinstance(preprocessor, layers.Discretization):
                    cardinality = preprocessor.num_bins
                elif isinstance(
                    preprocessor, (layers.HashedCrossing, layers.Hashing)
                ):
                    cardinality = preprocessor.num_bins
                else:
                    raise ValueError(
                        f"Feature '{name}' has `output_mode='one_hot'`. "
                        "However it isn't a standard feature and the "
                        "dimensionality of its output space is not known, "
                        "thus it cannot be one-hot encoded. "
                        "Try using `output_mode='int'`."
                    )
                if cardinality is not None:
                    encoder = layers.CategoryEncoding(
                        num_tokens=cardinality, output_mode="multi_hot"
                    )
                    self.one_hot_encoders[name] = encoder
                    feature = encoder(feature)

            if self.output_mode == "concat":
                dtype = feature.dtype
                if dtype.startswith("int") or dtype == "string":
                    raise ValueError(
                        f"Cannot concatenate features because feature '{name}' "

View on GitHub (pinned to 7a34a03db6)

Solutions

1. Switch the feature (or the whole FeatureSpace) to output_mode='int'
2. If you need one-hot, map the feature to a standard preprocessor that exposes num_bins (Hashing/HashedCrossing/CategoryEncoding)
3. Add a CategoryEncoding(num_tokens=known_cardinality) layer downstream to one-hot the int output yourself

Example fix

// before
fs = FeatureSpace(features={"cross": FeatureSpace.cross(feature_names=("a","b"), output_mode="one_hot")}, output_mode="concat")
// after
fs = FeatureSpace(features={"cross": FeatureSpace.cross(feature_names=("a","b"), output_mode="int")}, output_mode="int")

p = fs.preprocessors.get(name) or fs.crossers.get(name)
if not hasattr(p, "num_bins"):
    # cannot infer cardinality; use output_mode='int'

def one_hot_supported(p):
    return hasattr(p, "num_bins")

catch ValueError from get_encoded_features/__call__, then switch the FeatureSpace (or crossing) output_mode to 'int'

When it happens

Trigger: FeatureSpace(output_mode='one_hot') with a custom feature spec (e.g. FeatureSpace.feature(preprocessor=CustomLayer)) or a non-hashing crossing preprocessor, where the layer has no num_bins attribute.

Common situations: Custom crossing implementations, features built from Lambda or non-standard preprocessing layers, or crossing features without an underlying hashing/lookup layer.

Understand the failure class

Background: UnsupportedOperationException and "is not supported" errors: when a library deliberately refuses a call — this error's family across 30 libraries.

Related errors

• Feature '{name}' has `output_mode='one_hot'`. Thus its prepr
• `adapt()` can only be called on a tf.data.Dataset or a dict
• Cannot concatenate features because feature '{name}' has not
• You need to call `.adapt(dataset)` on the FeatureSpace befor
• A FeatureSpace can only be called with a dict. Received: dat

AI-assisted analysis of keras-team/keras@7a34a03db6 (2026-08-25). Data as JSON: /api/errors/280c3460cdb7056c. Report an issue: GitHub.